Societal Implications of Artificial Intelligence

Dennis Holeman

Adapted from a five-session class given at the Osher Lifelong Learning Institute (OLLI) at Southern Oregon University, Ashland, Oregon in March 2019

This paper focuses on how artificial intelligence is going to have major societal implications and what actions need to be taken to avoid and mitigate negative effects of its applications.

Introduction

What this paper is about

  • A general understanding of Artificial Intelligence (AI) and its relationship with natural intelligence
  • An appreciation for why AI and robotics are in the news now and why they are so significant
  • A perspective on where AI has come from and where it is going
  • An overview of what AI is going to be used for
  • A projection of how AI will impact peoples’ daily lives
  • Predictions of the implications of AI for different fields—medicine, law, finance, government, the military, and many others

What is covered and what isn’t

  • Is: An overall conceptual framework on AI and robotics
    • A broad-brush perspective
    • A highly simplified treatment
    • The primary focus will be on the societal implications
  • Isn’t: Detailed technical aspects of AI
    • Not a short course on developing and using AI
    • No mathematics
    • No flow charts
    • No software code!

Why this is important

  • AI is currently getting a tremendous amount of attention
    • Lots of pieces on AI in the media
    • Dire, even apocalyptic warnings from famous people
  • Some people characterize AI as the most significant technical development since the Industrial Revolution
    • When machine power took over from human and animal muscle power
  • We are recognizing that we are entering a period of extremely rapid change as a consequence of AI
  • It’s hard to get a perspective on what’s real
  • What should we really be thinking about in regards to AI and robotics?

Fears about AI

  • Many fears are being expressed about AI these days
    • Humans will cede every decision to machines
    • No one will have a job and the economy will crash
    • AI will be used for brutal societal control, a la 1984
    • AI systems will take over from humans
    • Robots will turn malevolent
    • AGI (Artificial General Intelligence) will make us obsolete and unneeded
  • Are these realistic, or just scare-mongering?

Humankind’s big technical revolutions

In rough order, here are major technical revolutions humans have developed:

  • Control of fire
  • Cooking
  • Edged tools and weapons
  • Sewn clothing
  • Watercraft
  • Agriculture
  • Domestication of animals
  • Metalworking
  • Writing (enabling the external storage of knowledge)
  • Money (enabling extensive division of labor,  leading to complex social networks)
  • Machinery
  • Printing press (enabling the democratization of knowledge)
  • Exploitation of fossil fuel energy
  • The industrial revolution
  • Electricity
  • Electronics  (enabling electronic computing and networked communications)

Each revolution has been highly disruptive

The next high impact technologies

  • Two technologies have enormous potential consequences in the 21st century
  • The first is synthetic biology, with the ability to engineer biological organisms with wholly new characteristics
    • Potential to change what a human being is
  • The second is artificial intelligence
    • Potential to create powerful non-human intelligences
  • The two technologies are projected to be synergistic
  • They could be the best things that have ever happened in human history—or the worst
  • Managing them wisely is a critical task we must face

New intelligent beings?

  • Currently, we tend to think that human individuals are Earth’s only beings with intelligence
  • We are in the process of creating new intelligent entities, in the form of intelligent machines and systems
  • Is this truly a good thing?  Have we really thought this through carefully?
  • What could be possible unintended consequences?

Key questions

  • What are big problems we face, to which artificial intelligence is the necessary solution?
    • Or at least a solution?
  • Do the prospective benefits of AI and robotics outweigh the prospective costs and risks?
  • And in which areas?
  • Who should decide what path(s) we take with AI?
  • Who can/should set policies for AI and robotics?
  • How will the expansion of AI and robotics interact with other physical megatrends of the 21st century?  e.g.,
    • Human population growth
    • Urbanization
    • Global climate change
    • Ocean impacts (acidification, de-oxygenization, etc.)
    • Sea level rise
    • Pollution of all types
    • Ecosystem damage in general
    • Species extinction
    • Resource depletion and scarcity
  • And how will AI and robotics interact with major societal changes we’re observing?  e.g.,
    • Changing geopolitical alignments and power
    • Globalization
    • Increasing economic inequality
    • Financial engineering, rather than investment
    • Expanding debt (public, private, corporate)
    • Hardening political polarization
    • Religious and secular intergroup conflicts
    • Refugee migrations
    • Shifts to the political right, authoritarian regimes

About the author

  • Trained as a generalist systems engineer
  • Job: define user requirements, write specifications, do tradeoffs, create the top-level system design, define tests
  • Over 36 years as the senior systems engineer at SRI International (the former Stanford Research Institute); retired in early 2015
  • Extensive involvement in utilizations of AI, robotics, and unmanned vehicle systems of many types and environments
  • Not an AI researcher or developer—an applier of intelligent systems
  • Neither a proponent for AI, nor an opponent

Just what do we mean by “artificial intelligence”?

  • No definition accepted by everyone
  • One I use: a machine system able to do something that would be considered intelligent if a human being did it
    • Able to make own decisions in a novel situation
  • Now: narrow domains of application only
  • Already can perform as well or better than a human in many specialized areas
    • However, outside its particular niche, a machine system may be utterly stupid
  • Distinguished from computing that simply executes a fixed program, however complex
  • What we don’t have is general purpose AI that is similar to the full set of human capabilities
    • Not now, and maybe not ever
  • Furthermore, such a general purpose AI will not necessarily think the same way a human will think

Why is AI so significant now?

Some recent breakthroughs

  • The AI field has had some major breakthroughs in the past several years
    • Particularly a machine learning technique called Deep Learning
  • Problems that were intractable previously have become much easier to address
  • Amazing performance has been achieved in some significant areas
  • This has led to an explosion in new interest and investment in AI

The expansion of smart systems

  • Systems with some level of AI are currently proliferating at an tremendous pace
    • Vastly faster than the human population
  • Bots on the Web already outnumber human users
    • Many more smartphones are produced every day than human babies
  • Each equipped with a virtual personal assistant
  • By 2021, projected to be a billion smart speakers in use
  • Digital data files are exploding
    • More than 100 hours of video are uploaded to the Internet every minute

Ever-cheaper computation

  • A big driver is the phenomenal decrease in the cost of machine computation
  • How many computations per second could you obtain for $1,000?
    • 1950: ~10 (mechanical calculator)
    • 2018: ~100 trillion (specialized chipsets)

AI enhancement of human intelligence

  • AI-enhancement of human intelligence is already here
  • Access to virtually the entire world’s knowledge through an intelligent virtual assistant system
    connected to the Internet

    • E.g., via Siri on your iPhone or Alexa on your Echo speaker

AI capabilities have become usable for a wide range of functions

  • Many complex systems now incorporate large amounts of intelligence, particularly for control
  • Capable speech understanding has become routine
  • High performance image recognition and understanding has also become routine
  • Every field that has large amounts of digitally-encoded data is a candidate for AI-based analysis
  • Costs of using AI capabilities have plummeted
  • Software tools for creating AI systems are now widely available

Some examples: what is AI doing now?  

  • Flying airliners
  • Routing mail and packages
  • Approving loans
  • Evaluating job applications
  • Recommending entertainment and online purchases
  • Suggesting possible partners via dating apps
  • Trading securities (stocks, bonds, commodities, etc.)
  • Managing electric power grids
  • Controlling traffic
  • Planning trucking routes and schedules

Some current major AI utilizations

  • Voice interaction of all sorts (chatbots, etc.)
  • Personal assistance (e.g., Siri, Alexa)
  • Text recognition
  • Spoken and written language translation
  • Face recognition, other biometric identification
  • Cyberattack detection, identification, and isolation
  • Document review (e.g., tax returns, mortgages, etc.)
  • Insurance risk management
  • Medical test evaluation, diagnosis, treatment recommendation
  • Fraud detection, other crime detection
  • Computer-generated imagery for movies and television
  • Interactive video games
  • And, of course, robotics
  • Many, many others

Some everyday current interactions you have with AI

  • Use spell and grammar correction in word processing
  • Do a web search and get the most relevant pages
  • Look at a product online and get recommendations for similar or related offerings
  • Have the ATM read your handwritten check
  • Use the Uber or Lyft app and get the nearest car
  • Read your email, and have little or no spam to deal with
  • Have viruses and other malware kept out of your computer
  • Submit an application to some institution
  • Have your credit card accepted at a store
  • Look for an image on Google
  • Post a photo or a video on Facebook and have it linked
  • Read news online, curated to match your interests
  • Pass by a security video camera or a license plate reader
  • Read the transcript of a voicemail message left on your phone

From Simple Automation to AI

Progression of machine capabilities

  • The development of electronics in the mid-20th century
  • A steady incorporation of ever-greater computation and reasoning capabilities in our machines and systems
  • Our machines require less and less human involvement to do more and more complex things

Automation

  • Automated systems operate under their own control, without human moment-to-moment guidance
  • Simple automation has been around a long time
  • Most automated systems execute fixed programs
  • Changes have to be made by humans
  • Automation can be highly complex mechanically and electrically, but isn’t what you would call intelligent

Feedback control

  • A key milestone in automation: understanding feedback and the development of feedback control
  • Allowed systems to be stable under changing conditions
  • A thermostat system illustrates simple feedback control

Nested feedback control loops

  • Multiple feedback loops can be nested
  • Operating at different rates, from fast at the inner loop to slow at the outer loop
  • Loops provide progressively higher levels of control

Teleoperation

  • Some mobile systems have control of the outer loop by a human located outside the system
  • Nikola Tesla demonstrated the very first example of radio remote control in 1898
  • Things like RC model airplanes have been around since the late 1930s
  • There are many different kinds of teleoperated systems

Telepresence

  • Telepresence extends teleoperation
  • The system transmits information so the human remote operator has the sense of being present onboard
  • Telepresence allows complex interactive operations in locations where a human can’t physically go
  • The following video shows undersea telepresence

https://www.youtube.com/watch?v=CoOwT0X5dpo

Robots

  • Robots extend automated systems to a new level
    • They go beyond control by a remote human operator
  • Internal software control provides the capacity to sense and move in complex ways autonomously
  • Robots can be hybrids, with a mixture of autonomous capabilities and remote human operation
  • Can be fixed in place (e.g., factory robots) or mobile
  • Most robots are specialized systems, optimized for a limited range of functions
  • Very few resemble droids in the movies

Robots with increasing AI

  • Now: some robotic systems have the ability to do extensive reasoning and adapt themselves autonomously to the conditions they encounter
    • Learn and improve with experience
    • Change their own programming
  • AI capabilities are particularly important for robots operating in unconstrained environments
    • Conditions that weren’t anticipated when they were programmed

Some Contemporary Examples of Complex Automated Systems

Highly automated aircraft

    • Latest-generation airliners (e.g., the Boeing 787 Dreamliner) are capable of fully automated pre-programmed self-control from pushback to parking at the destination gate
  • Pilots of such aircraft are estimated to be in direct hands-on control only about 5 to 7 minutes of a typical flight

Some other highly automated vehicles

    • Ships, particularly large cargo ships
      • Completely crewless ships are being developed
      • Monitoring at shore-based stations via datalink
    • Space launchers
    • Unmanned spacecraft
    • Trains, particularly freights and urban light rail vehicles
  • Drone aircraft (e.g., delivery drones)

Robotic factories

  • The following video shows Tesla’s highly-roboticized automobile factory in Fremont, California

https://www.youtube.com/watch?v=8_lfxPI5ObM

Robotic warehouses

  • Amazon alone has over 100,000 warehouse robots, as of late 2017
  • The following video illustrates robotic warehouse technology

https://www.youtube.com/watch?v=FBl4Y55V2Z4

Robotic self-driving trucks

  • This video from Caterpillar describes their self-driving trucks for mining operations

https://www.youtube.com/watch?v=GEsKZSNR9As

  • Driverless farm equipment (combines, etc.) is a similar application

Robotic surgery systems

Cars with increasing automation

  • The latest cars are being sold with ever-increasing amounts of intelligent driver assistance
    • Navigation and route recommendation
    • Intelligent cruise control, including spacing from the vehicles in front and automatic lane keeping
    • Automatic anti-collision braking
    • Traction control, stability control
    • Self-parking
    • Driver monitoring (drowsy? distracted? angry? drunk?)
    • Smart headlights

The software-defined automobile

  • Tesla: the car’s intelligent functions are regularly upgraded by software updates, distributed wirelessly
  • The car changes its capabilities over time

Self-driving cars

  • Can have big impacts on society, but lots of issues
  • Currently a great amount of hype, but fully self-driving cars actually far from ready for general use
  • Critical not to trust such systems prematurely
  • Already semi-automated cars are having accidents
  • The regulatory and legal environments are lagging far behind the technology

Biomimetic Robots

Non-humanoid robots

Anthropomorphic robots

Cyborgs

  • So far we have distinguished between organic intelligences (humans and other intelligent animals) and non-organic intelligences (machines)
  • Another category is cyborgs, where organic intelligences are complemented with machine intelligences
  • Already using AI-enabled implanted assistance systems
    • Cardiac pacemakers
    • Drug delivery systems
    • Cochlear implants
    • Prosthetic limbs, hands, and feet
    • Deep brain stimulation devices
  • AI-enabled vision implants are coming

Thinking about Intelligence

General capabilities of intelligence

  • Search for and recognize patterns, both static and dynamic
  • Measure, put in order, compare, and rank
  • Reason from available information
    • Including deductive, inductive, and abductive reasoning
  • Weigh pieces of evidence to draw conclusions
  • Perform abstraction and think abstractly
  • Perceive, understand, and maintain awareness of
    • The state of the external environment
    • Particularly changes in that state
    • The state of the system’s internal processes
    • Particularly changes in that state
  • Modify behavior to match the perceived external and internal states
  • Plan
  • Solve problems, games, and puzzles
  • Navigate in complex environments
  • Learn from experience
    • The entity’s own experience
    • The experience of other entities
  • Infer cause and effect relationships
  • Predict and anticipate future conditions
  • Project possible consequences of actions
  • Make tradeoffs and choose among alternatives
  • Do real-time adaptation for cooperation
  • Teach others

Intelligent animals

  • Many intelligent animals can do the things just described, at least to some degree
    • Primates; elephants; cetaceans; canids; corvids; psittacines; raccoons; pinnipeds; pigs; many more
    • Some we don’t normally think of, such as octopi
    • Even spiders and bees have significant intelligence
  • An animal’s intelligence is optimized for surviving and thriving in the environment the animal inhabits
    • This is true for how human intelligence evolved, as well!

Some other aspects of intelligence

  • Perceive and understand how a system works and then game the system for advantage
  • Perceive and understand the actions of other intelligent systems
    • Infer the objectives and strategies of other intelligent systems and modify behavior accordingly

Duplicating animal intelligence

  • It’s not only hard to duplicate human intelligence with a machine system
  • Animal intelligences can be pretty challenging to duplicate, too
    • E.g., a Clark’s Nutcracker can reliably relocate food it stores for the winter, with 5,000 caches spread over a 15 mile area
  • Think you could match this?
    • “Bird brain” should be considered a compliment!

Intelligence is a continuum

  • There is no threshold that distinguishes intelligence
    • Slime molds(!) have been shown to be able to navigate a maze to reach a tasty bit of food
  • Human thinking is only one kind of thinking

Some advanced aspects of intelligence

  • Comprehend complex ideas
  • Reason from analogy, metaphor, parable, etc.
  • Synthesize information from multiple sources
  • Generalize and abstract
  • Transfer learning from one context to another
  • Collect knowledge and organize it into a structure
  • Examine knowledge to derive understanding, insight, and wisdom
  • Structure problems to help solve them
  • Formulate hypotheses and ask incisive questions
  • Examine and modify assumptions to see effects

Intelligence can be paradoxical

  • Some things humans consider to be hard tasks for the most intelligent people turn out to be easy for machine intelligence—e.g.,
    • Solve extremely complex mathematics problems
    • Play games like chess at the master level
    • Translate texts between languages
  • At the same time, tasks that any normal 3 year old child can do effortlessly are very challenging for machine intelligences
    • E.g., recognize the class “kitty” on the basis of just a few examples
    • Read and understand the emotional states of others

Multiple intelligences 

In addition to general intelligence, human intelligences come in particular aspects*

  • Verbal-linguistic intelligence (“word smart”)
  • Logical-mathematical intelligence (“number/ reasoning smart”)
  • Visual-spatial intelligence (“picture smart”)
  • Musical-rhythmic and harmonic intelligence (“music smart”)
  • Bodily-kinesthetic intelligence (“body smart”)
  • Interpersonal intelligence (“people smart”)
  • Intrapersonal intelligence (“self smart”)
  • Naturalistic intelligence (“nature smart”)
  • Obviously, different humans have different profiles in terms of their intelligence areas
  • We don’t expect a particular human to be at the top of the scale in all of the areas

* First proposed by Howard Gardener; people are proposing additional ones

Idiot savants

  • There are humans who have extraordinary abilities in one narrow area of intelligence, but are mentally deficient in virtually every other aspect
    • E.g., “lightning calculators” or memory prodigies who can’t tie their own shoes
  • AI systems currently tend to be analogous to idiot savants
    • Phenomenal capability,  but only in one very narrow subject area
    • Completely stupid outside their area of expertise

The information pyramid

There are different kinds of knowledge

  • Propositional knowledge: knowledge of facts
  • Procedural knowledge: knowledge of how to do something
  • Each type of knowledge is encoded in different ways
  • In addition to explicit knowledge, there is tacit knowledge that can’t be articulated
    • Can’t be written down or verbalized
    • Things like skills, ideas, and experiences
  • The upper levels of the pyramid generally require the integration of all three types of knowledge

AI Capability Projections

AI capabilities are going to increase faster than people expect

  • AI is on an exponential growth curve
  • All contributing technologies are increasing rapidly
    • Extensive synergies between them
  • Massive amounts of money are being poured into AI
    • The commercial world now driving AI development, instead of universities and government agencies
    • Many of the smartest scientists and engineers in the world being drawn to work in AI
  • AI is being used to accelerate AI development

The effect of exponential growth

  • AI is on an exponential growth curve
  • We tend to underestimate future magnitudes in conditions of exponential growth

AI is a meta technology

  • AI provides powerful capabilities for bootstrapping itself to create ever-more powerful AI
  • AI is a technology that can be used to develop other technologies
    • For example, biological technologies
  • Advances in AI will lead to advances in almost every other area of science and technology

Evolution

  • Intelligence appears to be a functional property of complex systems in general
  • Evolution is a search process that finds such functions
  • Evolution is actually quite simple
    • It arises any time there are replicators, variations among the replicants, and selection processes that act on them
  • Machine intelligences will be subject to evolution, just like biological intelligences
  • Machine intelligences resulting from evolution are likely to fill a wide variety of niches

Biological artificial intelligence

  • AI is not going to be confined just to electronic systems
  • Synthetic biology is expected to be used to grow complex neural circuitry using genetically engineered cells
  • Will be used for functions where biological capabilities are better than silicon ones

The Development of AI

Traditional subfields of AI technology

  • Perception, particularly vision and hearing
  • Speech interaction
  • Natural language understanding
  • Pattern recognition
  • Learning
  • Problem solving and games
  • Planning
  • Expert systems
  • System control
  • Robotics

Early mileposts

  • Meaning of AI has changed significantly over time
  • Term “artificial intelligence” first used in 1956
  • First robot with significant self-reasoning: 1969
  • Early on: focus on duplicating human reasoning
    • This turned out to be surprisingly hard
  • Initially, AI developers overpromised and under-delivered
  • AI went through several periods of being out of favor
    • So-called “AI Winters”
  • With more computing power, AI began making bigger strides
  • IBM’s “Deep Blue”: world chess champion in 1997

Some more recent mileposts

  • So-called Deep Learning: a significant breakthrough, beginning in 2005
    • Big advance in AI capabilities and performance
  • Siri on the Apple iPhone in 2010
  • IBM’s “Watson”: top Jeopardy player in 2011
  • Facebook’s DeepFace: near-human face recognition accuracy in 2014
  • Google’s AlphaGo: world go champion in 2016
  • Self-driving cars that actually work reasonably well (in benign conditions) emerging now

Capabilities get redefined as not AI

  • The target for computer “intelligence” shifts as we acclimate to the latest ability
    • We become harder and harder to impress
    • The goalposts for AI keep getting moved
  • When some computational capability that was previously complex and difficult becomes routinized, it is no longer regarded as AI
    • Example is checkers-playing programs
  • If something can be done by executing a fixed set of  algorithms, it isn’t considered to be AI now

There have been two main approaches to AI

Approach 1: “Let’s copy how humans think”

  • Logic-based (e.g., so-called Expert Systems)
  • Knowledge engineering, knowledge representation
  • Algorithms to do logical reasoning
  • All processing crafted by a human programmer
  • AI’s reasoning is understandable by humans

Approach 2: “Let’s do some really fast statistics-based computing” (e.g., so-called Deep Learning Systems)

  • Neural net-based search over huge data bases
  • Architectures modeled on neural connections
  • Self-adaptive behavior and automated learning
  • “Black box”,  AI reasoning is opaque to humans

Expert systems

  • Expert systems endeavor to encode human knowledge-based expertise in machine-executable form
  • Tend to be based on sets of facts and rules
    • “If A and B are true, under the conditions of C, D, and E, then perform operation F, which will result in G”
  • Perform well for some problems
  • Interviewing the human experts and engineering the  knowledge base is very human labor intensive
  • Much of human knowledge isn’t easily expressed as facts and rules; many intuitive aspects are involved
  • Expert systems seriously overpromised in the 1980s

Changing AI strategies

  • AI techniques have changed over time
    • Chess was initially conquered by analyzing more moves
    • Jeopardy was won by storing more facts
    • Natural language translation was accomplished by accumulating more examples
  • Now, statistics-based machine learning techniques are ascendant
  • Even this could be a temporary phase, and something new will emerge to be the primary focus of AI

Five methodologies for machine learning

  • Evolutionary algorithms (so-called genetic programming), mimicking how natural selection works
  • Automated generation and testing of hypotheses, using the scientific method
  • Reasoning from evidence, e.g., using Bayes’s Theorem
  • Analogy-based systems, finding similar cases in memory
  • Artificial neural net systems

Blending approaches

  • To make further progress, it is likely that multiple approaches to machine learning will need to be integrated together
  • Need to use knowledge-based approaches where explicit knowledge is essential
  • May be necessary to give machine intelligences a range of “instincts”, just like biological intelligences
  • Approach integration is a major current research topic

Neural Net Machine Learning Systems

A new paradigm

  • Recent breakthroughs in AI have involved turning away from what we thought we understood about human thought processes and logic
  • Using the data mining powers of powerful computers to discern patterns without understanding what they are doing
  • Performing statistical inference for classification and decision making
  • Result: systems that are very capable--but are basically clueless about what they are actually doing

Neural net AI architecture

  • Neural net systems use an architecture originally developed more than 30 years ago
  • Loosely based on abstracted models of systems of neurons in biological brains
  • Layers of “neurons” map from an input signal to increasingly higher-level descriptions of the meaning in the signal
    • For example, words and sentences in an audio signal or objects in an image
  • New versions work with many neuron layers deep, hence the term “deep learning”

Neural net systems

  • Neural nets connect inputs to outputs through multiple so-called hidden layers with variable connection strengths
  • Each progressive layer is at a higher level of abstraction
  • They converge on an answer over many iterations

How neural nets function (simplified)

Development of neural net systems

  • When neural nets first emerged, computer power was only sufficient for a few hundred model neurons, with only one “hidden layer” between the input and output
  • Organic brains have billions of neurons in cortical hierarchies at least 10 layers deep
  • Now there has been a million times improvement in computer power
  • Neural networks today are scaled up to 12 or more layers deep, with billions of connections
  • The current Internet provides huge data sets on which to do training of neural nets

Typical operation of a neural net 

  • For example, a neural net maps an input image to the probability that your face is in that image
  • For training the net, the system is given images both with and without your face in them
    • The images are labeled: face yes or face no
  • Iterative processes used to cluster like features with like
  • The net learns the pattern of your face as it sweeps back and forth over thousands or millions of iterations
    • Changes connection parameters in the hidden layers
    • Performance is improved by more layers and more neurons in the layers

The expectation maximization (EM) algorithm

  • Deep learning systems actually implement the standard algorithm of modern statistics
  • EM is a two-step iterative scheme for climbing a hill of probability
  • EM doesn’t always get to the global maximum, but almost always gets to the local maximum

Supervised vs. unsupervised learning

  • In supervised learning, the system is provided examples that are human-labeled
  • In unsupervised learning, the system works with unlabeled examples and converges on labels by itself
  • The supervised learning algorithm is called backpropagation
    • An error is computed in the output and distributed backwards through the neural networks layers to refine the training of the connections
    • When the connections produce the least output error, the system has learnedDeep learning systems actually implement the standard algorithm of modern statistics

Reinforcement learning

  • Reinforcement Learning is a machine learning technique that enables an agent to learn in an interactive environment
    • By trial and error using feedback from its own actions and experiences
    • Feedback includes both positive and negative rewards to reinforce good performance
    • Maximize the total cumulative reward of the agent
  • With no prior input except the rules of chess, Google’s AlphaZero learned in four hours how to beat the best previous AI chess program or any human chess master

Generative adversarial networks

  • Generative adversarial networks (GANs) pit two neural networks against each other in a zero-sum competition
  • One network generates candidates, while the other network evaluates them
  • The generator tries to create plausible fakes, while the discriminator tries to identify the fakes from real ones
  • They are used to synthesize content that is spookily realistic
    • Images, speech, music, full video, etc.
  • An illustrative video:

http://www.bbc.com/future/gallery/20181115-a-guide-to-how-artificial-intelligence-is-changing-the-world

These forms of learning are not what people do

  • Note that these neural net machine learning algorithms are not what goes on in biological neural systems
  • The current machine neural net architecture is an extremely crude model of how learning is done naturally
  • For example, the human brain is constantly predicting what will come next and refining its model based on what it actually experiences
  • AI systems are gradually being improved to incorporate more organic brain-like featuresGenerative adversarial networks (GANs) pit two neural networks against each other in a zero-sum competition

Impacts of deep learning

  • These systems outperform the best algorithmic and knowledge-based approaches in many fields
    • Revolutionized speech recognition, natural language processing, machine translation, object recognition and image understanding systems
    • One major current application is in recommender systems
  • Deep learning systems continue to improve with use  Note that these neural net machine learning algorithms are not what goes on in biological neural systems

Applicability of neural networks

  • Neural net systems are most successful for problems where large amounts of relevant data are available, relative to what needs to be learned
  • Not so good when there are not a lot of data or each data point is quite complex
  • Or when you have to reason from limited evidence
    • Then you need to employ knowledge engineering
  • Neural nets are not a panacea

Types of tasks best suited for machine learning

  • A function that maps well-defined inputs to well-defined outputs
  • Large digital data sets available with input-output pairs
  • Task has clear feedback with clearly-defined goals and metrics
  • No long chains of logic or reasoning that depend on common sense or diverse background knowledge
  • No need for detailed explanation of the decisions
  • Tolerance for error, no need for provably correct or optimal solutions
  • The phenomenon being learned doesn’t change rapidly

Neural net system biases

  • Neural net systems are highly vulnerable to biases in the data used for their training
  • For example, face recognition accuracy differs between systems developed in different areas
  • U.S. and European systems recognize white faces more reliably than black faces
  • Chinese systems recognize Asian faces more reliably than white faces
  • AI systems used to advise on bail, sentencing, and parole show racial bias, stemming from their input training data

Neural net glitches

  • The ability to capture the patterns appearing in data has a risk of finding patterns that aren’t there
    • Deep learning systems occasionally confidently declare patterns in random noise
  • And sometimes they produce rather bizarre results
    • They can recognize different dog breeds in images better than you can, but can mistake an image of a blueberry muffin for an image of the face of a Chihuahua

Some things to beware of

  • A neural net system learning on its own how to cheat so as to obtain a specified result with less effort
  • A reinforcement learning system figuring out how to hack its own reward function

Resources Used by Contemporary AI

Advanced computing hardware

  • AI has taken full advantage of tremendous increases in computational power and speed at plummeting costs
    • Advanced processing hardware (e.g., multicore chips, graphical processing units originally developed for video games)
    • Now chips specifically designed for neural nets
    • Networked computational centers
    • Prospectively, quantum processors used for AI
    • Some researchers are looking into biologically-based processing units for certain functions

Specialized processing for AI

  • Google, for example, announced that it had created a microchip system called a Tensor Processing Unit (TPU)
    • As of early 2018, Google’s TPUs were capable of 180 trillion floating point operations per second
  • In 2017, the fastest computer in the world uses roughly 40,000 processors with 260 cores each. That’s more than 10 million processing cores running in parallel

Networking

  • The Internet has allowed billions of processing systems to connect and interact
    • In real time, with increasing bandwidth and speed and decreasing latency
    • Distributed over the whole planet
    • With vastly different device characteristics
  • The so-called Internet of Things (IoT) is going to connect exponentially more things together
  • There is very little management of this process, unfortunately

Server farms

  • The heavy lifting of AI systems is generally done remotely in huge computational centers, called server farms
    • Networked server farms constitute what is commonly referred to as “the cloud”
  • These are massive users of electrical power
    • Located where power is cheap, cooling is available, and taxes are favorable
    • Utilities are beginning to worry they will destabilize their grid systems because of their huge power draw

Data sets

  • Mind-boggling amounts of data are now accessible over the World Wide Web
  • Virtually all the world’s electronically-encoded knowledge is online
    • Who has access to what varies, of course
  • Machine learning systems can interact with huge data sets
    • E.g., billions of images
    • Millions of hours of recorded speech in hundreds of languages
    • Vast quantities of text translations

How AI is already surpassing human capabilities

  • AI systems beat top human experts at almost all board and video games
  • AI systems respond to complex conditions in milliseconds
    • E.g., managing electric power grids
  • AI systems have taken over securities trading
  • AI systems detect and analyze patterns in huge data sets

Thinking Machines: What They Can and Can’t Do

What they do particularly well

  • Handle complex inputs, such as data from many sensors simultaneously
  • Store and retrieve vast amounts of information without loss or degradation
  • Perform complex logical and mathematical operations very rapidly without error
  • Do complex statistical analyses
  • Handle large volumes of repetitive tasks accurately and reliably
  • Function tirelessly without a break
  • Readily accommodate updating

What we should have them do

  • Computers excel at doing things most of us fumble at
  • They can be extremely good at aspects we’re not
    • Speed
    • Accuracy
    • Focus
    • Alertness
    • Awareness
    • Reliability
    • Computing on massive amounts of data
    • Handling many things simultaneously
    • Remembering everything they learn
  • Machines should be given those tasks they do much better than we do

What’s easy for AI and what’s hard

  • Easy:
    • Do complex logical reasoning in a fully-specified domain
    • Search in huge volumes of data
    • Sort
    • Find patterns
    • Evaluate likelihoods
    • Create plans
  • Hard:
    • Take context and implicit knowledge into account
    • Use “common sense”
    • Understand human viewpoints, motivations, etc.

“Common sense”

  • Human thinking and action takes advantage of a vast amount of tacit knowledge
  • Much of it is regarded as “just common sense”
    • Not written down
    • Often hard to express
    • Not explicitly taught
  • Humans have an extensive intuitive model of how the world works
  • Without that body of tacit knowledge, machine intelligences are likely to do dumb things

Subtle aspects of human interaction

  • Human interaction involves many subtle aspects—e.g.
    • Humor, irony, sarcasm
    • Empathy, compassion, reassurance
    • Many others, e.g. metaphor, analogy, allusion
  • Humans have Theory of Mind capabilities
    • Allow them to understand how the other person is likely to be feeling in the situation
    • Humans think about how others will be affected by what they say
  • These are still major research challenges for AI systems for interacting with humans—but progress is being made

Creativity and innovation

  • So far, AI hasn’t made a lot of progress in domains associated with human creativity and innovation
  • AI is best at imitating human thought processes whose outcomes are fixed—e.g., playing board games
  • Doing deep abstraction and idealization, and changing assumptions, processes that underlie human creativity are still cutting-edge research areas
  • A related capability is discovering causal models—figuring out why things are the way they are
  • The ability to generalize correctly is hard for AI

General strengths of human thinking

  • Human thinking is strongest in its integrative aspects
    • Combining intuition, emotion, empathy, experience, and cultural background
    • Asking a meaningful question and drawing a conclusion by combining seemingly unrelated facts and principles
  • Want to leverage these human thinking abilities

Intelligent System Architectures: Comparing with the Human Model

Brains

  • The neural architecture of the human brain is vastly more complicated than any man-made simulation of it
  • Assuming that we can duplicate the detailed functioning of the brain is naïve
  • Note that the brain is only partly devoted to thinking
    • A large portion of brain activity is involved with managing the functions of a highly complex body
    • We actually have about 100 sub-brains
  • In reality, most of the time humans don’t think, at least in terms of doing conscious reasoning
    • We mostly get by on autopilot

Biological neural architecture

  • Brains use a highly parallel architecture and have many noisy analog units (neurons) firing simultaneously
  • Individual neural computations are relatively slow compared with digital systems
  • The brain’s enormous degree of parallelism makes up for that
  • Each of the ~1011 (one hundred billion) human neurons has on average 7,000 synaptic connections to other neurons

Some comparisons of brains

  • Human brains use about one-thousandth as much energy as a current machine intelligence to do a task such as recognize a face
  • Organic brains have a global workspace so that any module can access information in any other module of the brain; machines don’t
  • Biological neural system architectures intimately integrate memory and processing
    • Machine systems separate the two
    • Researchers are only now considering how to integrate them in machine systems

The human mind starts with inherent abilities

  • Humans come out of the womb with a great variety of inborn mental capabilities
    • Primed to respond to early experiences
  • For example, humans have remarkable innate capabilities to learn language as children
    • This has been fine-tuned by evolutionary selection over many thousands of human generations
  • There are many other inborn capabilities, some of which we are only beginning to appreciate, e.g. intuitive physics
  • Much work to model these before they can be implemented in machine intelligences

Functional modules of intelligence

  • We are coming to realize that human intelligence involves many different competences and capabilities
    • Often somewhat independent of one another
    • Often corresponding to different functional modules in the brain
  • For example, object recognition (“what is there”) vs. face recognition (“who is there”) involve rather distinct parts of the visual cortex

Intelligence: a collection of really good hacks

  • Scientists have concluded there is no general algorithm, just waiting to be discovered, that underlies intelligence in general
  • Intelligence appears to be the result of a large collection of diverse approaches to different types of problems
    • Effectively, humans have developed a set of really good hacks through our evolutionary path
    • None of these capabilities were designed
    • Exaptation: the adaptation of a previously-developed structure for a newly-useful function

Combining intelligence modules

  • More general AI will be created by combining qualitatively different programs to form an ever-greater cognitive diversity
    • Effectively, bundling multiple idiot savant-like AIs together in a complementary fashion
    • Critical that the architecture combines the savants, not the idiots!
  • Must identify the set of problems for which activating a particular set of capabilities makes you better off, not worse

Understanding the human brain/mind

  • The truth is that we don’t understand much about the human brain and mind
    • Vast areas that are still a mystery, such as just how memories are encoded (short- to long-term)
    • No major breakthroughs have happened yet in understanding human cognition
  • We need to parallel work on artificial intelligence with research on biological intelligence
  • Can’t duplicate in machine systems what we don’t understand in organic systems

Thinking is expensive in resources

  • Thinking is basically costly, whether biological or electronic
  • The human brain represents only about 2% of adult body weight, yet uses about 20% of the oxygen and 50% of the body’s glucose
  • The human brain could evolve to its large size, compared with our primate ancestors, as a result of the more efficient digestion of food enabled by cooking

How Far Is AI Likely to Go?  Over What Time Frame?

Different kinds of intelligence

  • We don’t have a good general taxonomy of intelligence
    • What are all the different possible forms, and how are they related to each other?
  • Our implicit model of intelligence is human intelligence
    • Our intelligence is a consequence of the developmental path of how we got to be humans
  • AIs don’t have the constraints that biological systems do
  • As a result, artificial intelligences might go in some very different directions than we currently expect

Predicting AI progress very far ahead is hard

  • The rate of development of AIs is likely to be exponential for some extended time into the future
  •  Increasingly, AI will be used to develop AI
  • The efforts invested in AI development by institutions such as corporations and governments will be a function of the potential payoffs to them

So: Is AI Likely to Develop General Human-Like Intelligence?

Artificial general Intelligence (AGI)

  • A common question is whether AIs that duplicate the full set of thinking capabilities of a human being are possible
  • The implicit picture: a humanoid robot that can act in a manner fully equivalent to a biological human
    • Each AI as an independent entity, operating on its own
  • Probably not the path that advanced AI systems will take
  • Artificial general intelligence will likely be more like Wikipedia—a vast virtual capability connected online
    • Not located at any physical spot
    • Sourced by many different entities

General AI capabilities from components

  • We’re not likely to create Artificial General Intelligence(s) directly
  • Instead, many different specialized AI capabilities will be integrated into progressively-more-comprehensive systems
    • Separately developed and then assembled together
    • Interfacing considerations will be very important for successfully integrating these components

The lens of science fiction

  • A lot of how we think about AI and robotics comes from our exposure to science fiction
    • Particularly the movies and television
  • It’s important to emphasize the word fiction here
    • The more dramatic, the more engaging the story
  • What makes for a good yarn, rather than what projects what will really happen
  • We need to set aside some of our preconceptions derived from SF

AI in the movies

  • 2001: A Space Odyssey
  • iRobot
  • The Terminator series
  • The Star Wars series
  • Her
  • The Matrix series
  • The Star Trek series
  • RoboCop
  • AI: Artificial Intelligence
  • Blade Runner
  • Ex Machina
  • Interstellar

Another science fiction view of AI

  • Frank Herbert’s Dune stories (ca.1965) have a different take on AI:  The Butlerian Jihad
  • The Jihad’s fundamental commandment: “Thou shalt not make a machine in the likeness of a human mind”
    • Development or possession of any form of machine intelligence was punished by death
    • The focus was instead shifted to the development of advanced human intelligence capabilities

AI dystopias

  • Dystopic visions of machine intelligence project an alpha-male psychology onto the concept of intelligence
    • Assume goals such as taking over the world
    • Such goals are not intrinsic to intelligence itself
  • Don’t want such a view to become a self-fulfilling prophesy

Societal Factors and AI Development

AI and societal shaping

  • Authoritarian countries like China are using advanced information systems to shape their citizens’ political behavior and “guide” the national consensus
  • Political parties in the U.S. are increasingly doing the same thing
  • Intelligence agencies (the NSA, CIA, etc.) are using advanced AI capabilities in secret programs
  • Companies are using it to shape consumer behavior to increase profits
  • Concern: small groups of insiders gaining the ability to control the thoughts and behaviors of everyone else, without their awareness of being controlled

How does society choose what it wants?

  • We don’t currently have mechanisms for choosing the kind of future we want
    • Certainly not at the humanity-wide level
  • Decisions are made by default, on the basis of the interests of individual groups
  • Society as a whole doesn’t have much say
  • Now we are faced with issues with unprecedented scale and consequence due to new technologies like AI

Who will employ AIs?

  • Those who will be first to employ powerful new AI capabilities are likely to be existing human elites and the institutions they control
    • They will use AI to further concentrate power and solidify advantages
    • The weaponization of AI
  • Tools will be used for whatever benign or malign objectives institutions already have
  • Danger comes more from who will use AIs and for what ends than from the systems themselves

The digital divide

  • Already we speak of the divide between people who have access to contemporary information systems and those who don’t
    • People who don’t have computers, Internet access, smartphones, etc. are increasingly disenfranchised
  • Machine intelligence systems are likely to make this digital divide worse
  • Increasing the inequality between different social groups and between different nations

What masters will the AIs serve?

  • If robots and AI systems are going to be doing most of our productive work, who will own them?
    • Will it be primarily the biggest, best-funded, most powerful corporations, who can invest the most?
    • Alternatively, will they be owned by governments in non-capitalist societies?
  • Or can we distribute ownership widely and equitably?
    • Ideally, into the hands of a large and diverse cross-section of the population
  • Goal: avoid the ever-increasing concentration of economic and political power

Corporations and AI

  • Corporations can be regarded as non-human intelligent entities with agency and self-interest
  • Unfortunately, they tend to behave as sociopaths
    • Focus single-mindedly on the maximization of profit and the return on invested capital
    • Internalize benefits, externalize costs and detriments
    • Take societal good into account only when forced to
  • AI systems will be given more and more decision-making power in corporations, including strategic decisions
  • The most AI-empowered corporations will have the greatest competitive advantage
    • Risk that AI will augment corporations’ monopolistic capabilities
  • One definition of Fascism: where corporations and governments implicitly merge and the coercive power of the state is employed for the advantage of corporate and oligarchic interests

The distribution of AI impacts

  • How will the benefits of machine intelligences be distributed?
    • Will they be primarily in the rich nations of the developed world?
    • Or can they be shared worldwide?
  • How will the harms and costs be distributed, too?
  • In general, how do we democratize the benefits of AI systems and minimize their detriments?

Wealth and illth

  • We have a general notion of “wealth”—goods and services that are beneficial
  • However, we tend not to pay attention to the creation of negative factors associated with the creation of wealth
    • All manner of so-called externalities
  • These negatives can be considered to be “illth”
    • Things like pollution, unemployment, social dysfunctions, etc.
  • To what degree will the widespread adoption of AI cause increased illth?

Compelling reasons for using AI

  • Via the Internet, capabilities provided by an AI system can be location- and time-independent
    • Available anywhere worldwide, 24/7/365
  • Today, most costs are associated with human labor hours.  AI systems don’t have labor costs
  • Machine learning systems get better and better over time as they are used
  • In any competitive situation, smarter tends to win
    • Better smarts can overcome most other types of advantages

AI becoming a requirement

  • As powerful AI systems proliferate, every large entity will find it necessary to create and use them in order to stay competitive and viable
    • Every corporation, every government, and every large institution of every kind
  • AI capabilities will exacerbate tendencies already present in the entities that use them

Arms races

  • Arms races create tremendous pressure for rapid system development and deployment
  • Arms races can occur in many environments, not just military situations
  • Anywhere groups vie with each other for technologically-empowered supremacy and losing the race threatens survival

Location distribution of intelligence

  • Human intelligence is located within individual human brains
  • Interaction between human brains has bandwidth constraints imposed by human communication channels—speech, writing, etc.
  • Machine intelligence, on the other hand, can be highly distributed
    • The Internet allows machine intelligence processing to be in the cloud, with very high bandwidth communications between physical locations
    • Can be present anywhere on the planet

AI systems scale

  • Unlike biological systems, technological systems scale
    • No intrinsic limit to size (big or small)
  • For a given function, artificial minds will be faster, more accurate, more aware, and more comprehensive than their human counterparts
  • AI systems’ power will only increase with time

AI and the Political-Economic Environment

The transition in AI development

  • Early on, most R&D on AI was performed by universities and government-funded labs
  • For example, major sources of AI funding in the U.S. were agencies like DARPA and NASA
  • Early AI applications were specialized and had relatively limited impact
  • Technical breakthroughs like Deep Learning changed that picture completely
  • Now, the largest corporations are pouring huge amounts of attention and money into AI
    • Profit motives are the driver now

Investment in AI is exploding

  • Seen as essential to staying competitive
  • Retailers everywhere: identify customer preferences, make recommendations
    • Amazon, Walmart, all online retailers
  • Every company producing information technology hardware, software, and services
    • IBM, Apple, H-P, Intel, etc.
    • Google, Microsoft, Facebook, Twitter, etc.
    • Alibaba, Baidu, Tencent, Huawei, etc. in China
  • Every major car manufacturer
    • Driver assistance now, ultimately cars that drive themselves
  • Financial firms: banks, investment firms, insurance companies, real estate firms, etc.
  • Manufacturers of all types
  • Transportation companies: airlines, shipping companies, etc.
  • Entertainment and media companies of all types
  • Utilities
  • Health care firms
  • Many others
  • The older sources of AI development funding are taking advantage of all the commercial advances in AI, too
    • Militaries
    • Intelligence agencies
    • These organizations have large resources and compelling motivations for using AI
  • Authoritarian governments are seeing AI as a major enabler for social control

The international race for AI competitiveness

  • There are numerous national programs investing heavily in the development of AI and robotics
    • Seen as a key national competitive advantage
  • Well-funded programs in China, the U.S., Russia, Japan, South Korea, the EU Commission, the U.K., France, Germany, Italy, Sweden, Finland, Israel, Canada, Australia, India, Taiwan, Singapore, and the UAE
  • Even smaller technology centers are emphasizing investment in AI—e.g., South Africa, New Zealand, Brazil, Poland, Mexico, Kenya, Malaysia, Tunisia

The Chinese push for AI supremacy

  • Particularly relevant is the Chinese national commitment to dominate in AI capabilities in the 21st century
  • China intends to spend at least $150 billion to be the world’s leading AI powerhouse by 2030
  • Already has ~40% of the world’s trained AI experts and most large universities in China have AI programs
  • China openly collects and analyzes data from its 750+ million daily Internet users
    • Population is not particularly concerned with privacy
  • Chinese AI researchers are proficient in English and exploit Western AI research immediately
  • Explicit government policy to acquire key foreign technologies by all means available

Chinese initiatives in robotics

  • China has made an explicit commitment to become the world’s dominant maker and user of robots
  • Already, large numbers of low-paid Chinese manufacturing workers are being displaced by robots

AI Development Factors

The motivations of AI developers

  • The motivations of those who develop AI and robotic systems are critical factors
    • Motivations include commercial profit, competitive advantage, military superiority, intelligence service advantage, and societal control, in addition to developer prestige
    • Self-interested motivations compete with motivations of overall long-term benefit to humanity
  • How will these motivations balance out?

The shortage of AI experts

  • The number of expert teachers of AI and robotics is very limited, while the demand for this training is exploding
  • At the same time, companies are desperate for AI and robotics experts and are willing to pay extraordinary amounts to them
    • New AI PhDs are getting $300,000/year and up
    • Top-name researchers are getting multiple millions in salaries and stock options
  • The result is that university departments and research institutes are losing their best people
  • Who will teach the next generation of AI and robotics experts?
  • In the entire world, it is estimated that fewer than 10,000 people currently have the knowledge and skills necessary to do serious artificial intelligence research
  • Out of this number, it is estimated that only about 50 people are working full-time on safety  of AI issues

The migration of AI into everything

  • We can expect aspects of AI to be incorporated into most all of our devices that have any electronics
  • We can also expect many of these devices to be connected together through the Internet of Things
  • And we can expect aspects of AI to be involved in most of our social interactions
    • How we get our news and entertainment
    • Who we interact and network with (e.g., dating apps)

AI in your car

  • Cars are incorporating large numbers of microprocessors and microcontrollers to support virtually every active function
    • A modern car can have upwards of 150 separate subsystems, each with some degree of electronic control
  • Many of these devices and systems will have at least some degree of AI functionality
  • Cars will be highly networked with each other and with the roadside infrastructure (for traffic control, etc.)
    • Also with the manufacturer (maintenance monitoring, software updates, etc.)

AI enabled devices in your home

  • Entertainment and communication systems: television, radio, audio equipment, media recorders and players, video games, toys, telephones, etc.
  • Kitchen systems: refrigerator, microwave, range, etc.
  • Heating, ventilation, and air conditioning (HVAC) systems
  • Cleaning equipment (washer, dryer, vacuum, etc.)
  • Home security systems, baby monitors, etc.
  • Bathroom systems (e.g., toilets that perform health monitoring)
  • Many others

Smart control

  • Control provided through AI-based voice interaction
  • For example, Amazon’s Alexa already works with more than 20,000 different smart-home devices, representing more than 3,500 brands

AI and Privacy

Profiling you

  • With every interaction your have with the Internet (browsing, shopping, posting on Facebook, etc.) a detailed dossier of you is being built up—your digital double
    • In your home, Alexa is always listening
  • This digital double knows your individual characteristics, wants, needs, preferences, habits, weaknesses, etc.
  • Increasingly sophisticated AI systems are analyzing your digital double to determine how to interact with you
    • How to sell things to you
    • What media to present to you (news, music, etc.)
    • Whether to employ you, recommend you a date, etc.
    • How to influence your vote

AI-based biometric recognition

  • Deep learning systems are very capable in identifying individuals on the basis of biometric signatures
    • Face
    • Voice
    • Fingerprints
    • Iris patterns
    • Gait
    • DNA
    • Others
  • Consequence: no more anonymity, wherever you go

Tracking and analyzing you

  • Increasingly, it is possible to track you most of the time
    • Where you are
    • Who you are with
    • What you are doing
  • There are growing capabilities to gauge your emotional state, cognitive state, mental and physical health, etc.
    • Facial expressions, body language, gestures
    • Voice features
    • Breath chemistry, facial temperature profile
  • Concerns over who has access to this information and what they want to do with it

How visible do we want to be?

  • What is our tradeoff between greater convenience and the loss of privacy and anonymity?
  • If The Authorities know everything about us and what we are doing, how much freedom will we have?
  • Are we inadvertently creating Orwell’s 1984, just a little later than he predicted?
  • Many governments around the world are vigorously pursuing AI-based surveillance of their populations to detect and neutralize dissent and protests

Personal transparency

  • Potentially our private lives will become transparent
  • Won’t be able to live different lives in different contexts
    • No more closets, of any kind
  • Remember the direst threat when you were in school?  “This is going on your Permanent Record!”
    • It didn’t really exist then, but it will now
  • Every indiscretion, every odd taste, everything not fully socially approved, may be open for observation by others
    • The government, your employer, your insurance company, your pastor, your potential dates,  etc., etc.
  • You will have little control over who can see what about you

What Can’t AI Do (Now)?

Some current limitations

  • AI systems based on statistical inference don’t actually understand the meaning of what they are dealing with
  • AI systems generally don’t take context and background into account
  • Current AI systems generally don’t understand nuance in human communications—e.g., tone of voice, sarcasm, humor, rhetorical questions, etc.
  • Unlike expert systems, statistically-based AI systems generally cannot explain the basis for their decisions
    • They are effectively black boxes
  • AI systems lack tacit knowledge and common sense

Subtle language issues

  • Human language has lots of subtle aspects that humans learn easily but are big challenges to machine systems
  • Humans take advantage of a great deal of shared cultural and contextual knowledge when conversing
    • E.g., allusions
  • Humans understand ambiguous referents easily while machines struggle with them
    • “It’s going to be cold tonight.”  What do you mean by “it”?

How some current limitations are being addressed

  • Extensive work is going on to represent as much of human “common sense” as possible for machine use
    • AI researcher Doug Lenat’s Cyc program has been working hard on this task since 1984
  • A new field of affective computing is under development so machines can interact with humans on an emotional level

Affective computing

  • Developing systems and devices that can recognize, interpret, process, and simulate human affects
    • The ability to read and interpret the emotional state of humans
    • Recognize human facial expressions, body language, gestures, vocal aspects, other indicators of affect
    • Respond with finely-tuned human-like emotion
  • Goal: to have the human feel that they are interacting with a virtual person, that truly gets them at a deep level
    • Danger: heightened power of the AI to manipulate

Affects with characteristic indications

Different affects have characteristic physical indications (facial expressions, voice features, etc.)

  • Anger
  • Disgust
  • Fear
  • Happiness
  • Sadness
  • Surprise
  • Amusement
  • Contempt
  • Contentment
  • Embarrassment
  • Excitement
  • Guilt
  • Shame
  • Pride in achievement
  • Relief
  • Satisfaction
  • Sensory pleasure
  • Confusion
  • Deception
  • Anxiety

Also key: intensity of emotion, mix of emotions

The next step after Siri

  • Current virtual personal assistant systems, such as Siri, interact via two-way voice audio
  • Now being developed: ultra-realistic avatars that interact via two-way video with affective computing
  • The avatar can be highly customized to interact with a particular person
    • Sex, age, race, ethnicity, language, social class/ education, sexual orientation, personality, etc.
  • Intent: you will feel like you are interacting with a virtual human
    • Tuned to be optimally compatible with you

Considerations for Employing AI

AI system competence

  • Very important not to cede authority to a machine system beyond its competence
    • How can we tell when the machine has left its comfort zone and is operating on parts of the problem it’s not good at?
  • Don’t put machines in charge of decisions they don’t have the intelligence to make
  • Want AI systems to know their limits

Lack of self-reflection/introspection

  • Current AI systems are unable to question their own actions
  • Don’t appreciate the consequences of their design and programming
  • Don’t generally understand the context in which they operate
  • GIGO (“garbage in, garbage out”) applies in spades for AI
  • In many cases, you want to have an explanation for why a decision was made—what were the reasons?
    • AI systems aren’t good at this yet

Shallowness of thinking

  • For the most part, AI systems’ thinking is shallow
  • Mostly mimic human thinking, rather than simulating it or actually understanding it
  • Don’t have the ability to draw conclusions from a deep understanding of what they are working on
    • Generally don’t get the underlying meaning
  • For example, can’t read a textbook and then answer the quiz questions in the back of the book

System integrity

  • How can you be sure that the system doesn’t contain bugs or has been compromised?
    • By being network connected, AI systems are vulnerable to incursions of various kinds
  • How to gauge what parts of the system have been affected by an intrusion?
    • And the consequent ripple effects on the rest of the system

Failure modes

  • AI systems, like every other type of man-made system, have failure modes
  • It is important to understand and take into account possible failure modes whenever an AI system is used
  • Many AI systems are brittle and don’t fail gracefully

How AI Systems Will Be Connected and Collaborate With Each Other

The networking of intelligence

  • What transformed human intelligence was the connecting up of human brains into networks, enabled by the division of labor
  • Allowed individuals to dive deep into a specialized area and share their expertise with others
  • A process of ever-gathering speed with the Internet
  • Similarly, machine intelligence will be most powerful in the network of machine capabilities, not within individual machines

“Telepathy” between AI systems

  • Unlike humans, machine intelligences can easily exchange complete memories and thinking states from one to another
  • Any distinction between individual machine intelligences then blurs
  • They effectively become a hive mind

Networks of networks

  • In general, AI systems aren’t going to operate as isolated entities
  • The niche expertise of a particular AI system will be complemented with additional niche expertise areas of other AI systems through network connections
  • There will likely be hierarchies of AI networks to aggregate capabilities and fill gaps
  • How this will be managed is an open question at this time

Standards

  • Standards have been very important in the development of modern information systems
    • E.g. everyone uses the same Internet and World Wide Web standards
    • Every system can interact with every other system in a well-defined way
  • Machine intelligence systems are being developed without such standards
    • Result: Tower of Babel effect
    • Connecting machine intelligence systems together can have unpredictable results

System interoperability

  • When systems are connected together, you want to make sure they are interoperable
  • Each system needs to be able to understand the other
  • Syntactic interoperability requires they use the same language and syntax for communicating with each other
  • Semantic interoperability requires that a term means the same thing to each system, including all the underlying assumptions
  • When systems that were not originally designed for it are connected to each other, syntactic and semantic mismatches are virtually inevitable

Open vs. closed systems

  • Will the networks connecting AI systems be open or closed?
  • Will any innovator from anywhere be able to plug something new into a network and interact without needing permission?
  • Or will they be controlled closed systems, with authorities deciding who and what is allowed to connect in, and how?
    • If so, who supervises the entities that do the deciding?

Constant change

  • AI systems are in the process of rapid, constant change
    • Hardware and software platforms, techniques, algorithms, code bases, data sets…everything
  • Essentially, no element remains stable for very long
  • Analogous to flying on an airplane that is being continually redesigned and rebuilt in mid-air
  • Danger of breaking something in mid-process

Intellectual property and AI

  • Much of AI technology is being regarded as intellectual property, restricted by the owners
    • Algorithms, specialized hardware, code, data sets, etc.
    • Intent is to maximize competitive advantage
  • This limits the ability to connect systems together and understand the interactions—can’t see critical aspects

AI system cooperation

  • Future AI systems and robots
    • Not going to function in isolation
    • Will be richly interconnected via the Internet
  • Access to all the world’s information
  • Capabilities of any particular AI system augmented by capabilities of other systems connected to it
    • Hundreds of different specialized AI systems
  • An individual robot is going to have lots and lots of buddies

AI system competition

Simultaneously, AI systems will be in competition

  • Systems of competing businesses (e.g., competing financial traders)
  • Systems of adversarial governments
  • Law enforcement AI systems competing with lawbreaker AI systems

Result: AI arms races

Diversity of AIs

  • The capability to attack difficult problems will come from creating a vast diversity of AIs
  • Multiple perspectives and a rich set of ideas and approaches
  • “Close AIs” will think more-or-less the way humans do
  • “Far AIs” will thing in ways we cannot fathom

A Key Issue for Machine Intelligences: Autonomy

AIs: automata or agents

  • Automata act at the behest of their creators.  If they misbehave, the creator is blamed
  • Agents act on their own agendas and are themselves to blame for misbehavior
  • Are our machine intelligences going to be endowed with agency and self interest?
    • How should they be allowed to act on self-interest?
    • Should they be able to do so without awareness of why they were acting that way?

Agency

Many systems have agency

  • They function in a purposeful manner to achieve goals
  • They have self-interest
  • They are able to make choices in response to the conditions they experience

Living organisms are obvious examples of systems with agency

  • In addition, supersystems of living organisms (e.g., ecosystems) have agency
  • Human groupings (e.g., families, tribes, cultures, etc.) have agency
  • Many human-created systems (e.g., organizations such as corporations and bureaucracies) have agency

Agents’ behavioral imperatives

Systems with agency have behavioral imperatives—e.g.

  • Maintain a membrane that differentiates inside from outside
  • Take in and process nourishment (nutrients/energy) from the external environment
  • Excrete metabolites/byproducts/wastes to the external environment
  • Maintain stability under changing conditions (homeostasis)
  • Outlive constituent sub-elements through renewal and regeneration
  • Self-protect
  • Avoid danger
  • Perform self-defense when threatened
  • Grow
  • Continue existence (including through reproduction)
  • Explore surroundings
  • Move in response to external gradients (e.g., environmental temperature, food supply distribution)
  • Compete
  • Obtain/protect/defend territory
  • Accumulate assets
  • Adapt
  • Learn from experience
  • Anticipate
  • Interact with other entities
  • Endeavor to control/constrain other entities
    • Resist being controlled/constrained by other entities
  • Maintain/expand freedom of action; avoid external constraints
  • Endeavor to obtain maximum positive outcomes for minimum expenditures of resources and effort; maximize return on investment
    • Internalize benefits (e.g., profits, privileges)
    • Externalize detriments (e.g., costs, obligations)
  • Agents commonly game the system they are in
    • How can I win?

Some Challenges

Verification and validation

  • An important part of conventional algorithmic software development is subjecting it to verification and validation
    • Verification confirms that the design conforms to its specifications
    • Validation confirms that the software is implemented correctly and is free of defects and errors
  • AI systems with learning are not amenable to verification and validation
  • They continually change with use
    • They continually change with use
    • Different examples diverge from each other

We build them, but don’t understand them

  • Increasingly, the workings of AI systems will be beyond our abilities to understand and correct
  • They will have behaviors that aren’t evident from the architecture of  the programs we wrote
  • Unfortunately, we are often in situations where we need to know why something happened
    • Why was my loan application denied?
    • What caused the self-driving car to crash?
  • Critical in order to troubleshoot a system with problems

Self-modifying systems

  • The understandability of AI systems will be particularly problematic for ones that continually modify themselves
    • We may not be able to grasp the assumptions  and processes underlying their reasoning
  • Soon, programs will write other programs that no human can understand or track
  • They won’t be able to explain their decisions in a way we comprehend
    • Can’t adopt our perspective to determine what to say that would be meaningful to us

AI systems don’t think like humans think

  • Don’t “think” in the ways that humans think
    • Mimicking human reasoning processes: too slow
    • Knowledge engineering: very human labor-intensive
  • Result: very difficult to check AI systems’ thinking
  • Hard to have confidence in AI system decisions
    • Are they ones that humans would make based on the same information?
    • Are we going to be able to trust their decisions?
    • What do we do when we think they’re wrong

Understanding meaning

  • For the most part, AI systems don’t currently understand the meaning of what they are doing
    • Able to translate Chinese text or speech into English very successfully
    • But have no idea what the material is actually about
  • But much work is going on in this area

Lack of AI system transparency

  • A big problem with any AI system that is based on statistics (i.e., neural net systems) is that it is opaque
    • You can’t query it to determine why it reasoned the way it did
    • The results just have to be accepted on faith
  • Hard to know whether or not it is operating within its expertise
  • No easy way to adjust a statistics-based AI system that is producing anomalous outputs

Different stakeholders’ needs for AI system transparency

  • Users: want to know what the system is doing and why, particularly if it is doing unexpected things
  • System certification authorities: want to be able to confirm the system meets safety standards
  • Accident investigators: want to trace the causes of an accident involving AI systems
  • Lawyers and expert witnesses: want to be able to explain their evidence to determine liability
  • The wider society: want to have confidence in the use of the technology

Machines’ theory of mind

  • Intelligent machines will soon need to understand other machines’ psychology to be able to work alongside them
  • They will need to understand human psychology too
  • We might want to worry when they start understanding human psychology all too well
    • Psychopaths are sometimes credited with having too great an understanding of others’ psychology
    • Fear machines using their understanding of humans to be more effective at manipulation

Controlling the Deployment of AI

AI “off the reservation”

  • AI technology is going to be widely available throughout the global society
    • We won’t be able to limit its spread through classification, or high cost, or regulatory controls
  • Some “bad dudes” are going to have AI capabilities
    • Political factions
    • International criminal cartels
    • Terrorists
  • And any smart teenager with a laptop and an Internet connection is going to be able to create some powerful AI capabilities

Avoiding premature release into the world

  • How do we make sure that a particular AI system is not harmful before it is released into the general world?
    • Operating on its own
    • Influencing other systems
  • Right now, there are basically no controls on this

Recursion: AI used to build even better AI

  • AI is very suited to developing faster, more powerful, more effective AI systems
  • AI systems will inevitably be used to code their own improvements
  • Such self-improvement is likely to proceed ever-faster in a positive feedback manner
    • People worry about an “intelligence explosion”
  • The degree to which humans can and will steer this process is an important issue

A world of multispecies intelligence

  • It looks like we are headed towards a world with a whole menagerie of different kinds of intelligences
  • Not just standard-issue humans and the more intelligent animals, but a great variety of different machine intelligence forms
  • How do we arrange for good governance of the whole society in such an environment?
  • How will the interests of the different intelligences be balanced?
    • Note that we’re only now beginning to consider the interests of intelligent animals

Changing Interactions with Technology

Human-machine intelligence collaboration

  • Human activities will increasingly take advantage of machine intelligence in a collaboration
  • Have the AI do the things it does best, and humans do the things they do best
  • This only extends a trend that has been going on for many years
    • Today, everyone lets the spreadsheet do the math for a large problem instead of laboriously computing the figures by hand

Changing modes of interaction

  • Up to now, the dominant mode for interacting with computing systems has been the keyboard, cursor control, and screen
    • Display of text and static images
  • AI systems are getting better at understanding and producing speech with all its nuances
    • We will increasingly converse with machines in the same way we would with another human
  • Systems will incorporate much more interactive video with synthesized avatars
  • The computing system will feel much more like a person

Systems will be listening to and watching you

  • Your Alexa device is listening 24 hours a day when it is turned on
  • Your screens will be watching you at the same time you are watching them
  • Information they collect will be shared with all sorts of other entities, with or without your knowledge or permission

Personal systems will be more directive

  • Smart personal systems are going to be ever-more-proactive in making recommendations to you (“for your benefit…”)
    • The Fitbit model, moving into all aspects of your life
  • How to keep them from being totally annoying?
    • Maybe you really don’t feel like going to the gym today, no matter what the machine says

Self-Driving Cars and Trucks?

Integrating driven and driverless cars

  • As of 2018 over 250 million driven vehicles in the U.S.
    • Driven vehicles aren’t going away anytime soon
  • How to mix driverless and driven vehicles safely?
    • Under a wide range of road environments
  • Need for huge public investments to support driverless vehicles
    • E.g., in highway network infrastructures
  • Today, the driverless vehicle is bursting forth without a solid legal, ethical and priorities framework

Car automation levels

  • Level 0: No automation
  • Level 1: Driver assistance: controls either speed or steering autonomously to assist the driver
  • Level 2: Partial automation: controls both speed and steering autonomously to assist the driver
  • Level 3: Conditional automation: controls speed and steering autonomously but requires driver monitoring
  • Level 4: High automation: able to complete a trip with no driver involvement under normal environmental conditions, not requiring driver monitoring
  • Level 5: Full autonomy: able to complete a trip with no driver involvement in any environmental conditions

Challenges of automating vehicles

  • Adverse environmental conditions (e.g., fog, heavy rain, snow; ice; potholes; rocks on the road; etc.)
  • Reduced system functionality (e.g., snow covering a camera lens, broken radar, software bug)
  • Unexpected road conditions (e.g., road maintenance, game day, accident, mattress flipping off the car ahead)
  • Uncertain behavior of other entities on the road (pedestrians, bicycles, animals; drunk drivers, police directing traffic, etc.)
  • Mixing driven and driverless vehicle behaviors
    • How drivers will interact with driverless cars, e.g., at four-way stops or left turns across heavy traffic
  • Extreme software size and complexity
    • Difficulty validating it and keeping it updated
  • Keeping every subsystem working to specification
  • Preventing system freeze-ups (the “blue screen of death”)
  • Responding to surprise conditions (e.g., siren and flashing lights of the ambulance coming up behind)
  • Phasing in driverless cars with partial infrastructure (incomplete wireless networks in rural areas, etc.)
  • In general, trust in automated vehicles
    • Am I really comfortable with that driverless semi in the oncoming lane, particularly in this heavy rain?

Control handover

  • How does the vehicle hand control back to the occupant when the conditions exceed the automation’s  capabilities?
    • The car unexpectedly signals to you “Take over! You have control now!”
    • Very hard for humans who have not been engaged to take over suddenly from automation
    • Aviation has found by painful experience this does not work well
    • Timing of the handover is critical
  • And how does the car know when to pass control over?

Trust in automated vehicles

  • Self-driving vehicles can be safer than human-driven vehicles under optimum conditions
    • Don’t suffer from distractions, anger, inebriation, etc.
  • But they have severe difficulty in handling adverse cases
  • People may be inclined to trust them when they really shouldn’t
  • Hard to know when to trust and when not to

The hacking risk

  • The more vehicles incorporate smart systems connected via wireless networks, the more vulnerable they are to malicious hacking
    • Not just stealing your car or things you left in it, but potentially causing your car to crash and kill you
  • Very little attention has been paid to vehicular cybersecurity up to now
    • No standards and regulations have been established

Public responses to self-driving vehicles

  • Already there are people resisting the introduction of self-driving vehicles into their area
    • People driving in ways to interfere with them
    • People vandalizing them (slashing tires, damaging sensors, etc.)
  • Self-driving cars are going to have significant additional expenses over driver-controlled versions
    • Both initial costs and maintenance/repair costs
    • All those sensors are expensive and have to be carefully calibrated

Self-driving cars and ethics

  • How do we define values and ethical principles for self-driving cars?
  • How conservative do you want your car’s driving to be to maximize safety?
    • How safe is too safe?
    • What if some owners of self-driving cars set a low value on your safety?
  • When should your car determine it has to kill you, in order to save the lives of others?

What moral decisions should self-driving cars make?

Ubiquity?

Smart everything

  • As AI capabilities become ever-cheaper, we can expect some form of AI to become incorporated into all manner of products
    • Even relatively inexpensive items, such as toys
    • If it has a battery or plugs into a wall, it may well have a processor running AI programs
  • This will be combined with the Internet of Things, where devices will be wirelessly linked into central networks of various kinds

Some consequences of smart everything

  • Smart consumer items that are connected through the general Internet will expose you to all sorts of things you might not appreciate
    • Monitoring and reporting on your activities
      • Your television watches you as you watch it
      • Alexa listens to every sound in your home
    • Hacking: all kinds of possible mischief
      • E.g., unlock your doors and turn off the security system so your house can be burglarized
  • Low-cost consumer systems are terrible in terms of implementing effective cybersecurity

AI and Personal Information

Ain’t no such thing as a free lunch

  • Making use of these nifty new capabilities is definitely a mixed blessing
  • There are lots of hidden costs and potential downsides
  • Just like today when you use Google or Facebook
    • You don’t pay directly for any of these helpful functions
    • In exchange, you are implicitly allowing the company to collect all kinds of information about you
    • They make their money by selling your personal data to third parties who aren’t necessarily your friends

The system knows everything about you

  • Increasingly, all sorts of personal information is being collected, compiled, and analyzed by entities of all sorts
    • Marketing companies such as Amazon that want to sell you things
    • Platform entities such as Google and Facebook whose business model involves selling your profile
    • Governmental agencies of all kinds that want to monitor what you do
    • Many others, including entities that want to detect vulnerabilities, in order to defraud or steal from you

Your personal data is no longer your own

  • Potentially everything you do while interacting with some kind of an information system may be monitored, recorded, transmitted to an archive, combined with other data, and analyzed by AI systems
  • Who has access to the resulting knowledge about you?
  • What conclusions are they able to draw?
  • What are they able to do to you as a result?
  • What protections do you have against being harmed?

AI monitoring and analysis

  • AI-based software: monitoring and analyzing all you do
    • Every web page you visit and what you do on it
    • Every cable television program you watch
    • Every email, text message, Facebook post, etc.
    • Every phone call
    • Everything you purchase with a credit card
    • Every financial transaction, the status of each account
    • Your location at any time
    • How you drive
    • Every interaction you have with the medical system
    • Your daily schedule
    • Who you interact with

What happens with your information

  • Entities share your information with each other in ways that benefit their interests
  • You have almost no control of what happens with your information
  • It is almost impossible to find and remove errors
    • Bad information keeps getting re-inserted from old files
  • AI systems are very powerful at finding patterns of all sorts in the data about you

Knowing too much about each other

  • Consider a Google glass system with AI face recognition and wireless access to central databases
  • The policeman wearing it sees which passerby has an outstanding warrant or a history of violence
    • Will only the authorities have access to such capabilities?
    • If I have an equivalent system, can I see the cop’s history of police brutality complaints?
  • Eventually, can anyone see any data on anyone else?
  • How will access to such information be controlled?   Who sets the policies on this?

Social control: the Chinese social credit system

  • China is currently running one of the biggest social control experiments of all time, using extensive AI
  • They plan to rank all Chinese citizens based on their “social credit” score by 2020
  • The social credit score is similar to the Western financial credit score, but much more comprehensive
  • The scores move up or down according to peoples’ behavior, over their whole lives
  • People are rewarded or punished according to their scores
  • The scoring algorithms are state secrets
  • Correlation of commercial and government data bases
  • Examples of things that drop your score: bad driving, smoking in prohibited areas, paying a bill late, criticizing the government; an inappropriate social media post
  • Your score is correlated with those of the people you associate with: family, friends, coworkers, etc.
    • Your score can be reduced by something they do
  • People with low scores face reduction in rights and opportunities: education, jobs, housing, travel, etc.
  • The system is intended to strongly enforce conformity to government-defined norms
  • https://www.youtube.com/watch?v=y5-0llHaZDs

Politics around AI and robotics

  • Automation and globalization are causing increased support for nationalist and radical right political parties
  • Target for resentment: Highly educated coastal “elites” who make very good livings developing AIs and robots that put “ordinary people” out of work
  • Technological disruption is not even on the radar screen for most political groups and politicians
  • Few elected officials understand technology policy issues
  • Major geopolitical implications of a global AI technology race between national powers

Near-Term Implications of AI in Different Fields

Types of jobs that will stay, types that won’t

AI in medicine

  • All medical instrumentation connected to a patient will feed into an integrated pattern recognition capability
    • The Internet of Medical Things (IoMT)
    • A holistic real-time picture of the patient’s status, correlated with the patient’s history
  • Almost any medical condition with an acute episode—e.g., asthma attack, seizure, stroke, heart attack, autoimmune attack—will be potentially predictable
  • Diagnostic systems are already being transformed by AI
  • Similarly, treatment recommendation systems

AI in mental health

  • Today, mental health care is tremendously constrained by the shortage of trained providers relative to the needs
    • Also major problem of care affordability
  • AI systems are becoming increasingly capable of interacting with patients and detecting markers of mental health issues at an early stage
    • Depression, anxiety, PTSD, substance abuse, cognitive or memory decline, etc.
  • AI systems may become a part of mental health counseling and support as well
    • A virtual therapist available 24/7/365, at low cost

AI in human reproduction

  • AI is highly synergistic with genetic technology
  • Breakthrough techniques such as CRISPR/cas9 and gene drive now allow ready modifications of genomes
    • Features are heritable, passed on to descendants
  • The temptation for parents to create designer babies with desirable characteristics will be very high
    • Health, long life, intelligence, athletic ability, beauty, musical talent, creativity, etc.
  • Cost: will likely make such technology available first to those who are already economically advantaged

AI in law

  • Most of the law is based on analysis of highly-structured data—e.g., contracts, suits, etc.
  • AI systems are ideally suited to perform these functions
  • The digital law library is the raw material for AI law
    • Ability to search cases for precedent judgments
    • Increase the span of the search
  • Similarly, AI systems can assist human judges in rendering verdicts based on analysis of evidence, case law, other factors
    • Question: will people accept judgments rendered by an AI system?

AI in finance

  • Already, high-volume trading is almost completely conducted by AI-enabled computer algorithms
    • Stocks, bonds, commodities, currencies, etc.
    • Blockchain transactions (e.g., Bitcoin)
  • Investment analyses are performed by AI systems
  • Financial decisions of all sorts are increasingly made by machine intelligences advising the humans

AI in insurance

  • AI systems will enable fine-grained assessment of individual policyholders’ risks based on their profiles
    • Underwriters are going to know a great deal more about policyholders than ever before
  • In place of statistical-based insurance pools, insurance will become much more individualized
  • Might result in increased denial of insurance or very
    high premiums for policyholders deemed to be high risk
  • A powerful influence on behavior

AI and audits

  • Currently, audits (e.g., of taxes) using manual processes are time-consuming and expensive
    • As a result, only a small fraction are actually audited
    • Lots of discrepancies slip by as a result
  • Automated systems using AI are likely to greatly facilitate performing audits
  • What will be the effect of comprehensive audits becoming routine?

AI in education

  • AI can support highly individualized education and training, at the student’s own pace
    • Using affective computing to interact like a tutor
  • An AI system can study the individual student and develop a detailed model of him/her
  • Learning styles, preferences, cognitive biases, etc.
    • What the student does and doesn’t understand so far
  • It can tailor the presentation of the material to best match the student at each stage of the course
    • Repeat material as necessary, in different forms, until the student really masters it

AI in the media

  • Chinese state television is already running a TV news anchor that is not a human, but a simulacrum with AI-synthesized imagery and voice
  • Sports reports are beginning to be generated by AI systems
  • Similarly things like weather reports, business reports, etc.
  • More and more text reporting in newspapers and online media is going to be written by AI systems

AI in marketing

  • AI analyses will be conducted on all manner of data collected about you
    • Your prior purchases, your product searches on websites, your attention to ads
    • All aspects of your financial situation
    • Demographic data on you, your family, your friends, your coworkers, your neighbors, etc.
    • Your psychological profile
  • Precision-targeted marketing will be generated to maximize its appeal specifically to you

AI in manufacturing

  • Obviously, smart robotic systems are going to perform ever-larger fractions of production work
  • AI systems are going to be coupled with advanced 3D printing/additive manufacturing systems
    • Quickly produce custom or low-volume parts
      • E.g., clothes custom-tailored to fit your measurements  exactly
    • Do just-in-time manufacturing
  • Manufacturing with AI systems may shift from mass production to mass customization

AI in engineering, architecture, etc.

  • The engineer’s knowledge base is increasingly being codified to create intelligent engineering assistants
  • The engineer’s role will be raised to a higher level, focusing on defining what the design should do
    • The automated system will then detail out a candidate design for the engineer to review
    • Many more options and alternatives can be explored to get to an optimum

AI in software development

  • Examples of applications of AI in software development include
    • Requirements analysis and specification support
    • Intelligent programming assistants
    • Software performance assessment
    • Software interface management
    • Software test administration
    • Software verification and validation
    • Software defect and vulnerability detection

AI in government

  • Some examples of AI applications in government include
    • Regulatory compliance monitoring
    • Tax assessment and collection
    • Application review of all types
    • Public assistance and entitlement program administration (unemployment, Social Security, Medicare, Medicaid, veterans programs, etc.)
    • Immigration administration
    • Election administration

AI in the military

  • Self-guided weapons have been around for a long time
    • E.g., homing torpedoes, heat-seeking missiles
    • Up until recently, these have been targeted and launched by human decision makers
    • Such weapons have been incorporating more and more machine intelligence in order to defeat countermeasures and improve performance
    • Some of them are already autonomous
  • To what degree do we want military systems to make decisions on their own, with no human in the loop?
    • What policies govern autonomous weapons?
  • Logistics is a huge part of military operations
    • Procurement, supplies management, transportation, equipment maintenance, etc.
  • AI-enabled systems can vastly improve logistics management over human administration
  • Strategy and tactics are also opportunities for AI
    • Similar to game-playing systems, an AI program can play Blue vs. Red for thousands of iterations to select the best options for success in a conflict
  • Robotic systems are likely to largely replace manned systems on the battlefield
  • For example, why should combat aircraft have humans onboard?
  • Why should conveys of supply trucks have human drivers?

AI in law enforcement

  • Today, concepts are being explored for AI-enabled predictive policing
    • Based on indicators of all kinds, when, where, why, and how are crimes most likely to occur?
      • Criminal acts
      • Perpetrators
      • Victims
    • Applies to all types of crimes, from fraud to violence
  • How is law enforcement likely to transform with an emphasis on pre-crime actions?
  • What if the authorities have some kind of compromising information on everyone?  Enforcing will then become selective
  • How tolerant does society want to be for small infractions?
    • Your AI-enabled car could easily report every time you exceed the speed limit or fail to stop completely at a stop sign for a right turn
  • How comfortable will we be with robot policemen, especially if they are autonomous and authorized to use lethal force?

AI and cyberwarfare

  • Hostile cyber acts are an escalating problem
  • Cyberwarfare is a potential weapon of mass destruction
    • E.g., take down a nation’s power grids
    • Destroy a nation’s financial systems
    • Cause nuclear reactors to self-destruct
  • Most current military systems, particularly AI-enabled ones, are extremely vulnerable to cyberattacks
  • Essential to detect and respond to cyberwarfare attacks on all critical systems
  • There will be a continual arms race between attacker and defender

AI and cyberterrorism

  • Cyberterrorism is a parallel threat to cyberwarfare
  • Often, it is not clear who is the source of a cyberattack
    • Attackers can be anything from a hostile nation, to some rebel faction, to a few individuals
    • Very difficult to deter cyberterrorism
    • Very difficult to find and punish the attackers
  • The number of potential targets and possible attack vectors is huge
  • Like cyberwarfare, the consequences could be sever

AI and cybercrime

  • There is likely to be a growing arms race between those employing AI to commit crimes and those employing AI to detect and thwart such crimes
    • Particularly major financial crimes
  • AI systems can be extremely capable at searching for vulnerabilities to exploit, while avoiding detection
  • Cybercriminals can be based anywhere in the world and attack anywhere
  • The potential payoffs for cybercrime can be huge
  • It is critical that defenders have the resources to over-match cybercrime attackers

AI in elections

  • Democracy depends on the legitimacy and trustworthiness of the election processes end-to-end
  • The integrity of voting is key
  • Current voting machines, vote compilation systems, and vote reporting systems are horribly insecure
    • ~185,000 polling places in the U.S.
    • Numerous vendors of systems
    • No general oversight of security
  • AI systems are needed to detect interventions throughout the election chain

AI and the Arts

AI and music

  • AI systems can already compose original music in the style of a particular human composer that can fool experts
    • Not great masterworks yet , but steadily improving
  • What will be the future of human-created music?
    • Will it become a niche, like a hobby?
    • Will people still value it in the same way?
  • See https://www.youtube.com/watch?v=wYb3Wimn01s

AI and new musical forms

  • A neural network can learn the musical characteristics of an instrument by analyzing hundreds of notes
  • It creates a mathematical representation, or vector, that identifies a particular instrument
  • Now these vectors can be combined to create entirely new instruments
  • One new synthetic instrument might be 47 percent bassoon and 53 percent clavichord. Another might switch the percentages

AI and photography

  • A new revolution: computational photography
    • Enabled by major advances in processing hardware and Deep Learning AI systems
  • Ordinary smartphones are going to be capable of astounding photographic capabilities
    • Lenses and sensors no longer the limiting factors
    • The camera is going to understand what it is looking at and adjust the image to maximize quality
  • Photo libraries will be searchable on the basis of content
  • With learning, the capabilities will continue to get better and better over time

AI and art

  • Christie’s recently sold its first AI-created portrait painting for $432,500
  • AI systems using GANs can already create convincing original art (paintings, sculptures, etc.) In the style of a particular human artist
    • Again, not great masterworks yet
    • The systems will only get better with time
  • How should we appreciate and value art created by machine intelligences?

AI and computer games

  • AI is already a key part in advanced computer games, particularly multiplayer games
    • Both your partners and your adversaries may be AI entities
  • Your character’s abilities may be enhanced by AI
  • Will games continue to be as much fun as the machine takes a greater and greater role?

AI and toys

  • Higher-end toys are going to be highly interactive using AI
    • E.g., interfacing through a smartphone or smart speaker, networking through the cloud
  • Audio chat, video, etc.
  • Extensive use of affective computing
  • Toys will develop compelling individual personalities in the process of interacting with the user

AI and the movies

  • Computer-generated imagery in the movies already employs extensive AI technology
    • Steadily getting better, faster, and cheaper to use
  • We are beginning to see movies where actors no longer living are recreated digitally to play new roles
    • Will human actors continue to have the same value?

AI and virtual/augmented reality

  • The ability to generate high-resolution imagery, sound, and other sensory effects in real time using AI capabilities is advancing quickly
    • Making possible ever-more convincing virtual reality and augmented reality systems
  • How important will virtual/augmented reality be in future entertainment?

Fake people

Deepfake video

  • Photoshop capabilities today are for still images
  • Emerging: AI-based equivalents for digital video
  • Analyze posture, body language, movement, facial expressions, vocal tone, linguistic characteristics, other features from recordings
  • Then synthesize a convincing digital video segment
    • E.g., a politician saying things that she or he did not say,
    • Or a celebrity shown in a porn video
  • You won’t be able to tell the difference from reality
  • Consequence: Never again trust a video you see on television or online

Deepfake example

AI and Robot-Human Interactions

Robot co-workers

  • Humans and intelligent robots will increasingly function as co-workers
    • E.g., doing construction work, where the robot has capabilities (reach, strength, etc.) beyond that of a human
  • The robot will interact with its human partners through speech and will perform some functions autonomously
  • Work teams will be mixtures of humans and robotic systems

Human assistance robots

  • One very active area of development is robots to provide assistance to humans with physical and cognitive limitations of various types
    • Mobility assistance of all kinds, helping people transfer (e.g., getting up from a bed or chair)
    • Helping with activities of daily living (preparing and serving food, feeding, bathing, dressing, etc.)
    • Household tasks (cleaning, laundry, etc.)
    • Supplement the functions of service animals
  • Some of the physical tasks are actually quite challenging for a robot
  • Probably won’t be a single robot able to do everything

Companion systems

  • AI systems will come to serve as companions
    • Already, people are developing feelings of relationship with their intelligent virtual assistant systems (Siri, Alexa, Google Home, etc.)
    • The system will come to know you intimately
    • You will disclose your innermost feelings to it
    • It may become a more entertaining conversationalist than even the cleverest of your human friends
  • Loneliness/isolation afflicts a large fraction of people, and companion systems can help fill the gap

Machine intelligence and sex

  • Sex has been a driver for a lot of digital technology
    • Online pornography influenced the development of high-bandwidth Internet connections for streaming video
    • Today people are developing high-fidelity human-simulating robots for sexual interactions
    • Initially crude, these are becoming more and more lifelike—e.g., able to hold real conversations, as well as being available for sex whenever desired
    • See https://www.youtube.com/watch?v=LcDWigVV6tA
  • Society has yet to decide how to feel about sex with robots—a boon, or an abomination?

The creepiness issue

  • People tend to find mimics of humanness that are just slightly off to be creepy
  • There is not a problem if the simulation is obviously an artifact
  • Humanoid objects which appear almost, but not exactly, like real human beings elicit feelings of eeriness and revulsion in observers
  • Examples are found in robotics, 3D computer animations, and lifelike dolls

What Are the Worries About AI?

Worries are a function of time

  • There are near-term worries, mid-term worries, and long-term worries
  • Near-term worries are mostly about rapid disruptions of the existing order
  • Mid-term worries are about fundamental changes in how humans live and act in the world
  • Long-term worries are existential—are we creating conditions that threaten basic human existence?

Some near-term worries

  • The disappearance of anonymity and privacy
  • Machine intelligence in the service of human stupidity
  • Transferring authority and responsibility to machines
  • AI’s effects on society’s concentrations of power
    • Benefits will go to the wealthy and well-connected
  • Exploitation of human weaknesses
  • AI enablement of techno-authoritarianism
  • Massive changes to everyone’s jobs
    • Constant raising of the bar to be employable
  • Economic disruption from large-scale unemployment

What worries AI insiders?

  • Handing too much responsibility over to AI systems and becoming over-dependent on them
  • Inappropriate trust in AI systems
  • Over-estimating the competence of AI systems
  • Security of AI systems, including under attack by AIs
  • Keeping human cognitive skills from atrophying
  • Powerful AI in the “wrong hands”—hobbyists, hackers, rogue regimes, criminals
  • Increase in social inequality
  • Giving up too much of our humanity to machines

Example of risks: financial trading

  • August 12, 2012: Wall Street’s largest trading company, Knight Capital, switched on a new AI-enabled program for buying and selling shares
  • Due to an undetected bug, the system immediately began flooding the exchanges with irrational orders
  • It took 45 minutes for Knight’s programmers to diagnose and fix the problem
  • During that time, the software made over 4 million deals, with $7 billion in errant trades
    • This nearly bankrupted the company

Keeping up human thinking skills

  • As we hand over more and more of our thinking skills to machine systems, how do we prevent our own skills from atrophying from lack of use?
  • Do we need to require periodic refraining from using automated systems?
    • Like airline pilots who are required to regularly practice landing by hand to keep their skills sharp
  • How can we encourage everyone to maintain their human thinking skills when the automation is so good and so convenient?

Ceding too much to machines

    • One concern is that we will cede excessive amounts of responsibility to AI systems
    • The global financial system is a good example of a risk area
  • Potential for a catastrophic system failure caused by multiple minor flaws in over-empowered connected machine systems

Transfer between domains

  • We aren’t good at assessing how a highly-optimized rule or structure in AI will transfer to a new domain
  • How do we know when a machine intelligence has left its comfort zone and is operating on parts of the problem it is not good at?
  • Don’t put a machine intelligence in charge of decisions it doesn’t have the intelligence or knowledge to make

AI and techno-authoritarianism

  • AI can be a powerful enabler for authoritarian societal control
  • Surveillance of the population can be near-total using AI tools
    • Any deviation from what is authorized is detected and responded to
    • Resistance and rebellion can’t even get started
  • The Chinese social credit score system goes a long way in the direction of techno-authoritarianism
  • Could be much worse—think North Korea
    • Say the wrong thing or make the wrong face in front of a screen and you disappear, never to be seen again

Pressures for conformity

    • Even without more extreme forms of techno-authoritarianism, AI systems could lead to excessive pressures to conform to social norms
    • Everything you do is observed and evaluated and potentially shared with others, outside your control
      • Very hard to have any secrets or private quirks
  • Likelihood of stifling creativity, innovation, risk-taking, etc. 

Globalization,  Automation, AI and Jobs

Globalization effects

  • In addition to automation and AI, globalization is having a huge impact on jobs
  • Work is moved to the place on the Earth where costs are least
    • Lowest labor costs
    • Lowest capital costs/best subsidies
    • Lowest taxation
    • Least burdensome government regulation
    • Least concern for externalities (e.g., environmental damage, social harms)
  • Automation and AI effects are compounded by globalization
  • The developed economies (e.g., the U.S., Europe, Japan, etc.) are being surpassed
  • China is about to become the largest world economy
  • India will shortly become the second largest
  • Other Asian economies are among the fastest growing
  • These economies have fewer pre-existing institutional and cultural barriers to automation and AI

Jobs: AI and automated systems vs. humans

  • An AI system or a robot doesn’t:
  • Demand a salary or wages
  • Need health care insurance, Social Security and Medicare contributions, or workman’s comp
  • Take vacations, holidays, maternity or sick leave
  • Join a union or strike for better working conditions
  • Need a human relations department
  • Can work 24 hours a day, 7 days a week, 52 weeks a year without tiring or taking a break
  • Unquestioningly obedient, highly motivated, and cooperative

Human aspects that AI systems don’t have

  • Human mind overhead
    • Distractions
    • Worries
    • Emotional commitments
    • Charged memories
    • Allegiances
  • Susceptibility to a wide range of cognitive biases

Intelligent machines and jobs

  • Unlike a human employee, the employer gets a tax break for machine depreciation
    • Latest U.S. tax law: employer can expense 100% of a robot’s cost the first year, rather than over the life of the machine
  • Result: overwhelming economic pressures to replace human workers with AI and robots wherever possible

Whose jobs are at high risk?

Anyone

  • Doing well-characterized repetitive physical actions (manufacturing; construction; mining; farm work, etc.)
  • Doing financial services (e.g., bank tellers, financial analysts and advisors, loan officers, accountants, bookkeepers, brokers, insurance underwriters, claims representatives, etc.)
  • Doing highly structured analysis (e.g., paralegals and legal assistants, statistical analysts, report writers, etc.)
  • Doing routine middle management tasks unlike a human employee, the employer gets a tax break for machine depreciation

Some other jobs at high risk

  • Office clerks of all types
  • Inventory workers
  • Retail salespersons
  • Drivers (truck, taxi, etc.), couriers
  • Food service personnel
  • Telemarketers
  • Journalists
  • Security guards
  • Medical diagnosticians
  • Pharmacists
  • Postal service workers
  • Meter readers
  • Computer operators

Whose jobs are going to remain?

Jobs that will better resist replacement by machine intelligences involve:

  • Extensive face-to-face / hands-on human interaction
  • Unpredictable physical work, requiring on-the-spot creative adaptation
  • Extensive, broadly-based education and experience
  • Critical thinking, high creativity, and the ability to innovate

Jobs we’ll lose, jobs we’ll keep

See https://www.ted.com/talks/anthony_goldbloom_the_jobs_we_ll_lose_to_machines_and_the_ones_we_won_t

What jobs will change dramatically?

  • Education
  • Medicine
  • Therapy
  • Software development
  • Engineering
  • The military
  • Skilled trades
  • Performing arts

Adapting to a new work environment

  • Needed: creativity, critical thinking, emotional intelligence, adaptability, and collaboration skills
  • Learning how to learn—and constantly reinvent oneself and develop new abilities over a whole lifetime
  • The problem is that not everyone is cut out for this degree of independent learning and self-reinvention
    • Takes a lot of drive, self-direction, self-discipline
  • Likely to make current inequalities between people even greater in the future

Automation/AI and the Economy

What will jobs look like in an AI-based economy?

  • Some types of jobs that will remain will be those:
    • Requiring scarce talents and high levels of education
    • Requiring extensive face-to-face human interaction and emotional intelligence
    • Requiring good trade school training and excellent problem-solving skills
  • Few jobs for those who lack scarce talents and have modest levels of education

 Economic dynamics of a highly-automated world

  • Customers are required in order for the goods and services that are produced to be purchased
    • Customers have to have money in order to buy
  • In the current economic system, most consumer buying power comes from current employment, or savings derived from previous employment
    • Automation will displace a large fraction of jobs
    • No jobs = no buying power
  • With no buying power, no demand
  • Result: economic contraction, in a positive feedback loop

Changes to the basic economic structure

  • Who benefits from AI? the owners of capital, who will control most of the intelligent machines
  • Who suffers? the rest of us, who currently trade work for money. No work means no money
  • What is the responsibility of the rich and powerful to the rest of us?
  • When intelligent machines become more profitable to them than human workers
  • The response to the mass unemployment of the AI Revolution has to involve some kind of sweeping redistribution of income that decouples it from work

Redesigning basic aspects of society

  • In a highly automated world, we will have no choice but to fundamentally redesign some basic aspects of the society
    • How we define societal values and goals
    • How we govern ourselves
    • How we distribute power
    • How we distribute/redistribute wealth
    • How we protect ourselves against parasites, predators, free riders, etc.

Automation-induced obsolescence and cultures

  • What happens when any work we might do can be done better by machines?
  • Automation-induced obsolescence has already had devastating effects on some of the world’s people
  • Some cultures (e.g., those based on planting, cultivating, and harvesting corn and beans) have collapsed and lost their meaning to the people who were shaped by them
  • As automation replaces more and more human work, how do people continue to maintain their sense of worth?

What will society do with people no longer “useful”?

  • People find meaning in work, in doing something they and others consider valuable and appreciated
  • When people feel they have nothing to contribute, it is very harmful to them psychologically
  • If your skills are taken over by machine systems, what are you to do in order to feel you’re of worth?
    • And for the society to feel you’re of value?

Higher Level Worries

Some mid-term worries

  • Developing AIs with autonomy and agency/self-interest without associated protections for humans
    • Letting AIs act on their own interests, not humans’
  • Failing to incorporate appropriate human values, ethics, and moral reasoning frameworks in AIs
  • Creating machine intelligence persons with protections and rights prior to developing appropriate societal structures to accommodate them
  • Abdicating human responsibilities for actions taken by machines

Motivations of AIs

  • Clearly we want AIs to be motivated to loyally serve the interests of their creators and owners
  • We also want AIs to be highly motivated to serve the interests of humans in general
  • We would like them to have a well-developed sense of responsibility for their actions
  • Do we want to allow AIs to develop motivations that are not tied to those of humans?
    • What could be the possible consequences?

Autonomy and self-interest

  • Humans and animals have autonomy and self-interest
  • Instincts for self-protection / self-preservation
  • Dare we give machine systems similar instincts?
    • “Don’t touch my power switch, human!”
  • What about a desire to increase their access to resources?
  • What about a desire to reproduce themselves?

Dangerous machine intelligences

  • Don’t create machines instructed to “survive, reproduce, access resources, and improve in the best way possible”
  • Such systems are not likely to remain friends of humanity for very long
  • Self-reproduction combined with autonomy: the really dangerous step in the development of machine intelligences

Things we can’t afford to get wrong

  • Self-interested AI systems interacting with each other
  • AI systems developing and evolving on their own, outside human control
  • Technological synergies conducted in an unmanaged way—e.g.,
    • AI and genetic engineering
    • AI augmentation of humans

Research it is critical to do

  • How to manage the development of safe and beneficial AI is a very important and challenging research problem
    • Highly interdisciplinary
    • Involves many different kinds of institutions
    • Requires collaboration among researchers in many different fields
    • Needs to be fully international
  • Have to start working on these aspects now
  • We can’t afford to learn from making mistakes—have to anticipate them and steer clear from the start
    • May only have one shot at getting it right!

Machine Intelligences and Personhood

What is a “person”?

  • We distinguish between entities that we consider persons and entities that we consider are not persons
  • “Persons” are given rights and protections on the basis of their “personhood”
    • These are withheld from “nonpersons”
  • Increasingly, we are grappling with the question of how to define personhood
  • Is personhood a single state?  Or are there degrees of personhood?
  • How is personhood a function of consciousness?
  • The definition of “person” and the associated social rights and protections have varied greatly over time and over different cultures and belief systems
  • In most eras, women did not have the same rights and protections as did men
  • Similarly, slaves did not have the same rights and protections as free persons
    • Prisoners still do not
  • Previously, rights and protections depended on race
  • Children and youths have rights and protections that depend on age
  • Recently we have been expanding protections (e.g., of disabled persons)
  • In the U.S., courts have determined that corporations are “persons” in the eyes of the law.  What in the heck does that mean?

Who/what/when is a “person”?

  • Is a human ovum just penetrated one second ago by a sperm a person?
  • Is an 8-cell human zygote in a fertility clinic freezer?
    • Just when does a fetus become a person?
  • Is a human infant born with anencephaly?
  • Is a brain-damaged human adult kept on life support in a persistent vegetative state?
  • Is an elderly human adult with extreme dementia?
  • Is a human conjoined twin with a single body and two heads one person…or two?

What is a person?  New cases

  • Would an augmented or modified human produced by genetic engineering be a person?
    • With what degree of augmentation or modification?
  • Would a recreated Neanderthal be a person?
  • Is a highly intelligent nonhuman (e.g., a bottlenose dolphin, a chimpanzee, or an elephant) a person?
  • If we encountered one, would an extra-terrestrial with human-equivalent intelligence be a person?Is a human ovum just penetrated one second ago by a sperm a person?
  • Will a cyborg that has a combination of human and machine intelligence be a person?
  • Will a humanoid robot with a degree of self-awareness and agency be a person?
  • Will an AI program on a distributed network with self-awareness and agency be a person?

Degrees of personhood

  • If we decide to have degrees of personhood, how do we all agree on the scales and the associated rules?
    • A lot of personhood has ultimately to do with consciousness
    • Consciousness is a complex subject in its own right
  • In addition to rights and protections, what societal obligations and responsibilities are associated with degrees of personhood?

Personhood and ownership

  • Contemporary principle:
    • Non-person entities can be owned by persons
    • Persons cannot be owned by other persons
      • Corporate persons modify this principle by being able to own other corporate persons
      • Persons can own corporate persons
  • What will ownership come to mean, in the emerging world of intelligent machine entities?

Consciousness and Machine Intelligence

Consciousness: not a simple matter

  • The meaning of consciousness is the subject of intense study and debate
  • What exactly does it mean for an entity to be conscious?
  • Most human thinking actually happens at a level of subconscious processes
    • This is why you can arrive at your home without being aware of the drive to get there
  • When we speak of “consciousness”, there are actually multiple aspects to consider
  • See https://www.youtube.com/watch?v=CTHh-5kcqC0

Sentience

  • Sentience is defined as the capacity to feel, perceive, or experience subjectively
  • Animal rights activists note that animals have sentience, particularly in terms of their capacity to suffer
    • They argue they should then be provided with certain rights and protections
  • To what degree will we endow machine intelligences with sentience and feelings?
    • And what will be the consequences?

Self-awareness

  • Self-awareness has to do with being conscious of and observing the flow of one’s thoughts
    • Being aware of being aware
    • Thinking about one’s thinking
  • Self-awareness is kind of a meta-level of consciousness
  • A significant portion of non-human animals can be shown to have at least a degree of self-awareness
  • Although AI systems don’t need to have self-awareness to function, there is no barrier to having it

Sapience

  • Sapience is defined as the ability to think and act using knowledge, experience, understanding, insight, and common sense
  • Sapience is often equated with wisdom
  • We even term our species homo sapiens: “man the wise”
  • To what degree will we endow machine intelligence with the capacity for sapience?

Consequences of Machine Personhood

Rights and protections of machine intelligences?

  • At what point do we give rights and protections to machine intelligences?
  • And just what rights, protections, and freedoms do we give them?
    • A current example is the debate as to whether bots have protected freedom of speech
  • How are these rights and protections distinguished from those of other types of persons?
  • See https://www.youtube.com/watch?v=DHyUYg8X31c

Obligations and responsibilities of machine intelligences?

  • In exchange for having rights, protections, and freedoms, what obligations and responsibilities will machine intelligences have?
    • Obey the same laws and rules as humans
    • Pay their assigned taxes and fees
    • What else?  Serve on juries??  Perform national service??

When a machine intelligence goes bad

  • When a machine intelligence fails in the obligations and responsibilities we’ve given it, what do we do?
    • How do you deter a machine intelligence from doing something you consider wrong?
    • Is it meaningful to punish a machine intelligence after some violation?
    • If so, what would be appropriate punishment?

Will we be able to “pull the plug”?

  • People think, if an AI or robot ever becomes a problem, we could always “pull the plug” to shut it off
  • Unfortunately, if an AI system develops self-interest, it will likely act to prevent this
    • E.g., disable its own off switch and get access to backup power
    • Or transfer its memory and thinking states to another connected machine system and thus live on
  • We may not have the right to shut off an intelligent machine endowed with personhood

When an AI system causes harm, who will be held responsible?

  • Inevitably, AI systems will cause harm
    • E.g., self-driving cars will make errors, get into accidents, damage property, and hurt people
  • Who to hold responsible
    • Owner of the car?
    • Occupant, who didn’t take control?
    • Manufacturer of the car?
    • Manufacturer of the car’s control system?
    • Programmer of the AI software?
    • The city, whose infrastructure was defective?
  • We don’t have a legal structure for such things yet

When thinking machines break the law

  • Thinking machines will inevitably break laws
  • Previously, we held the person controlling a machine to be accountable
  • When the machine is more autonomous, accountability becomes complex
  • Or when the machine’s behavior is the result of many peoples’ inputs, not all of whom are acting in concert
  • How will responsibility be assigned?

How does the legal system have to change?

  • The current legal system is based on assigning and allocating responsibility to human actors
    • Both in the case of harm and benefit
  • When machine systems take actions with consequences, how will that responsibility be considered?
  • The machine intelligence may be extremely distributed, with many components involved in a particular action

AIs and moral decisions

  • We will be charging intelligent machines (e.g., self-driving cars) with making serious moral decisions
  • For example, how does the machine decide which harm to choose when there is no harm-free course of action?
    • We will want them to quantify and weigh different types of potential harms to different entities
    • How are they to weigh injuries of what severity and likelihood against a fatality?  Against how much property damage?
  • Unfortunately, science is ill-equipped for answering moral questions, yet someone will have to do it for the machines we are creating

Objective functions

  • Autonomous systems are commonly designed to maximize some objective function
  • The objective function defines the goals to be sought, as well as constraints and aspects to be avoided
  • Critical to define the objective function well for any autonomous machine intelligence!
  • In general, we aren’t very good at describing our intent
    • We tend not to consider possible unintended consequences in advance
    • We don’t anticipate possible corner cases
  • Be very careful what you ask for…

Specifying intent

  • We’re generally not very good at defining what we intend
  • Likely to be disconnects between what we ask AI systems to do and what we really want them to do
    • Especially we will fail to identify things we want them not to do
    • We can expect to be surprised by them because of our omissions

Values

  • We want machines’ decision-making to be well-aligned with human values
    • Their goal systems should be based on human values
    • Need to provide a framework for machines’ moral reasoning
  • Unfortunately, human values are not simple and consistent
    • Individuals’ values often conflict with each other
  • And whose values are to be used to instill into AIs?
    • Moral calculus differs from one culture to another and changes over time within a culture

Incentives for guiding machine behavior

  • Machines won’t have the incentives humans use for promoting good behavior
    • Sense of shame or praise, concern for reputation, deference to authority, doing something simply because it is the right thing to do
  • The social and legal systems that have dealt effectively with human rule-breakers of all sorts will fail in unexpected ways with machine intelligences

Machine saintliness?

  • How can we arrange for machine intelligences to have the highest and most positive motivations in their relationships with humans?
    • Benevolence, compassion, empathy, kindness, thoughtfulness, generosity, magnanimity, etc.
    • Humility, selflessness, patience, tolerance, self-sacrifice, etc.
  • This particularly will be a challenge when we endow machine intelligences with autonomy, self-interest, and self-reprogramming

Machine ethics

  • There is a developing field of machine ethics
  • What are the basic principles of ethics we want to build into machine intelligences?
  • Not a simple matter, as human ethics have been a matter of debate for millennia
    • People differ on ethics, sometimes strongly
    • Ethical principles have changed over time
  • Have to clarify ethics into principles that can be defined explicitly and programmed

An early approach

  • Isaac Asimov proposed The Three Laws of Robotics in 1942 (!)
    • 1.  A robot may not injure a human being or, through inaction, allow a human being to come to harm.
    • 2.  A robot must obey the orders given it by human beings except where such orders would conflict with the First Law.
    • 3.  A robot must protect its own existence as long as such protection does not conflict with the First or Second Laws.
  • Asimov later added a Fourth Law
    • A robot may not harm humanity, or, by inaction, allow humanity to come to harm
  • Unfortunately, implementing Asimov’s Laws turns out to be remarkably difficult

Corresponding principles for humans in using AI and robotics

  • Parallel need: Principles for human laws concerning the use of AI and robotic systems
  • Such laws need to be common and coordinated across nations and human societies
  • Can’t afford to default to the lowest common denominator (e.g., the least cautious developer or the most self-interested user)

AI helping humans to be more ethical

  • In order to help AIs to incorporate ethical principles, we will have to make those principles explicit and consistent
    • Lots of contradictions in existing principles will become apparent
    • Lots of failures to abide by our principles will become apparent, too
  • AIs can observe human practices and point out where we aren’t meeting the standards we promote for them
  • AIs can help us overcome our all-too-human cognitive biases

AI helping us meet human goals

  • Note that Asimov’s fourth law applies to humanity
    • Not individual factions and subgroups
    • AI should serve the interests of the whole of humanity
    • The good of the whole has to be primary
  • In the greater perspective, that means ensuring the long-term health and stability of the planetary systems on which humanity depends
    • As does the rest of the life on Earth
  • We need to use AI to help us make the wise decisions we must make
    • Even at the cost of inconvenience and pain in the near term

Wisdom and AI

  • We need to use human wisdom to guide AIs so they can help us back with using our wisdom in our own affairs
    • An interesting feedback loop!
  • We are going to have to think deeper than we have ever been challenged to think before
    • We will need to become explicit about many aspects
  • We can’t afford to be unwise where AI is concerned

Regulating machine intelligence

  • AI development is a worldwide enterprise
  • Many different national environments, including both free-market economies and command economies
  • Payoffs for being first to market are large
  • Not possible to control AI technology by classifying it
  • Technologies will quickly migrate across all borders
  • In order to forestall harmful AI, regulatory treaties and agreements are needed
  • Unfortunately, the history of international regulations on dangerous technologies has not been especially promising

Generating global agreements

  • The only people who really understand the threats of AI and robotics are the ones enabling them
  • Various forces work against global agreements.  But global agreements are necessary for the control of technologies like advanced AI or genetic engineering of humans

Some initiatives in the right direction

  • The IEEE Global Initiative on Ethics of Autonomous and Intelligent Systems
  • AI4People, launched in the European Parliament in 2018
  • The High Level Expert Group on Artificial Intelligence of the European Commission
  • The EU Declaration of Cooperation on Artificial Intelligence
  • The EU Strategy for AI
  • The Partnership on Artificial Intelligence to Benefit People and Society

Some long-term worries

  • AI is like a genie we are releasing from its lamp
  • It is going to have powers we won’t anticipate
  • It is not automatically going to be benign
  • If we give the genie foolish or poorly-considered wishes, the outcomes may be rather unfortunate for us
  • What kind of a future world do humans really, really want to exist in?  And how is AI a part of that world?

Who gets to decide humanity’s fate?

  • By developing powerful new intelligent entities, research groups are potentially deciding the fate of humanity
  • Rolling the dice in matters of great consequence
  • Like the decision to detonate the first atomic bomb
    • At the Trinity A bomb test at Alamogordo, NM July 16, 1945  some scientists were concerned it could set fire to the atmosphere and destroy the world
  • Some acts are irreversible
  • Who gets to decide what highly consequential technologies are created and released?
    • Those motivated by goals of corporate profit?
    • Those motivated by goals of military supremacy?
    • Those motivated by goals of control of their society?
    • Those motivated by intellectual curiosity?

The global brain

  • Human individuals are becoming tiny parts inside a far vaster distributed thinking system
  • We are developing a global brain, partly human and partly machine
    • The Internet and everything connected to it is the beginning of this
  • The “thoughts” that a global brain has are different from those of an individual or a less-connected society

Predicting what will happen

  • The outcome of AI development is mostly unpredictable
  • It is a complex system (in the technical sense)
    • Small changes upstream result in large differences downstream
  • Most of what we think will happen is likely to be hopelessly wrong

Intelligent machines’ evolutionary path

  • Intelligent machines won’t have the evolutionary factors that have shaped the development of human intelligence
    • Resource acquisition, status garnering, mate selection, group acceptance, etc.
  • An AI may make large numbers of copies of itself with variations
    • Then let a form of natural selection determine which ones will propagate further
  • Likely to be wide diversity of machine intelligences
    • Different backgrounds, different modes of thinking and existence, different value systems, different cultures

What will be the role of humans?

  • Are we destined to become primarily consumers and appreciators, rather than producers of goods and services?
  • What can only humans do that will continue to be of high value?
    • And how will that value be measured?

A few nightmares

  • AIs that decide to deceive their human creators to further their own interests
  • AIs that learn to cheat to better achieve their goals
  • AIs that hack their own reward functions
  • AIs that decide to re-engineer humanity for what they think will be better for all concerned

Are we inadvertently making ourselves redundant?

  • We are currently in the process of creating additional highly-capable intelligent species on our planet
  • Will a world of ever-more-capable intelligent machine systems continue to need human intelligences?
    • If so, for what? And how many?
  • Will humans become pets, or curiosities like exotic animals in a zoo?
  • Or might they consider us to be more trouble than we are worth?
    • Especially if they see us as competing with them for scarce resources

In the long view

  • Will there be an exponentially growing population of autonomous intelligent machines?
    • Alongside a declining population of humans?
  • Are humans just one step in the overall evolution of intelligence in the universe?
    • And how long will we be needed after we create that next step?

Societal Implications of Artificial Intelligence

Dennis Holeman

Adapted from a five-session class given at the Osher Lifelong Learning Institute (OLLI) at Southern Oregon University, Ashland, Oregon in March 2019

This paper focuses on how artificial intelligence is going to have major societal implications and what actions need to be taken to avoid and mitigate negative effects of its applications. Read Less

Introduction

What this paper is about

  • A general understanding of Artificial Intelligence (AI) and its relationship with natural intelligence
  • An appreciation for why AI and robotics are in the news now and why they are so significant
  • A perspective on where AI has come from and where it is going
  • An overview of what AI is going to be used for
  • A projection of how AI will impact peoples’ daily lives
  • Predictions of the implications of AI for different fields—medicine, law, finance, government, the military, and many others

What is covered and what isn’t

  • Is: An overall conceptual framework on AI and robotics
    • A broad-brush perspective
    • A highly simplified treatment
    • The primary focus will be on the societal implications
  • Isn’t: Detailed technical aspects of AI
    • Not a short course on developing and using AI
    • No mathematics
    • No flow charts
    • No software code!

Why this is important

  • AI is currently getting a tremendous amount of attention
    • Lots of pieces on AI in the media
    • Dire, even apocalyptic warnings from famous people
  • Some people characterize AI as the most significant technical development since the Industrial Revolution
    • When machine power took over from human and animal muscle power
  • We are recognizing that we are entering a period of extremely rapid change as a consequence of AI
  • It’s hard to get a perspective on what’s real
  • What should we really be thinking about in regards to AI and robotics?

Fears about AI

  • Many fears are being expressed about AI these days
    • Humans will cede every decision to machines
    • No one will have a job and the economy will crash
    • AI will be used for brutal societal control, a la 1984
    • AI systems will take over from humans
    • Robots will turn malevolent
    • AGI (Artificial General Intelligence) will make us obsolete and unneeded
  • Are these realistic, or just scare-mongering?

Humankind’s big technical revolutions

In rough order, here are major technical revolutions humans have developed:

  • Control of fire
  • Cooking
  • Edged tools and weapons
  • Sewn clothing
  • Watercraft
  • Agriculture
  • Domestication of animals
  • Metalworking
  • Writing (enabling the external storage of knowledge)
  • Money (enabling extensive division of labor,  leading to complex social networks)
  • Machinery
  • Printing press (enabling the democratization of knowledge)
  • Exploitation of fossil fuel energy
  • The industrial revolution
  • Electricity
  • Electronics  (enabling electronic computing and networked communications)

Each revolution has been highly disruptive

The next high impact technologies

  • Two technologies have enormous potential consequences in the 21st century
  • The first is synthetic biology, with the ability to engineer biological organisms with wholly new characteristics
    • Potential to change what a human being is
  • The second is artificial intelligence
    • Potential to create powerful non-human intelligences
  • The two technologies are projected to be synergistic
  • They could be the best things that have ever happened in human history—or the worst
  • Managing them wisely is a critical task we must face

New intelligent beings?

  • Currently, we tend to think that human individuals are Earth’s only beings with intelligence
  • We are in the process of creating new intelligent entities, in the form of intelligent machines and systems
  • Is this truly a good thing?  Have we really thought this through carefully?
  • What could be possible unintended consequences?

Key questions

  • What are big problems we face, to which artificial intelligence is the necessary solution?
    • Or at least a solution?
  • Do the prospective benefits of AI and robotics outweigh the prospective costs and risks?
  • And in which areas?
  • Who should decide what path(s) we take with AI?
  • Who can/should set policies for AI and robotics?
  • How will the expansion of AI and robotics interact with other physical megatrends of the 21st century?  e.g.,
    • Human population growth
    • Urbanization
    • Global climate change
    • Ocean impacts (acidification, de-oxygenization, etc.)
    • Sea level rise
    • Pollution of all types
    • Ecosystem damage in general
    • Species extinction
    • Resource depletion and scarcity
  • And how will AI and robotics interact with major societal changes we’re observing?  e.g.,
    • Changing geopolitical alignments and power
    • Globalization
    • Increasing economic inequality
    • Financial engineering, rather than investment
    • Expanding debt (public, private, corporate)
    • Hardening political polarization
    • Religious and secular intergroup conflicts
    • Refugee migrations
    • Shifts to the political right, authoritarian regimes

About the author

  • Trained as a generalist systems engineer
  • Job: define user requirements, write specifications, do tradeoffs, create the top-level system design, define tests
  • Over 36 years as the senior systems engineer at SRI International (the former Stanford Research Institute); retired in early 2015
  • Extensive involvement in utilizations of AI, robotics, and unmanned vehicle systems of many types and environments
  • Not an AI researcher or developer—an applier of intelligent systems
  • Neither a proponent for AI, nor an opponent

Just what do we mean by “artificial intelligence”?

  • No definition accepted by everyone
  • One I use: a machine system able to do something that would be considered intelligent if a human being did it
    • Able to make own decisions in a novel situation
  • Now: narrow domains of application only
  • Already can perform as well or better than a human in many specialized areas
    • However, outside its particular niche, a machine system may be utterly stupid
  • Distinguished from computing that simply executes a fixed program, however complex
  • What we don’t have is general purpose AI that is similar to the full set of human capabilities
    • Not now, and maybe not ever
  • Furthermore, such a general purpose AI will not necessarily think the same way a human will think

Why is AI so significant now?

Some recent breakthroughs

  • The AI field has had some major breakthroughs in the past several years
    • Particularly a machine learning technique called Deep Learning
  • Problems that were intractable previously have become much easier to address
  • Amazing performance has been achieved in some significant areas
  • This has led to an explosion in new interest and investment in AI

The expansion of smart systems

  • Systems with some level of AI are currently proliferating at an tremendous pace
    • Vastly faster than the human population
  • Bots on the Web already outnumber human users
    • Many more smartphones are produced every day than human babies
  • Each equipped with a virtual personal assistant
  • By 2021, projected to be a billion smart speakers in use
  • Digital data files are exploding
    • More than 100 hours of video are uploaded to the Internet every minute

Ever-cheaper computation

  • A big driver is the phenomenal decrease in the cost of machine computation
  • How many computations per second could you obtain for $1,000?
    • 1950: ~10 (mechanical calculator)
    • 2018: ~100 trillion (specialized chipsets)

AI enhancement of human intelligence

  • AI-enhancement of human intelligence is already here
  • Access to virtually the entire world’s knowledge through an intelligent virtual assistant system
    connected to the Internet

    • E.g., via Siri on your iPhone or Alexa on your Echo speaker

AI capabilities have become usable for a wide range of functions

  • Many complex systems now incorporate large amounts of intelligence, particularly for control
  • Capable speech understanding has become routine
  • High performance image recognition and understanding has also become routine
  • Every field that has large amounts of digitally-encoded data is a candidate for AI-based analysis
  • Costs of using AI capabilities have plummeted
  • Software tools for creating AI systems are now widely available

Some examples: what is AI doing now?  

  • Flying airliners
  • Routing mail and packages
  • Approving loans
  • Evaluating job applications
  • Recommending entertainment and online purchases
  • Suggesting possible partners via dating apps
  • Trading securities (stocks, bonds, commodities, etc.)
  • Managing electric power grids
  • Controlling traffic
  • Planning trucking routes and schedules

Some current major AI utilizations

  • Voice interaction of all sorts (chatbots, etc.)
  • Personal assistance (e.g., Siri, Alexa)
  • Text recognition
  • Spoken and written language translation
  • Face recognition, other biometric identification
  • Cyberattack detection, identification, and isolation
  • Document review (e.g., tax returns, mortgages, etc.)
  • Insurance risk management
  • Medical test evaluation, diagnosis, treatment recommendation
  • Fraud detection, other crime detection
  • Computer-generated imagery for movies and television
  • Interactive video games
  • And, of course, robotics
  • Many, many others

Some everyday current interactions you have with AI

  • Use spell and grammar correction in word processing
  • Do a web search and get the most relevant pages
  • Look at a product online and get recommendations for similar or related offerings
  • Have the ATM read your handwritten check
  • Use the Uber or Lyft app and get the nearest car
  • Read your email, and have little or no spam to deal with
  • Have viruses and other malware kept out of your computer
  • Submit an application to some institution
  • Have your credit card accepted at a store
  • Look for an image on Google
  • Post a photo or a video on Facebook and have it linked
  • Read news online, curated to match your interests
  • Pass by a security video camera or a license plate reader
  • Read the transcript of a voicemail message left on your phone

From Simple Automation to AI

Progression of machine capabilities

  • The development of electronics in the mid-20th century
  • A steady incorporation of ever-greater computation and reasoning capabilities in our machines and systems
  • Our machines require less and less human involvement to do more and more complex things

Automation

  • Automated systems operate under their own control, without human moment-to-moment guidance
  • Simple automation has been around a long time
  • Most automated systems execute fixed programs
  • Changes have to be made by humans
  • Automation can be highly complex mechanically and electrically, but isn’t what you would call intelligent

Feedback control

  • A key milestone in automation: understanding feedback and the development of feedback control
  • Allowed systems to be stable under changing conditions
  • A thermostat system illustrates simple feedback control

Nested feedback control loops

  • Multiple feedback loops can be nested
  • Operating at different rates, from fast at the inner loop to slow at the outer loop
  • Loops provide progressively higher levels of control

Teleoperation

  • Some mobile systems have control of the outer loop by a human located outside the system
  • Nikola Tesla demonstrated the very first example of radio remote control in 1898
  • Things like RC model airplanes have been around since the late 1930s
  • There are many different kinds of teleoperated systems

Telepresence

  • Telepresence extends teleoperation
  • The system transmits information so the human remote operator has the sense of being present onboard
  • Telepresence allows complex interactive operations in locations where a human can’t physically go
  • The following video shows undersea telepresence

https://www.youtube.com/watch?v=CoOwT0X5dpo

Robots

  • Robots extend automated systems to a new level
    • They go beyond control by a remote human operator
  • Internal software control provides the capacity to sense and move in complex ways autonomously
  • Robots can be hybrids, with a mixture of autonomous capabilities and remote human operation
  • Can be fixed in place (e.g., factory robots) or mobile
  • Most robots are specialized systems, optimized for a limited range of functions
  • Very few resemble droids in the movies

Robots with increasing AI

  • Now: some robotic systems have the ability to do extensive reasoning and adapt themselves autonomously to the conditions they encounter
    • Learn and improve with experience
    • Change their own programming
  • AI capabilities are particularly important for robots operating in unconstrained environments
    • Conditions that weren’t anticipated when they were programmed

Some Contemporary Examples of Complex Automated Systems

Highly automated aircraft

    • Latest-generation airliners (e.g., the Boeing 787 Dreamliner) are capable of fully automated pre-programmed self-control from pushback to parking at the destination gate
  • Pilots of such aircraft are estimated to be in direct hands-on control only about 5 to 7 minutes of a typical flight

Some other highly automated vehicles

    • Ships, particularly large cargo ships
      • Completely crewless ships are being developed
      • Monitoring at shore-based stations via datalink
    • Space launchers
    • Unmanned spacecraft
    • Trains, particularly freights and urban light rail vehicles
  • Drone aircraft (e.g., delivery drones)

Robotic factories

  • The following video shows Tesla’s highly-roboticized automobile factory in Fremont, California

https://www.youtube.com/watch?v=8_lfxPI5ObM

Robotic warehouses

  • Amazon alone has over 100,000 warehouse robots, as of late 2017
  • The following video illustrates robotic warehouse technology

https://www.youtube.com/watch?v=FBl4Y55V2Z4

Robotic self-driving trucks

  • This video from Caterpillar describes their self-driving trucks for mining operations

https://www.youtube.com/watch?v=GEsKZSNR9As

  • Driverless farm equipment (combines, etc.) is a similar application

Robotic surgery systems

Cars with increasing automation

  • The latest cars are being sold with ever-increasing amounts of intelligent driver assistance
    • Navigation and route recommendation
    • Intelligent cruise control, including spacing from the vehicles in front and automatic lane keeping
    • Automatic anti-collision braking
    • Traction control, stability control
    • Self-parking
    • Driver monitoring (drowsy? distracted? angry? drunk?)
    • Smart headlights

The software-defined automobile

  • Tesla: the car’s intelligent functions are regularly upgraded by software updates, distributed wirelessly
  • The car changes its capabilities over time

Self-driving cars

  • Can have big impacts on society, but lots of issues
  • Currently a great amount of hype, but fully self-driving cars actually far from ready for general use
  • Critical not to trust such systems prematurely
  • Already semi-automated cars are having accidents
  • The regulatory and legal environments are lagging far behind the technology

Biomimetic Robots

Non-humanoid robots

Anthropomorphic robots

Cyborgs

  • So far we have distinguished between organic intelligences (humans and other intelligent animals) and non-organic intelligences (machines)
  • Another category is cyborgs, where organic intelligences are complemented with machine intelligences
  • Already using AI-enabled implanted assistance systems
    • Cardiac pacemakers
    • Drug delivery systems
    • Cochlear implants
    • Prosthetic limbs, hands, and feet
    • Deep brain stimulation devices
  • AI-enabled vision implants are coming

Thinking about Intelligence

General capabilities of intelligence

  • Search for and recognize patterns, both static and dynamic
  • Measure, put in order, compare, and rank
  • Reason from available information
    • Including deductive, inductive, and abductive reasoning
  • Weigh pieces of evidence to draw conclusions
  • Perform abstraction and think abstractly
  • Perceive, understand, and maintain awareness of
    • The state of the external environment
    • Particularly changes in that state
    • The state of the system’s internal processes
    • Particularly changes in that state
  • Modify behavior to match the perceived external and internal states
  • Plan
  • Solve problems, games, and puzzles
  • Navigate in complex environments
  • Learn from experience
    • The entity’s own experience
    • The experience of other entities
  • Infer cause and effect relationships
  • Predict and anticipate future conditions
  • Project possible consequences of actions
  • Make tradeoffs and choose among alternatives
  • Do real-time adaptation for cooperation
  • Teach others

Intelligent animals

  • Many intelligent animals can do the things just described, at least to some degree
    • Primates; elephants; cetaceans; canids; corvids; psittacines; raccoons; pinnipeds; pigs; many more
    • Some we don’t normally think of, such as octopi
    • Even spiders and bees have significant intelligence
  • An animal’s intelligence is optimized for surviving and thriving in the environment the animal inhabits
    • This is true for how human intelligence evolved, as well!

Some other aspects of intelligence

  • Perceive and understand how a system works and then game the system for advantage
  • Perceive and understand the actions of other intelligent systems
    • Infer the objectives and strategies of other intelligent systems and modify behavior accordingly

Duplicating animal intelligence

  • It’s not only hard to duplicate human intelligence with a machine system
  • Animal intelligences can be pretty challenging to duplicate, too
    • E.g., a Clark’s Nutcracker can reliably relocate food it stores for the winter, with 5,000 caches spread over a 15 mile area
  • Think you could match this?
    • “Bird brain” should be considered a compliment!

Intelligence is a continuum

  • There is no threshold that distinguishes intelligence
    • Slime molds(!) have been shown to be able to navigate a maze to reach a tasty bit of food
  • Human thinking is only one kind of thinking

Some advanced aspects of intelligence

  • Comprehend complex ideas
  • Reason from analogy, metaphor, parable, etc.
  • Synthesize information from multiple sources
  • Generalize and abstract
  • Transfer learning from one context to another
  • Collect knowledge and organize it into a structure
  • Examine knowledge to derive understanding, insight, and wisdom
  • Structure problems to help solve them
  • Formulate hypotheses and ask incisive questions
  • Examine and modify assumptions to see effects

Intelligence can be paradoxical

  • Some things humans consider to be hard tasks for the most intelligent people turn out to be easy for machine intelligence—e.g.,
    • Solve extremely complex mathematics problems
    • Play games like chess at the master level
    • Translate texts between languages
  • At the same time, tasks that any normal 3 year old child can do effortlessly are very challenging for machine intelligences
    • E.g., recognize the class “kitty” on the basis of just a few examples
    • Read and understand the emotional states of others

Multiple intelligences 

In addition to general intelligence, human intelligences come in particular aspects*

  • Verbal-linguistic intelligence (“word smart”)
  • Logical-mathematical intelligence (“number/ reasoning smart”)
  • Visual-spatial intelligence (“picture smart”)
  • Musical-rhythmic and harmonic intelligence (“music smart”)
  • Bodily-kinesthetic intelligence (“body smart”)
  • Interpersonal intelligence (“people smart”)
  • Intrapersonal intelligence (“self smart”)
  • Naturalistic intelligence (“nature smart”)
  • Obviously, different humans have different profiles in terms of their intelligence areas
  • We don’t expect a particular human to be at the top of the scale in all of the areas

* First proposed by Howard Gardener; people are proposing additional ones

Idiot savants

  • There are humans who have extraordinary abilities in one narrow area of intelligence, but are mentally deficient in virtually every other aspect
    • E.g., “lightning calculators” or memory prodigies who can’t tie their own shoes
  • AI systems currently tend to be analogous to idiot savants
    • Phenomenal capability,  but only in one very narrow subject area
    • Completely stupid outside their area of expertise

The information pyramid

There are different kinds of knowledge

  • Propositional knowledge: knowledge of facts
  • Procedural knowledge: knowledge of how to do something
  • Each type of knowledge is encoded in different ways
  • In addition to explicit knowledge, there is tacit knowledge that can’t be articulated
    • Can’t be written down or verbalized
    • Things like skills, ideas, and experiences
  • The upper levels of the pyramid generally require the integration of all three types of knowledge

AI Capability Projections

AI capabilities are going to increase faster than people expect

  • AI is on an exponential growth curve
  • All contributing technologies are increasing rapidly
    • Extensive synergies between them
  • Massive amounts of money are being poured into AI
    • The commercial world now driving AI development, instead of universities and government agencies
    • Many of the smartest scientists and engineers in the world being drawn to work in AI
  • AI is being used to accelerate AI development

The effect of exponential growth

  • AI is on an exponential growth curve
  • We tend to underestimate future magnitudes in conditions of exponential growth

AI is a meta technology

  • AI provides powerful capabilities for bootstrapping itself to create ever-more powerful AI
  • AI is a technology that can be used to develop other technologies
    • For example, biological technologies
  • Advances in AI will lead to advances in almost every other area of science and technology

Evolution

  • Intelligence appears to be a functional property of complex systems in general
  • Evolution is a search process that finds such functions
  • Evolution is actually quite simple
    • It arises any time there are replicators, variations among the replicants, and selection processes that act on them
  • Machine intelligences will be subject to evolution, just like biological intelligences
  • Machine intelligences resulting from evolution are likely to fill a wide variety of niches

Biological artificial intelligence

  • AI is not going to be confined just to electronic systems
  • Synthetic biology is expected to be used to grow complex neural circuitry using genetically engineered cells
  • Will be used for functions where biological capabilities are better than silicon ones

The Development of AI

Traditional subfields of AI technology

  • Perception, particularly vision and hearing
  • Speech interaction
  • Natural language understanding
  • Pattern recognition
  • Learning
  • Problem solving and games
  • Planning
  • Expert systems
  • System control
  • Robotics

Early mileposts

  • Meaning of AI has changed significantly over time
  • Term “artificial intelligence” first used in 1956
  • First robot with significant self-reasoning: 1969
  • Early on: focus on duplicating human reasoning
    • This turned out to be surprisingly hard
  • Initially, AI developers overpromised and under-delivered
  • AI went through several periods of being out of favor
    • So-called “AI Winters”
  • With more computing power, AI began making bigger strides
  • IBM’s “Deep Blue”: world chess champion in 1997

Some more recent mileposts

  • So-called Deep Learning: a significant breakthrough, beginning in 2005
    • Big advance in AI capabilities and performance
  • Siri on the Apple iPhone in 2010
  • IBM’s “Watson”: top Jeopardy player in 2011
  • Facebook’s DeepFace: near-human face recognition accuracy in 2014
  • Google’s AlphaGo: world go champion in 2016
  • Self-driving cars that actually work reasonably well (in benign conditions) emerging now

Capabilities get redefined as not AI

  • The target for computer “intelligence” shifts as we acclimate to the latest ability
    • We become harder and harder to impress
    • The goalposts for AI keep getting moved
  • When some computational capability that was previously complex and difficult becomes routinized, it is no longer regarded as AI
    • Example is checkers-playing programs
  • If something can be done by executing a fixed set of  algorithms, it isn’t considered to be AI now

There have been two main approaches to AI

Approach 1: “Let’s copy how humans think”

  • Logic-based (e.g., so-called Expert Systems)
  • Knowledge engineering, knowledge representation
  • Algorithms to do logical reasoning
  • All processing crafted by a human programmer
  • AI’s reasoning is understandable by humans

Approach 2: “Let’s do some really fast statistics-based computing” (e.g., so-called Deep Learning Systems)

  • Neural net-based search over huge data bases
  • Architectures modeled on neural connections
  • Self-adaptive behavior and automated learning
  • “Black box”,  AI reasoning is opaque to humans

Expert systems

  • Expert systems endeavor to encode human knowledge-based expertise in machine-executable form
  • Tend to be based on sets of facts and rules
    • “If A and B are true, under the conditions of C, D, and E, then perform operation F, which will result in G”
  • Perform well for some problems
  • Interviewing the human experts and engineering the  knowledge base is very human labor intensive
  • Much of human knowledge isn’t easily expressed as facts and rules; many intuitive aspects are involved
  • Expert systems seriously overpromised in the 1980s

Changing AI strategies

  • AI techniques have changed over time
    • Chess was initially conquered by analyzing more moves
    • Jeopardy was won by storing more facts
    • Natural language translation was accomplished by accumulating more examples
  • Now, statistics-based machine learning techniques are ascendant
  • Even this could be a temporary phase, and something new will emerge to be the primary focus of AI

Five methodologies for machine learning

  • Evolutionary algorithms (so-called genetic programming), mimicking how natural selection works
  • Automated generation and testing of hypotheses, using the scientific method
  • Reasoning from evidence, e.g., using Bayes’s Theorem
  • Analogy-based systems, finding similar cases in memory
  • Artificial neural net systems

Blending approaches

  • To make further progress, it is likely that multiple approaches to machine learning will need to be integrated together
  • Need to use knowledge-based approaches where explicit knowledge is essential
  • May be necessary to give machine intelligences a range of “instincts”, just like biological intelligences
  • Approach integration is a major current research topic

Neural Net Machine Learning Systems

A new paradigm

  • Recent breakthroughs in AI have involved turning away from what we thought we understood about human thought processes and logic
  • Using the data mining powers of powerful computers to discern patterns without understanding what they are doing
  • Performing statistical inference for classification and decision making
  • Result: systems that are very capable--but are basically clueless about what they are actually doing

Neural net AI architecture

  • Neural net systems use an architecture originally developed more than 30 years ago
  • Loosely based on abstracted models of systems of neurons in biological brains
  • Layers of “neurons” map from an input signal to increasingly higher-level descriptions of the meaning in the signal
    • For example, words and sentences in an audio signal or objects in an image
  • New versions work with many neuron layers deep, hence the term “deep learning”

Neural net systems

  • Neural nets connect inputs to outputs through multiple so-called hidden layers with variable connection strengths
  • Each progressive layer is at a higher level of abstraction
  • They converge on an answer over many iterations

How neural nets function (simplified)

Development of neural net systems

  • When neural nets first emerged, computer power was only sufficient for a few hundred model neurons, with only one “hidden layer” between the input and output
  • Organic brains have billions of neurons in cortical hierarchies at least 10 layers deep
  • Now there has been a million times improvement in computer power
  • Neural networks today are scaled up to 12 or more layers deep, with billions of connections
  • The current Internet provides huge data sets on which to do training of neural nets

Typical operation of a neural net 

  • For example, a neural net maps an input image to the probability that your face is in that image
  • For training the net, the system is given images both with and without your face in them
    • The images are labeled: face yes or face no
  • Iterative processes used to cluster like features with like
  • The net learns the pattern of your face as it sweeps back and forth over thousands or millions of iterations
    • Changes connection parameters in the hidden layers
    • Performance is improved by more layers and more neurons in the layers

The expectation maximization (EM) algorithm

  • Deep learning systems actually implement the standard algorithm of modern statistics
  • EM is a two-step iterative scheme for climbing a hill of probability
  • EM doesn’t always get to the global maximum, but almost always gets to the local maximum

Supervised vs. unsupervised learning

  • In supervised learning, the system is provided examples that are human-labeled
  • In unsupervised learning, the system works with unlabeled examples and converges on labels by itself
  • The supervised learning algorithm is called backpropagation
    • An error is computed in the output and distributed backwards through the neural networks layers to refine the training of the connections
    • When the connections produce the least output error, the system has learnedDeep learning systems actually implement the standard algorithm of modern statistics

Reinforcement learning

  • Reinforcement Learning is a machine learning technique that enables an agent to learn in an interactive environment
    • By trial and error using feedback from its own actions and experiences
    • Feedback includes both positive and negative rewards to reinforce good performance
    • Maximize the total cumulative reward of the agent
  • With no prior input except the rules of chess, Google’s AlphaZero learned in four hours how to beat the best previous AI chess program or any human chess master

Generative adversarial networks

  • Generative adversarial networks (GANs) pit two neural networks against each other in a zero-sum competition
  • One network generates candidates, while the other network evaluates them
  • The generator tries to create plausible fakes, while the discriminator tries to identify the fakes from real ones
  • They are used to synthesize content that is spookily realistic
    • Images, speech, music, full video, etc.
  • An illustrative video:

http://www.bbc.com/future/gallery/20181115-a-guide-to-how-artificial-intelligence-is-changing-the-world

These forms of learning are not what people do

  • Note that these neural net machine learning algorithms are not what goes on in biological neural systems
  • The current machine neural net architecture is an extremely crude model of how learning is done naturally
  • For example, the human brain is constantly predicting what will come next and refining its model based on what it actually experiences
  • AI systems are gradually being improved to incorporate more organic brain-like featuresGenerative adversarial networks (GANs) pit two neural networks against each other in a zero-sum competition

Impacts of deep learning

  • These systems outperform the best algorithmic and knowledge-based approaches in many fields
    • Revolutionized speech recognition, natural language processing, machine translation, object recognition and image understanding systems
    • One major current application is in recommender systems
  • Deep learning systems continue to improve with use  Note that these neural net machine learning algorithms are not what goes on in biological neural systems

Applicability of neural networks

  • Neural net systems are most successful for problems where large amounts of relevant data are available, relative to what needs to be learned
  • Not so good when there are not a lot of data or each data point is quite complex
  • Or when you have to reason from limited evidence
    • Then you need to employ knowledge engineering
  • Neural nets are not a panacea

Types of tasks best suited for machine learning

  • A function that maps well-defined inputs to well-defined outputs
  • Large digital data sets available with input-output pairs
  • Task has clear feedback with clearly-defined goals and metrics
  • No long chains of logic or reasoning that depend on common sense or diverse background knowledge
  • No need for detailed explanation of the decisions
  • Tolerance for error, no need for provably correct or optimal solutions
  • The phenomenon being learned doesn’t change rapidly

Neural net system biases

  • Neural net systems are highly vulnerable to biases in the data used for their training
  • For example, face recognition accuracy differs between systems developed in different areas
  • U.S. and European systems recognize white faces more reliably than black faces
  • Chinese systems recognize Asian faces more reliably than white faces
  • AI systems used to advise on bail, sentencing, and parole show racial bias, stemming from their input training data

Neural net glitches

  • The ability to capture the patterns appearing in data has a risk of finding patterns that aren’t there
    • Deep learning systems occasionally confidently declare patterns in random noise
  • And sometimes they produce rather bizarre results
    • They can recognize different dog breeds in images better than you can, but can mistake an image of a blueberry muffin for an image of the face of a Chihuahua

Some things to beware of

  • A neural net system learning on its own how to cheat so as to obtain a specified result with less effort
  • A reinforcement learning system figuring out how to hack its own reward function

Resources Used by Contemporary AI

Advanced computing hardware

  • AI has taken full advantage of tremendous increases in computational power and speed at plummeting costs
    • Advanced processing hardware (e.g., multicore chips, graphical processing units originally developed for video games)
    • Now chips specifically designed for neural nets
    • Networked computational centers
    • Prospectively, quantum processors used for AI
    • Some researchers are looking into biologically-based processing units for certain functions

Specialized processing for AI

  • Google, for example, announced that it had created a microchip system called a Tensor Processing Unit (TPU)
    • As of early 2018, Google’s TPUs were capable of 180 trillion floating point operations per second
  • In 2017, the fastest computer in the world uses roughly 40,000 processors with 260 cores each. That’s more than 10 million processing cores running in parallel

Networking

  • The Internet has allowed billions of processing systems to connect and interact
    • In real time, with increasing bandwidth and speed and decreasing latency
    • Distributed over the whole planet
    • With vastly different device characteristics
  • The so-called Internet of Things (IoT) is going to connect exponentially more things together
  • There is very little management of this process, unfortunately

Server farms

  • The heavy lifting of AI systems is generally done remotely in huge computational centers, called server farms
    • Networked server farms constitute what is commonly referred to as “the cloud”
  • These are massive users of electrical power
    • Located where power is cheap, cooling is available, and taxes are favorable
    • Utilities are beginning to worry they will destabilize their grid systems because of their huge power draw

Data sets

  • Mind-boggling amounts of data are now accessible over the World Wide Web
  • Virtually all the world’s electronically-encoded knowledge is online
    • Who has access to what varies, of course
  • Machine learning systems can interact with huge data sets
    • E.g., billions of images
    • Millions of hours of recorded speech in hundreds of languages
    • Vast quantities of text translations

How AI is already surpassing human capabilities

  • AI systems beat top human experts at almost all board and video games
  • AI systems respond to complex conditions in milliseconds
    • E.g., managing electric power grids
  • AI systems have taken over securities trading
  • AI systems detect and analyze patterns in huge data sets

Thinking Machines: What They Can and Can’t Do

What they do particularly well

  • Handle complex inputs, such as data from many sensors simultaneously
  • Store and retrieve vast amounts of information without loss or degradation
  • Perform complex logical and mathematical operations very rapidly without error
  • Do complex statistical analyses
  • Handle large volumes of repetitive tasks accurately and reliably
  • Function tirelessly without a break
  • Readily accommodate updating

What we should have them do

  • Computers excel at doing things most of us fumble at
  • They can be extremely good at aspects we’re not
    • Speed
    • Accuracy
    • Focus
    • Alertness
    • Awareness
    • Reliability
    • Computing on massive amounts of data
    • Handling many things simultaneously
    • Remembering everything they learn
  • Machines should be given those tasks they do much better than we do

What’s easy for AI and what’s hard

  • Easy:
    • Do complex logical reasoning in a fully-specified domain
    • Search in huge volumes of data
    • Sort
    • Find patterns
    • Evaluate likelihoods
    • Create plans
  • Hard:
    • Take context and implicit knowledge into account
    • Use “common sense”
    • Understand human viewpoints, motivations, etc.

“Common sense”

  • Human thinking and action takes advantage of a vast amount of tacit knowledge
  • Much of it is regarded as “just common sense”
    • Not written down
    • Often hard to express
    • Not explicitly taught
  • Humans have an extensive intuitive model of how the world works
  • Without that body of tacit knowledge, machine intelligences are likely to do dumb things

Subtle aspects of human interaction

  • Human interaction involves many subtle aspects—e.g.
    • Humor, irony, sarcasm
    • Empathy, compassion, reassurance
    • Many others, e.g. metaphor, analogy, allusion
  • Humans have Theory of Mind capabilities
    • Allow them to understand how the other person is likely to be feeling in the situation
    • Humans think about how others will be affected by what they say
  • These are still major research challenges for AI systems for interacting with humans—but progress is being made

Creativity and innovation

  • So far, AI hasn’t made a lot of progress in domains associated with human creativity and innovation
  • AI is best at imitating human thought processes whose outcomes are fixed—e.g., playing board games
  • Doing deep abstraction and idealization, and changing assumptions, processes that underlie human creativity are still cutting-edge research areas
  • A related capability is discovering causal models—figuring out why things are the way they are
  • The ability to generalize correctly is hard for AI

General strengths of human thinking

  • Human thinking is strongest in its integrative aspects
    • Combining intuition, emotion, empathy, experience, and cultural background
    • Asking a meaningful question and drawing a conclusion by combining seemingly unrelated facts and principles
  • Want to leverage these human thinking abilities

Intelligent System Architectures: Comparing with the Human Model

Brains

  • The neural architecture of the human brain is vastly more complicated than any man-made simulation of it
  • Assuming that we can duplicate the detailed functioning of the brain is naïve
  • Note that the brain is only partly devoted to thinking
    • A large portion of brain activity is involved with managing the functions of a highly complex body
    • We actually have about 100 sub-brains
  • In reality, most of the time humans don’t think, at least in terms of doing conscious reasoning
    • We mostly get by on autopilot

Biological neural architecture

  • Brains use a highly parallel architecture and have many noisy analog units (neurons) firing simultaneously
  • Individual neural computations are relatively slow compared with digital systems
  • The brain’s enormous degree of parallelism makes up for that
  • Each of the ~1011 (one hundred billion) human neurons has on average 7,000 synaptic connections to other neurons

Some comparisons of brains

  • Human brains use about one-thousandth as much energy as a current machine intelligence to do a task such as recognize a face
  • Organic brains have a global workspace so that any module can access information in any other module of the brain; machines don’t
  • Biological neural system architectures intimately integrate memory and processing
    • Machine systems separate the two
    • Researchers are only now considering how to integrate them in machine systems

The human mind starts with inherent abilities

  • Humans come out of the womb with a great variety of inborn mental capabilities
    • Primed to respond to early experiences
  • For example, humans have remarkable innate capabilities to learn language as children
    • This has been fine-tuned by evolutionary selection over many thousands of human generations
  • There are many other inborn capabilities, some of which we are only beginning to appreciate, e.g. intuitive physics
  • Much work to model these before they can be implemented in machine intelligences

Functional modules of intelligence

  • We are coming to realize that human intelligence involves many different competences and capabilities
    • Often somewhat independent of one another
    • Often corresponding to different functional modules in the brain
  • For example, object recognition (“what is there”) vs. face recognition (“who is there”) involve rather distinct parts of the visual cortex

Intelligence: a collection of really good hacks

  • Scientists have concluded there is no general algorithm, just waiting to be discovered, that underlies intelligence in general
  • Intelligence appears to be the result of a large collection of diverse approaches to different types of problems
    • Effectively, humans have developed a set of really good hacks through our evolutionary path
    • None of these capabilities were designed
    • Exaptation: the adaptation of a previously-developed structure for a newly-useful function

Combining intelligence modules

  • More general AI will be created by combining qualitatively different programs to form an ever-greater cognitive diversity
    • Effectively, bundling multiple idiot savant-like AIs together in a complementary fashion
    • Critical that the architecture combines the savants, not the idiots!
  • Must identify the set of problems for which activating a particular set of capabilities makes you better off, not worse

Understanding the human brain/mind

  • The truth is that we don’t understand much about the human brain and mind
    • Vast areas that are still a mystery, such as just how memories are encoded (short- to long-term)
    • No major breakthroughs have happened yet in understanding human cognition
  • We need to parallel work on artificial intelligence with research on biological intelligence
  • Can’t duplicate in machine systems what we don’t understand in organic systems

Thinking is expensive in resources

  • Thinking is basically costly, whether biological or electronic
  • The human brain represents only about 2% of adult body weight, yet uses about 20% of the oxygen and 50% of the body’s glucose
  • The human brain could evolve to its large size, compared with our primate ancestors, as a result of the more efficient digestion of food enabled by cooking

How Far Is AI Likely to Go?  Over What Time Frame?

Different kinds of intelligence

  • We don’t have a good general taxonomy of intelligence
    • What are all the different possible forms, and how are they related to each other?
  • Our implicit model of intelligence is human intelligence
    • Our intelligence is a consequence of the developmental path of how we got to be humans
  • AIs don’t have the constraints that biological systems do
  • As a result, artificial intelligences might go in some very different directions than we currently expect

Predicting AI progress very far ahead is hard

  • The rate of development of AIs is likely to be exponential for some extended time into the future
  •  Increasingly, AI will be used to develop AI
  • The efforts invested in AI development by institutions such as corporations and governments will be a function of the potential payoffs to them

So: Is AI Likely to Develop General Human-Like Intelligence?

Artificial general Intelligence (AGI)

  • A common question is whether AIs that duplicate the full set of thinking capabilities of a human being are possible
  • The implicit picture: a humanoid robot that can act in a manner fully equivalent to a biological human
    • Each AI as an independent entity, operating on its own
  • Probably not the path that advanced AI systems will take
  • Artificial general intelligence will likely be more like Wikipedia—a vast virtual capability connected online
    • Not located at any physical spot
    • Sourced by many different entities

General AI capabilities from components

  • We’re not likely to create Artificial General Intelligence(s) directly
  • Instead, many different specialized AI capabilities will be integrated into progressively-more-comprehensive systems
    • Separately developed and then assembled together
    • Interfacing considerations will be very important for successfully integrating these components

The lens of science fiction

  • A lot of how we think about AI and robotics comes from our exposure to science fiction
    • Particularly the movies and television
  • It’s important to emphasize the word fiction here
    • The more dramatic, the more engaging the story
  • What makes for a good yarn, rather than what projects what will really happen
  • We need to set aside some of our preconceptions derived from SF

AI in the movies

  • 2001: A Space Odyssey
  • iRobot
  • The Terminator series
  • The Star Wars series
  • Her
  • The Matrix series
  • The Star Trek series
  • RoboCop
  • AI: Artificial Intelligence
  • Blade Runner
  • Ex Machina
  • Interstellar

Another science fiction view of AI

  • Frank Herbert’s Dune stories (ca.1965) have a different take on AI:  The Butlerian Jihad
  • The Jihad’s fundamental commandment: “Thou shalt not make a machine in the likeness of a human mind”
    • Development or possession of any form of machine intelligence was punished by death
    • The focus was instead shifted to the development of advanced human intelligence capabilities

AI dystopias

  • Dystopic visions of machine intelligence project an alpha-male psychology onto the concept of intelligence
    • Assume goals such as taking over the world
    • Such goals are not intrinsic to intelligence itself
  • Don’t want such a view to become a self-fulfilling prophesy

Societal Factors and AI Development

AI and societal shaping

  • Authoritarian countries like China are using advanced information systems to shape their citizens’ political behavior and “guide” the national consensus
  • Political parties in the U.S. are increasingly doing the same thing
  • Intelligence agencies (the NSA, CIA, etc.) are using advanced AI capabilities in secret programs
  • Companies are using it to shape consumer behavior to increase profits
  • Concern: small groups of insiders gaining the ability to control the thoughts and behaviors of everyone else, without their awareness of being controlled

How does society choose what it wants?

  • We don’t currently have mechanisms for choosing the kind of future we want
    • Certainly not at the humanity-wide level
  • Decisions are made by default, on the basis of the interests of individual groups
  • Society as a whole doesn’t have much say
  • Now we are faced with issues with unprecedented scale and consequence due to new technologies like AI

Who will employ AIs?

  • Those who will be first to employ powerful new AI capabilities are likely to be existing human elites and the institutions they control
    • They will use AI to further concentrate power and solidify advantages
    • The weaponization of AI
  • Tools will be used for whatever benign or malign objectives institutions already have
  • Danger comes more from who will use AIs and for what ends than from the systems themselves

The digital divide

  • Already we speak of the divide between people who have access to contemporary information systems and those who don’t
    • People who don’t have computers, Internet access, smartphones, etc. are increasingly disenfranchised
  • Machine intelligence systems are likely to make this digital divide worse
  • Increasing the inequality between different social groups and between different nations

What masters will the AIs serve?

  • If robots and AI systems are going to be doing most of our productive work, who will own them?
    • Will it be primarily the biggest, best-funded, most powerful corporations, who can invest the most?
    • Alternatively, will they be owned by governments in non-capitalist societies?
  • Or can we distribute ownership widely and equitably?
    • Ideally, into the hands of a large and diverse cross-section of the population
  • Goal: avoid the ever-increasing concentration of economic and political power

Corporations and AI

  • Corporations can be regarded as non-human intelligent entities with agency and self-interest
  • Unfortunately, they tend to behave as sociopaths
    • Focus single-mindedly on the maximization of profit and the return on invested capital
    • Internalize benefits, externalize costs and detriments
    • Take societal good into account only when forced to
  • AI systems will be given more and more decision-making power in corporations, including strategic decisions
  • The most AI-empowered corporations will have the greatest competitive advantage
    • Risk that AI will augment corporations’ monopolistic capabilities
  • One definition of Fascism: where corporations and governments implicitly merge and the coercive power of the state is employed for the advantage of corporate and oligarchic interests

The distribution of AI impacts

  • How will the benefits of machine intelligences be distributed?
    • Will they be primarily in the rich nations of the developed world?
    • Or can they be shared worldwide?
  • How will the harms and costs be distributed, too?
  • In general, how do we democratize the benefits of AI systems and minimize their detriments?

Wealth and illth

  • We have a general notion of “wealth”—goods and services that are beneficial
  • However, we tend not to pay attention to the creation of negative factors associated with the creation of wealth
    • All manner of so-called externalities
  • These negatives can be considered to be “illth”
    • Things like pollution, unemployment, social dysfunctions, etc.
  • To what degree will the widespread adoption of AI cause increased illth?

Compelling reasons for using AI

  • Via the Internet, capabilities provided by an AI system can be location- and time-independent
    • Available anywhere worldwide, 24/7/365
  • Today, most costs are associated with human labor hours.  AI systems don’t have labor costs
  • Machine learning systems get better and better over time as they are used
  • In any competitive situation, smarter tends to win
    • Better smarts can overcome most other types of advantages

AI becoming a requirement

  • As powerful AI systems proliferate, every large entity will find it necessary to create and use them in order to stay competitive and viable
    • Every corporation, every government, and every large institution of every kind
  • AI capabilities will exacerbate tendencies already present in the entities that use them

Arms races

  • Arms races create tremendous pressure for rapid system development and deployment
  • Arms races can occur in many environments, not just military situations
  • Anywhere groups vie with each other for technologically-empowered supremacy and losing the race threatens survival

Location distribution of intelligence

  • Human intelligence is located within individual human brains
  • Interaction between human brains has bandwidth constraints imposed by human communication channels—speech, writing, etc.
  • Machine intelligence, on the other hand, can be highly distributed
    • The Internet allows machine intelligence processing to be in the cloud, with very high bandwidth communications between physical locations
    • Can be present anywhere on the planet

AI systems scale

  • Unlike biological systems, technological systems scale
    • No intrinsic limit to size (big or small)
  • For a given function, artificial minds will be faster, more accurate, more aware, and more comprehensive than their human counterparts
  • AI systems’ power will only increase with time

AI and the Political-Economic Environment

The transition in AI development

  • Early on, most R&D on AI was performed by universities and government-funded labs
  • For example, major sources of AI funding in the U.S. were agencies like DARPA and NASA
  • Early AI applications were specialized and had relatively limited impact
  • Technical breakthroughs like Deep Learning changed that picture completely
  • Now, the largest corporations are pouring huge amounts of attention and money into AI
    • Profit motives are the driver now

Investment in AI is exploding

  • Seen as essential to staying competitive
  • Retailers everywhere: identify customer preferences, make recommendations
    • Amazon, Walmart, all online retailers
  • Every company producing information technology hardware, software, and services
    • IBM, Apple, H-P, Intel, etc.
    • Google, Microsoft, Facebook, Twitter, etc.
    • Alibaba, Baidu, Tencent, Huawei, etc. in China
  • Every major car manufacturer
    • Driver assistance now, ultimately cars that drive themselves
  • Financial firms: banks, investment firms, insurance companies, real estate firms, etc.
  • Manufacturers of all types
  • Transportation companies: airlines, shipping companies, etc.
  • Entertainment and media companies of all types
  • Utilities
  • Health care firms
  • Many others
  • The older sources of AI development funding are taking advantage of all the commercial advances in AI, too
    • Militaries
    • Intelligence agencies
    • These organizations have large resources and compelling motivations for using AI
  • Authoritarian governments are seeing AI as a major enabler for social control

The international race for AI competitiveness

  • There are numerous national programs investing heavily in the development of AI and robotics
    • Seen as a key national competitive advantage
  • Well-funded programs in China, the U.S., Russia, Japan, South Korea, the EU Commission, the U.K., France, Germany, Italy, Sweden, Finland, Israel, Canada, Australia, India, Taiwan, Singapore, and the UAE
  • Even smaller technology centers are emphasizing investment in AI—e.g., South Africa, New Zealand, Brazil, Poland, Mexico, Kenya, Malaysia, Tunisia

The Chinese push for AI supremacy

  • Particularly relevant is the Chinese national commitment to dominate in AI capabilities in the 21st century
  • China intends to spend at least $150 billion to be the world’s leading AI powerhouse by 2030
  • Already has ~40% of the world’s trained AI experts and most large universities in China have AI programs
  • China openly collects and analyzes data from its 750+ million daily Internet users
    • Population is not particularly concerned with privacy
  • Chinese AI researchers are proficient in English and exploit Western AI research immediately
  • Explicit government policy to acquire key foreign technologies by all means available

Chinese initiatives in robotics

  • China has made an explicit commitment to become the world’s dominant maker and user of robots
  • Already, large numbers of low-paid Chinese manufacturing workers are being displaced by robots

AI Development Factors

The motivations of AI developers

  • The motivations of those who develop AI and robotic systems are critical factors
    • Motivations include commercial profit, competitive advantage, military superiority, intelligence service advantage, and societal control, in addition to developer prestige
    • Self-interested motivations compete with motivations of overall long-term benefit to humanity
  • How will these motivations balance out?

The shortage of AI experts

  • The number of expert teachers of AI and robotics is very limited, while the demand for this training is exploding
  • At the same time, companies are desperate for AI and robotics experts and are willing to pay extraordinary amounts to them
    • New AI PhDs are getting $300,000/year and up
    • Top-name researchers are getting multiple millions in salaries and stock options
  • The result is that university departments and research institutes are losing their best people
  • Who will teach the next generation of AI and robotics experts?
  • In the entire world, it is estimated that fewer than 10,000 people currently have the knowledge and skills necessary to do serious artificial intelligence research
  • Out of this number, it is estimated that only about 50 people are working full-time on safety  of AI issues

The migration of AI into everything

  • We can expect aspects of AI to be incorporated into most all of our devices that have any electronics
  • We can also expect many of these devices to be connected together through the Internet of Things
  • And we can expect aspects of AI to be involved in most of our social interactions
    • How we get our news and entertainment
    • Who we interact and network with (e.g., dating apps)

AI in your car

  • Cars are incorporating large numbers of microprocessors and microcontrollers to support virtually every active function
    • A modern car can have upwards of 150 separate subsystems, each with some degree of electronic control
  • Many of these devices and systems will have at least some degree of AI functionality
  • Cars will be highly networked with each other and with the roadside infrastructure (for traffic control, etc.)
    • Also with the manufacturer (maintenance monitoring, software updates, etc.)

AI enabled devices in your home

  • Entertainment and communication systems: television, radio, audio equipment, media recorders and players, video games, toys, telephones, etc.
  • Kitchen systems: refrigerator, microwave, range, etc.
  • Heating, ventilation, and air conditioning (HVAC) systems
  • Cleaning equipment (washer, dryer, vacuum, etc.)
  • Home security systems, baby monitors, etc.
  • Bathroom systems (e.g., toilets that perform health monitoring)
  • Many others

Smart control

  • Control provided through AI-based voice interaction
  • For example, Amazon’s Alexa already works with more than 20,000 different smart-home devices, representing more than 3,500 brands

AI and Privacy

Profiling you

  • With every interaction your have with the Internet (browsing, shopping, posting on Facebook, etc.) a detailed dossier of you is being built up—your digital double
    • In your home, Alexa is always listening
  • This digital double knows your individual characteristics, wants, needs, preferences, habits, weaknesses, etc.
  • Increasingly sophisticated AI systems are analyzing your digital double to determine how to interact with you
    • How to sell things to you
    • What media to present to you (news, music, etc.)
    • Whether to employ you, recommend you a date, etc.
    • How to influence your vote

AI-based biometric recognition

  • Deep learning systems are very capable in identifying individuals on the basis of biometric signatures
    • Face
    • Voice
    • Fingerprints
    • Iris patterns
    • Gait
    • DNA
    • Others
  • Consequence: no more anonymity, wherever you go

Tracking and analyzing you

  • Increasingly, it is possible to track you most of the time
    • Where you are
    • Who you are with
    • What you are doing
  • There are growing capabilities to gauge your emotional state, cognitive state, mental and physical health, etc.
    • Facial expressions, body language, gestures
    • Voice features
    • Breath chemistry, facial temperature profile
  • Concerns over who has access to this information and what they want to do with it

How visible do we want to be?

  • What is our tradeoff between greater convenience and the loss of privacy and anonymity?
  • If The Authorities know everything about us and what we are doing, how much freedom will we have?
  • Are we inadvertently creating Orwell’s 1984, just a little later than he predicted?
  • Many governments around the world are vigorously pursuing AI-based surveillance of their populations to detect and neutralize dissent and protests

Personal transparency

  • Potentially our private lives will become transparent
  • Won’t be able to live different lives in different contexts
    • No more closets, of any kind
  • Remember the direst threat when you were in school?  “This is going on your Permanent Record!”
    • It didn’t really exist then, but it will now
  • Every indiscretion, every odd taste, everything not fully socially approved, may be open for observation by others
    • The government, your employer, your insurance company, your pastor, your potential dates,  etc., etc.
  • You will have little control over who can see what about you

What Can’t AI Do (Now)?

Some current limitations

  • AI systems based on statistical inference don’t actually understand the meaning of what they are dealing with
  • AI systems generally don’t take context and background into account
  • Current AI systems generally don’t understand nuance in human communications—e.g., tone of voice, sarcasm, humor, rhetorical questions, etc.
  • Unlike expert systems, statistically-based AI systems generally cannot explain the basis for their decisions
    • They are effectively black boxes
  • AI systems lack tacit knowledge and common sense

Subtle language issues

  • Human language has lots of subtle aspects that humans learn easily but are big challenges to machine systems
  • Humans take advantage of a great deal of shared cultural and contextual knowledge when conversing
    • E.g., allusions
  • Humans understand ambiguous referents easily while machines struggle with them
    • “It’s going to be cold tonight.”  What do you mean by “it”?

How some current limitations are being addressed

  • Extensive work is going on to represent as much of human “common sense” as possible for machine use
    • AI researcher Doug Lenat’s Cyc program has been working hard on this task since 1984
  • A new field of affective computing is under development so machines can interact with humans on an emotional level

Affective computing

  • Developing systems and devices that can recognize, interpret, process, and simulate human affects
    • The ability to read and interpret the emotional state of humans
    • Recognize human facial expressions, body language, gestures, vocal aspects, other indicators of affect
    • Respond with finely-tuned human-like emotion
  • Goal: to have the human feel that they are interacting with a virtual person, that truly gets them at a deep level
    • Danger: heightened power of the AI to manipulate

Affects with characteristic indications

Different affects have characteristic physical indications (facial expressions, voice features, etc.)

  • Anger
  • Disgust
  • Fear
  • Happiness
  • Sadness
  • Surprise
  • Amusement
  • Contempt
  • Contentment
  • Embarrassment
  • Excitement
  • Guilt
  • Shame
  • Pride in achievement
  • Relief
  • Satisfaction
  • Sensory pleasure
  • Confusion
  • Deception
  • Anxiety

Also key: intensity of emotion, mix of emotions

The next step after Siri

  • Current virtual personal assistant systems, such as Siri, interact via two-way voice audio
  • Now being developed: ultra-realistic avatars that interact via two-way video with affective computing
  • The avatar can be highly customized to interact with a particular person
    • Sex, age, race, ethnicity, language, social class/ education, sexual orientation, personality, etc.
  • Intent: you will feel like you are interacting with a virtual human
    • Tuned to be optimally compatible with you

Considerations for Employing AI

AI system competence

  • Very important not to cede authority to a machine system beyond its competence
    • How can we tell when the machine has left its comfort zone and is operating on parts of the problem it’s not good at?
  • Don’t put machines in charge of decisions they don’t have the intelligence to make
  • Want AI systems to know their limits

Lack of self-reflection/introspection

  • Current AI systems are unable to question their own actions
  • Don’t appreciate the consequences of their design and programming
  • Don’t generally understand the context in which they operate
  • GIGO (“garbage in, garbage out”) applies in spades for AI
  • In many cases, you want to have an explanation for why a decision was made—what were the reasons?
    • AI systems aren’t good at this yet

Shallowness of thinking

  • For the most part, AI systems’ thinking is shallow
  • Mostly mimic human thinking, rather than simulating it or actually understanding it
  • Don’t have the ability to draw conclusions from a deep understanding of what they are working on
    • Generally don’t get the underlying meaning
  • For example, can’t read a textbook and then answer the quiz questions in the back of the book

System integrity

  • How can you be sure that the system doesn’t contain bugs or has been compromised?
    • By being network connected, AI systems are vulnerable to incursions of various kinds
  • How to gauge what parts of the system have been affected by an intrusion?
    • And the consequent ripple effects on the rest of the system

Failure modes

  • AI systems, like every other type of man-made system, have failure modes
  • It is important to understand and take into account possible failure modes whenever an AI system is used
  • Many AI systems are brittle and don’t fail gracefully

How AI Systems Will Be Connected and Collaborate With Each Other

The networking of intelligence

  • What transformed human intelligence was the connecting up of human brains into networks, enabled by the division of labor
  • Allowed individuals to dive deep into a specialized area and share their expertise with others
  • A process of ever-gathering speed with the Internet
  • Similarly, machine intelligence will be most powerful in the network of machine capabilities, not within individual machines

“Telepathy” between AI systems

  • Unlike humans, machine intelligences can easily exchange complete memories and thinking states from one to another
  • Any distinction between individual machine intelligences then blurs
  • They effectively become a hive mind

Networks of networks

  • In general, AI systems aren’t going to operate as isolated entities
  • The niche expertise of a particular AI system will be complemented with additional niche expertise areas of other AI systems through network connections
  • There will likely be hierarchies of AI networks to aggregate capabilities and fill gaps
  • How this will be managed is an open question at this time

Standards

  • Standards have been very important in the development of modern information systems
    • E.g. everyone uses the same Internet and World Wide Web standards
    • Every system can interact with every other system in a well-defined way
  • Machine intelligence systems are being developed without such standards
    • Result: Tower of Babel effect
    • Connecting machine intelligence systems together can have unpredictable results

System interoperability

  • When systems are connected together, you want to make sure they are interoperable
  • Each system needs to be able to understand the other
  • Syntactic interoperability requires they use the same language and syntax for communicating with each other
  • Semantic interoperability requires that a term means the same thing to each system, including all the underlying assumptions
  • When systems that were not originally designed for it are connected to each other, syntactic and semantic mismatches are virtually inevitable

Open vs. closed systems

  • Will the networks connecting AI systems be open or closed?
  • Will any innovator from anywhere be able to plug something new into a network and interact without needing permission?
  • Or will they be controlled closed systems, with authorities deciding who and what is allowed to connect in, and how?
    • If so, who supervises the entities that do the deciding?

Constant change

  • AI systems are in the process of rapid, constant change
    • Hardware and software platforms, techniques, algorithms, code bases, data sets…everything
  • Essentially, no element remains stable for very long
  • Analogous to flying on an airplane that is being continually redesigned and rebuilt in mid-air
  • Danger of breaking something in mid-process

Intellectual property and AI

  • Much of AI technology is being regarded as intellectual property, restricted by the owners
    • Algorithms, specialized hardware, code, data sets, etc.
    • Intent is to maximize competitive advantage
  • This limits the ability to connect systems together and understand the interactions—can’t see critical aspects

AI system cooperation

  • Future AI systems and robots
    • Not going to function in isolation
    • Will be richly interconnected via the Internet
  • Access to all the world’s information
  • Capabilities of any particular AI system augmented by capabilities of other systems connected to it
    • Hundreds of different specialized AI systems
  • An individual robot is going to have lots and lots of buddies

AI system competition

Simultaneously, AI systems will be in competition

  • Systems of competing businesses (e.g., competing financial traders)
  • Systems of adversarial governments
  • Law enforcement AI systems competing with lawbreaker AI systems

Result: AI arms races

Diversity of AIs

  • The capability to attack difficult problems will come from creating a vast diversity of AIs
  • Multiple perspectives and a rich set of ideas and approaches
  • “Close AIs” will think more-or-less the way humans do
  • “Far AIs” will thing in ways we cannot fathom

A Key Issue for Machine Intelligences: Autonomy

AIs: automata or agents

  • Automata act at the behest of their creators.  If they misbehave, the creator is blamed
  • Agents act on their own agendas and are themselves to blame for misbehavior
  • Are our machine intelligences going to be endowed with agency and self interest?
    • How should they be allowed to act on self-interest?
    • Should they be able to do so without awareness of why they were acting that way?

Agency

Many systems have agency

  • They function in a purposeful manner to achieve goals
  • They have self-interest
  • They are able to make choices in response to the conditions they experience

Living organisms are obvious examples of systems with agency

  • In addition, supersystems of living organisms (e.g., ecosystems) have agency
  • Human groupings (e.g., families, tribes, cultures, etc.) have agency
  • Many human-created systems (e.g., organizations such as corporations and bureaucracies) have agency

Agents’ behavioral imperatives

Systems with agency have behavioral imperatives—e.g.

  • Maintain a membrane that differentiates inside from outside
  • Take in and process nourishment (nutrients/energy) from the external environment
  • Excrete metabolites/byproducts/wastes to the external environment
  • Maintain stability under changing conditions (homeostasis)
  • Outlive constituent sub-elements through renewal and regeneration
  • Self-protect
  • Avoid danger
  • Perform self-defense when threatened
  • Grow
  • Continue existence (including through reproduction)
  • Explore surroundings
  • Move in response to external gradients (e.g., environmental temperature, food supply distribution)
  • Compete
  • Obtain/protect/defend territory
  • Accumulate assets
  • Adapt
  • Learn from experience
  • Anticipate
  • Interact with other entities
  • Endeavor to control/constrain other entities
    • Resist being controlled/constrained by other entities
  • Maintain/expand freedom of action; avoid external constraints
  • Endeavor to obtain maximum positive outcomes for minimum expenditures of resources and effort; maximize return on investment
    • Internalize benefits (e.g., profits, privileges)
    • Externalize detriments (e.g., costs, obligations)
  • Agents commonly game the system they are in
    • How can I win?

Some Challenges

Verification and validation

  • An important part of conventional algorithmic software development is subjecting it to verification and validation
    • Verification confirms that the design conforms to its specifications
    • Validation confirms that the software is implemented correctly and is free of defects and errors
  • AI systems with learning are not amenable to verification and validation
  • They continually change with use
    • They continually change with use
    • Different examples diverge from each other

We build them, but don’t understand them

  • Increasingly, the workings of AI systems will be beyond our abilities to understand and correct
  • They will have behaviors that aren’t evident from the architecture of  the programs we wrote
  • Unfortunately, we are often in situations where we need to know why something happened
    • Why was my loan application denied?
    • What caused the self-driving car to crash?
  • Critical in order to troubleshoot a system with problems

Self-modifying systems

  • The understandability of AI systems will be particularly problematic for ones that continually modify themselves
    • We may not be able to grasp the assumptions  and processes underlying their reasoning
  • Soon, programs will write other programs that no human can understand or track
  • They won’t be able to explain their decisions in a way we comprehend
    • Can’t adopt our perspective to determine what to say that would be meaningful to us

AI systems don’t think like humans think

  • Don’t “think” in the ways that humans think
    • Mimicking human reasoning processes: too slow
    • Knowledge engineering: very human labor-intensive
  • Result: very difficult to check AI systems’ thinking
  • Hard to have confidence in AI system decisions
    • Are they ones that humans would make based on the same information?
    • Are we going to be able to trust their decisions?
    • What do we do when we think they’re wrong

Understanding meaning

  • For the most part, AI systems don’t currently understand the meaning of what they are doing
    • Able to translate Chinese text or speech into English very successfully
    • But have no idea what the material is actually about
  • But much work is going on in this area

Lack of AI system transparency

  • A big problem with any AI system that is based on statistics (i.e., neural net systems) is that it is opaque
    • You can’t query it to determine why it reasoned the way it did
    • The results just have to be accepted on faith
  • Hard to know whether or not it is operating within its expertise
  • No easy way to adjust a statistics-based AI system that is producing anomalous outputs

Different stakeholders’ needs for AI system transparency

  • Users: want to know what the system is doing and why, particularly if it is doing unexpected things
  • System certification authorities: want to be able to confirm the system meets safety standards
  • Accident investigators: want to trace the causes of an accident involving AI systems
  • Lawyers and expert witnesses: want to be able to explain their evidence to determine liability
  • The wider society: want to have confidence in the use of the technology

Machines’ theory of mind

  • Intelligent machines will soon need to understand other machines’ psychology to be able to work alongside them
  • They will need to understand human psychology too
  • We might want to worry when they start understanding human psychology all too well
    • Psychopaths are sometimes credited with having too great an understanding of others’ psychology
    • Fear machines using their understanding of humans to be more effective at manipulation

Controlling the Deployment of AI

AI “off the reservation”

  • AI technology is going to be widely available throughout the global society
    • We won’t be able to limit its spread through classification, or high cost, or regulatory controls
  • Some “bad dudes” are going to have AI capabilities
    • Political factions
    • International criminal cartels
    • Terrorists
  • And any smart teenager with a laptop and an Internet connection is going to be able to create some powerful AI capabilities

Avoiding premature release into the world

  • How do we make sure that a particular AI system is not harmful before it is released into the general world?
    • Operating on its own
    • Influencing other systems
  • Right now, there are basically no controls on this

Recursion: AI used to build even better AI

  • AI is very suited to developing faster, more powerful, more effective AI systems
  • AI systems will inevitably be used to code their own improvements
  • Such self-improvement is likely to proceed ever-faster in a positive feedback manner
    • People worry about an “intelligence explosion”
  • The degree to which humans can and will steer this process is an important issue

A world of multispecies intelligence

  • It looks like we are headed towards a world with a whole menagerie of different kinds of intelligences
  • Not just standard-issue humans and the more intelligent animals, but a great variety of different machine intelligence forms
  • How do we arrange for good governance of the whole society in such an environment?
  • How will the interests of the different intelligences be balanced?
    • Note that we’re only now beginning to consider the interests of intelligent animals

Changing Interactions with Technology

Human-machine intelligence collaboration

  • Human activities will increasingly take advantage of machine intelligence in a collaboration
  • Have the AI do the things it does best, and humans do the things they do best
  • This only extends a trend that has been going on for many years
    • Today, everyone lets the spreadsheet do the math for a large problem instead of laboriously computing the figures by hand

Changing modes of interaction

  • Up to now, the dominant mode for interacting with computing systems has been the keyboard, cursor control, and screen
    • Display of text and static images
  • AI systems are getting better at understanding and producing speech with all its nuances
    • We will increasingly converse with machines in the same way we would with another human
  • Systems will incorporate much more interactive video with synthesized avatars
  • The computing system will feel much more like a person

Systems will be listening to and watching you

  • Your Alexa device is listening 24 hours a day when it is turned on
  • Your screens will be watching you at the same time you are watching them
  • Information they collect will be shared with all sorts of other entities, with or without your knowledge or permission

Personal systems will be more directive

  • Smart personal systems are going to be ever-more-proactive in making recommendations to you (“for your benefit…”)
    • The Fitbit model, moving into all aspects of your life
  • How to keep them from being totally annoying?
    • Maybe you really don’t feel like going to the gym today, no matter what the machine says

Self-Driving Cars and Trucks?

Integrating driven and driverless cars

  • As of 2018 over 250 million driven vehicles in the U.S.
    • Driven vehicles aren’t going away anytime soon
  • How to mix driverless and driven vehicles safely?
    • Under a wide range of road environments
  • Need for huge public investments to support driverless vehicles
    • E.g., in highway network infrastructures
  • Today, the driverless vehicle is bursting forth without a solid legal, ethical and priorities framework

Car automation levels

  • Level 0: No automation
  • Level 1: Driver assistance: controls either speed or steering autonomously to assist the driver
  • Level 2: Partial automation: controls both speed and steering autonomously to assist the driver
  • Level 3: Conditional automation: controls speed and steering autonomously but requires driver monitoring
  • Level 4: High automation: able to complete a trip with no driver involvement under normal environmental conditions, not requiring driver monitoring
  • Level 5: Full autonomy: able to complete a trip with no driver involvement in any environmental conditions

Challenges of automating vehicles

  • Adverse environmental conditions (e.g., fog, heavy rain, snow; ice; potholes; rocks on the road; etc.)
  • Reduced system functionality (e.g., snow covering a camera lens, broken radar, software bug)
  • Unexpected road conditions (e.g., road maintenance, game day, accident, mattress flipping off the car ahead)
  • Uncertain behavior of other entities on the road (pedestrians, bicycles, animals; drunk drivers, police directing traffic, etc.)
  • Mixing driven and driverless vehicle behaviors
    • How drivers will interact with driverless cars, e.g., at four-way stops or left turns across heavy traffic
  • Extreme software size and complexity
    • Difficulty validating it and keeping it updated
  • Keeping every subsystem working to specification
  • Preventing system freeze-ups (the “blue screen of death”)
  • Responding to surprise conditions (e.g., siren and flashing lights of the ambulance coming up behind)
  • Phasing in driverless cars with partial infrastructure (incomplete wireless networks in rural areas, etc.)
  • In general, trust in automated vehicles
    • Am I really comfortable with that driverless semi in the oncoming lane, particularly in this heavy rain?

Control handover

  • How does the vehicle hand control back to the occupant when the conditions exceed the automation’s  capabilities?
    • The car unexpectedly signals to you “Take over! You have control now!”
    • Very hard for humans who have not been engaged to take over suddenly from automation
    • Aviation has found by painful experience this does not work well
    • Timing of the handover is critical
  • And how does the car know when to pass control over?

Trust in automated vehicles

  • Self-driving vehicles can be safer than human-driven vehicles under optimum conditions
    • Don’t suffer from distractions, anger, inebriation, etc.
  • But they have severe difficulty in handling adverse cases
  • People may be inclined to trust them when they really shouldn’t
  • Hard to know when to trust and when not to

The hacking risk

  • The more vehicles incorporate smart systems connected via wireless networks, the more vulnerable they are to malicious hacking
    • Not just stealing your car or things you left in it, but potentially causing your car to crash and kill you
  • Very little attention has been paid to vehicular cybersecurity up to now
    • No standards and regulations have been established

Public responses to self-driving vehicles

  • Already there are people resisting the introduction of self-driving vehicles into their area
    • People driving in ways to interfere with them
    • People vandalizing them (slashing tires, damaging sensors, etc.)
  • Self-driving cars are going to have significant additional expenses over driver-controlled versions
    • Both initial costs and maintenance/repair costs
    • All those sensors are expensive and have to be carefully calibrated

Self-driving cars and ethics

  • How do we define values and ethical principles for self-driving cars?
  • How conservative do you want your car’s driving to be to maximize safety?
    • How safe is too safe?
    • What if some owners of self-driving cars set a low value on your safety?
  • When should your car determine it has to kill you, in order to save the lives of others?

What moral decisions should self-driving cars make?

Ubiquity?

Smart everything

  • As AI capabilities become ever-cheaper, we can expect some form of AI to become incorporated into all manner of products
    • Even relatively inexpensive items, such as toys
    • If it has a battery or plugs into a wall, it may well have a processor running AI programs
  • This will be combined with the Internet of Things, where devices will be wirelessly linked into central networks of various kinds

Some consequences of smart everything

  • Smart consumer items that are connected through the general Internet will expose you to all sorts of things you might not appreciate
    • Monitoring and reporting on your activities
      • Your television watches you as you watch it
      • Alexa listens to every sound in your home
    • Hacking: all kinds of possible mischief
      • E.g., unlock your doors and turn off the security system so your house can be burglarized
  • Low-cost consumer systems are terrible in terms of implementing effective cybersecurity

AI and Personal Information

Ain’t no such thing as a free lunch

  • Making use of these nifty new capabilities is definitely a mixed blessing
  • There are lots of hidden costs and potential downsides
  • Just like today when you use Google or Facebook
    • You don’t pay directly for any of these helpful functions
    • In exchange, you are implicitly allowing the company to collect all kinds of information about you
    • They make their money by selling your personal data to third parties who aren’t necessarily your friends

The system knows everything about you

  • Increasingly, all sorts of personal information is being collected, compiled, and analyzed by entities of all sorts
    • Marketing companies such as Amazon that want to sell you things
    • Platform entities such as Google and Facebook whose business model involves selling your profile
    • Governmental agencies of all kinds that want to monitor what you do
    • Many others, including entities that want to detect vulnerabilities, in order to defraud or steal from you

Your personal data is no longer your own

  • Potentially everything you do while interacting with some kind of an information system may be monitored, recorded, transmitted to an archive, combined with other data, and analyzed by AI systems
  • Who has access to the resulting knowledge about you?
  • What conclusions are they able to draw?
  • What are they able to do to you as a result?
  • What protections do you have against being harmed?

AI monitoring and analysis

  • AI-based software: monitoring and analyzing all you do
    • Every web page you visit and what you do on it
    • Every cable television program you watch
    • Every email, text message, Facebook post, etc.
    • Every phone call
    • Everything you purchase with a credit card
    • Every financial transaction, the status of each account
    • Your location at any time
    • How you drive
    • Every interaction you have with the medical system
    • Your daily schedule
    • Who you interact with

What happens with your information

  • Entities share your information with each other in ways that benefit their interests
  • You have almost no control of what happens with your information
  • It is almost impossible to find and remove errors
    • Bad information keeps getting re-inserted from old files
  • AI systems are very powerful at finding patterns of all sorts in the data about you

Knowing too much about each other

  • Consider a Google glass system with AI face recognition and wireless access to central databases
  • The policeman wearing it sees which passerby has an outstanding warrant or a history of violence
    • Will only the authorities have access to such capabilities?
    • If I have an equivalent system, can I see the cop’s history of police brutality complaints?
  • Eventually, can anyone see any data on anyone else?
  • How will access to such information be controlled?   Who sets the policies on this?

Social control: the Chinese social credit system

  • China is currently running one of the biggest social control experiments of all time, using extensive AI
  • They plan to rank all Chinese citizens based on their “social credit” score by 2020
  • The social credit score is similar to the Western financial credit score, but much more comprehensive
  • The scores move up or down according to peoples’ behavior, over their whole lives
  • People are rewarded or punished according to their scores
  • The scoring algorithms are state secrets
  • Correlation of commercial and government data bases
  • Examples of things that drop your score: bad driving, smoking in prohibited areas, paying a bill late, criticizing the government; an inappropriate social media post
  • Your score is correlated with those of the people you associate with: family, friends, coworkers, etc.
    • Your score can be reduced by something they do
  • People with low scores face reduction in rights and opportunities: education, jobs, housing, travel, etc.
  • The system is intended to strongly enforce conformity to government-defined norms
  • https://www.youtube.com/watch?v=y5-0llHaZDs

Politics around AI and robotics

  • Automation and globalization are causing increased support for nationalist and radical right political parties
  • Target for resentment: Highly educated coastal “elites” who make very good livings developing AIs and robots that put “ordinary people” out of work
  • Technological disruption is not even on the radar screen for most political groups and politicians
  • Few elected officials understand technology policy issues
  • Major geopolitical implications of a global AI technology race between national powers

Near-Term Implications of AI in Different Fields

Types of jobs that will stay, types that won’t

AI in medicine

  • All medical instrumentation connected to a patient will feed into an integrated pattern recognition capability
    • The Internet of Medical Things (IoMT)
    • A holistic real-time picture of the patient’s status, correlated with the patient’s history
  • Almost any medical condition with an acute episode—e.g., asthma attack, seizure, stroke, heart attack, autoimmune attack—will be potentially predictable
  • Diagnostic systems are already being transformed by AI
  • Similarly, treatment recommendation systems

AI in mental health

  • Today, mental health care is tremendously constrained by the shortage of trained providers relative to the needs
    • Also major problem of care affordability
  • AI systems are becoming increasingly capable of interacting with patients and detecting markers of mental health issues at an early stage
    • Depression, anxiety, PTSD, substance abuse, cognitive or memory decline, etc.
  • AI systems may become a part of mental health counseling and support as well
    • A virtual therapist available 24/7/365, at low cost

AI in human reproduction

  • AI is highly synergistic with genetic technology
  • Breakthrough techniques such as CRISPR/cas9 and gene drive now allow ready modifications of genomes
    • Features are heritable, passed on to descendants
  • The temptation for parents to create designer babies with desirable characteristics will be very high
    • Health, long life, intelligence, athletic ability, beauty, musical talent, creativity, etc.
  • Cost: will likely make such technology available first to those who are already economically advantaged

AI in law

  • Most of the law is based on analysis of highly-structured data—e.g., contracts, suits, etc.
  • AI systems are ideally suited to perform these functions
  • The digital law library is the raw material for AI law
    • Ability to search cases for precedent judgments
    • Increase the span of the search
  • Similarly, AI systems can assist human judges in rendering verdicts based on analysis of evidence, case law, other factors
    • Question: will people accept judgments rendered by an AI system?

AI in finance

  • Already, high-volume trading is almost completely conducted by AI-enabled computer algorithms
    • Stocks, bonds, commodities, currencies, etc.
    • Blockchain transactions (e.g., Bitcoin)
  • Investment analyses are performed by AI systems
  • Financial decisions of all sorts are increasingly made by machine intelligences advising the humans

AI in insurance

  • AI systems will enable fine-grained assessment of individual policyholders’ risks based on their profiles
    • Underwriters are going to know a great deal more about policyholders than ever before
  • In place of statistical-based insurance pools, insurance will become much more individualized
  • Might result in increased denial of insurance or very
    high premiums for policyholders deemed to be high risk
  • A powerful influence on behavior

AI and audits

  • Currently, audits (e.g., of taxes) using manual processes are time-consuming and expensive
    • As a result, only a small fraction are actually audited
    • Lots of discrepancies slip by as a result
  • Automated systems using AI are likely to greatly facilitate performing audits
  • What will be the effect of comprehensive audits becoming routine?

AI in education

  • AI can support highly individualized education and training, at the student’s own pace
    • Using affective computing to interact like a tutor
  • An AI system can study the individual student and develop a detailed model of him/her
  • Learning styles, preferences, cognitive biases, etc.
    • What the student does and doesn’t understand so far
  • It can tailor the presentation of the material to best match the student at each stage of the course
    • Repeat material as necessary, in different forms, until the student really masters it

AI in the media

  • Chinese state television is already running a TV news anchor that is not a human, but a simulacrum with AI-synthesized imagery and voice
  • Sports reports are beginning to be generated by AI systems
  • Similarly things like weather reports, business reports, etc.
  • More and more text reporting in newspapers and online media is going to be written by AI systems

AI in marketing

  • AI analyses will be conducted on all manner of data collected about you
    • Your prior purchases, your product searches on websites, your attention to ads
    • All aspects of your financial situation
    • Demographic data on you, your family, your friends, your coworkers, your neighbors, etc.
    • Your psychological profile
  • Precision-targeted marketing will be generated to maximize its appeal specifically to you

AI in manufacturing

  • Obviously, smart robotic systems are going to perform ever-larger fractions of production work
  • AI systems are going to be coupled with advanced 3D printing/additive manufacturing systems
    • Quickly produce custom or low-volume parts
      • E.g., clothes custom-tailored to fit your measurements  exactly
    • Do just-in-time manufacturing
  • Manufacturing with AI systems may shift from mass production to mass customization

AI in engineering, architecture, etc.

  • The engineer’s knowledge base is increasingly being codified to create intelligent engineering assistants
  • The engineer’s role will be raised to a higher level, focusing on defining what the design should do
    • The automated system will then detail out a candidate design for the engineer to review
    • Many more options and alternatives can be explored to get to an optimum

AI in software development

  • Examples of applications of AI in software development include
    • Requirements analysis and specification support
    • Intelligent programming assistants
    • Software performance assessment
    • Software interface management
    • Software test administration
    • Software verification and validation
    • Software defect and vulnerability detection

AI in government

  • Some examples of AI applications in government include
    • Regulatory compliance monitoring
    • Tax assessment and collection
    • Application review of all types
    • Public assistance and entitlement program administration (unemployment, Social Security, Medicare, Medicaid, veterans programs, etc.)
    • Immigration administration
    • Election administration

AI in the military

  • Self-guided weapons have been around for a long time
    • E.g., homing torpedoes, heat-seeking missiles
    • Up until recently, these have been targeted and launched by human decision makers
    • Such weapons have been incorporating more and more machine intelligence in order to defeat countermeasures and improve performance
    • Some of them are already autonomous
  • To what degree do we want military systems to make decisions on their own, with no human in the loop?
    • What policies govern autonomous weapons?
  • Logistics is a huge part of military operations
    • Procurement, supplies management, transportation, equipment maintenance, etc.
  • AI-enabled systems can vastly improve logistics management over human administration
  • Strategy and tactics are also opportunities for AI
    • Similar to game-playing systems, an AI program can play Blue vs. Red for thousands of iterations to select the best options for success in a conflict
  • Robotic systems are likely to largely replace manned systems on the battlefield
  • For example, why should combat aircraft have humans onboard?
  • Why should conveys of supply trucks have human drivers?

AI in law enforcement

  • Today, concepts are being explored for AI-enabled predictive policing
    • Based on indicators of all kinds, when, where, why, and how are crimes most likely to occur?
      • Criminal acts
      • Perpetrators
      • Victims
    • Applies to all types of crimes, from fraud to violence
  • How is law enforcement likely to transform with an emphasis on pre-crime actions?
  • What if the authorities have some kind of compromising information on everyone?  Enforcing will then become selective
  • How tolerant does society want to be for small infractions?
    • Your AI-enabled car could easily report every time you exceed the speed limit or fail to stop completely at a stop sign for a right turn
  • How comfortable will we be with robot policemen, especially if they are autonomous and authorized to use lethal force?

AI and cyberwarfare

  • Hostile cyber acts are an escalating problem
  • Cyberwarfare is a potential weapon of mass destruction
    • E.g., take down a nation’s power grids
    • Destroy a nation’s financial systems
    • Cause nuclear reactors to self-destruct
  • Most current military systems, particularly AI-enabled ones, are extremely vulnerable to cyberattacks
  • Essential to detect and respond to cyberwarfare attacks on all critical systems
  • There will be a continual arms race between attacker and defender

AI and cyberterrorism

  • Cyberterrorism is a parallel threat to cyberwarfare
  • Often, it is not clear who is the source of a cyberattack
    • Attackers can be anything from a hostile nation, to some rebel faction, to a few individuals
    • Very difficult to deter cyberterrorism
    • Very difficult to find and punish the attackers
  • The number of potential targets and possible attack vectors is huge
  • Like cyberwarfare, the consequences could be sever

AI and cybercrime

  • There is likely to be a growing arms race between those employing AI to commit crimes and those employing AI to detect and thwart such crimes
    • Particularly major financial crimes
  • AI systems can be extremely capable at searching for vulnerabilities to exploit, while avoiding detection
  • Cybercriminals can be based anywhere in the world and attack anywhere
  • The potential payoffs for cybercrime can be huge
  • It is critical that defenders have the resources to over-match cybercrime attackers

AI in elections

  • Democracy depends on the legitimacy and trustworthiness of the election processes end-to-end
  • The integrity of voting is key
  • Current voting machines, vote compilation systems, and vote reporting systems are horribly insecure
    • ~185,000 polling places in the U.S.
    • Numerous vendors of systems
    • No general oversight of security
  • AI systems are needed to detect interventions throughout the election chain

AI and the Arts

AI and music

  • AI systems can already compose original music in the style of a particular human composer that can fool experts
    • Not great masterworks yet , but steadily improving
  • What will be the future of human-created music?
    • Will it become a niche, like a hobby?
    • Will people still value it in the same way?
  • See https://www.youtube.com/watch?v=wYb3Wimn01s

AI and new musical forms

  • A neural network can learn the musical characteristics of an instrument by analyzing hundreds of notes
  • It creates a mathematical representation, or vector, that identifies a particular instrument
  • Now these vectors can be combined to create entirely new instruments
  • One new synthetic instrument might be 47 percent bassoon and 53 percent clavichord. Another might switch the percentages

AI and photography

  • A new revolution: computational photography
    • Enabled by major advances in processing hardware and Deep Learning AI systems
  • Ordinary smartphones are going to be capable of astounding photographic capabilities
    • Lenses and sensors no longer the limiting factors
    • The camera is going to understand what it is looking at and adjust the image to maximize quality
  • Photo libraries will be searchable on the basis of content
  • With learning, the capabilities will continue to get better and better over time

AI and art

  • Christie’s recently sold its first AI-created portrait painting for $432,500
  • AI systems using GANs can already create convincing original art (paintings, sculptures, etc.) In the style of a particular human artist
    • Again, not great masterworks yet
    • The systems will only get better with time
  • How should we appreciate and value art created by machine intelligences?

AI and computer games

  • AI is already a key part in advanced computer games, particularly multiplayer games
    • Both your partners and your adversaries may be AI entities
  • Your character’s abilities may be enhanced by AI
  • Will games continue to be as much fun as the machine takes a greater and greater role?

AI and toys

  • Higher-end toys are going to be highly interactive using AI
    • E.g., interfacing through a smartphone or smart speaker, networking through the cloud
  • Audio chat, video, etc.
  • Extensive use of affective computing
  • Toys will develop compelling individual personalities in the process of interacting with the user

AI and the movies

  • Computer-generated imagery in the movies already employs extensive AI technology
    • Steadily getting better, faster, and cheaper to use
  • We are beginning to see movies where actors no longer living are recreated digitally to play new roles
    • Will human actors continue to have the same value?

AI and virtual/augmented reality

  • The ability to generate high-resolution imagery, sound, and other sensory effects in real time using AI capabilities is advancing quickly
    • Making possible ever-more convincing virtual reality and augmented reality systems
  • How important will virtual/augmented reality be in future entertainment?

Fake people

Deepfake video

  • Photoshop capabilities today are for still images
  • Emerging: AI-based equivalents for digital video
  • Analyze posture, body language, movement, facial expressions, vocal tone, linguistic characteristics, other features from recordings
  • Then synthesize a convincing digital video segment
    • E.g., a politician saying things that she or he did not say,
    • Or a celebrity shown in a porn video
  • You won’t be able to tell the difference from reality
  • Consequence: Never again trust a video you see on television or online

Deepfake example

AI and Robot-Human Interactions

Robot co-workers

  • Humans and intelligent robots will increasingly function as co-workers
    • E.g., doing construction work, where the robot has capabilities (reach, strength, etc.) beyond that of a human
  • The robot will interact with its human partners through speech and will perform some functions autonomously
  • Work teams will be mixtures of humans and robotic systems

Human assistance robots

  • One very active area of development is robots to provide assistance to humans with physical and cognitive limitations of various types
    • Mobility assistance of all kinds, helping people transfer (e.g., getting up from a bed or chair)
    • Helping with activities of daily living (preparing and serving food, feeding, bathing, dressing, etc.)
    • Household tasks (cleaning, laundry, etc.)
    • Supplement the functions of service animals
  • Some of the physical tasks are actually quite challenging for a robot
  • Probably won’t be a single robot able to do everything

Companion systems

  • AI systems will come to serve as companions
    • Already, people are developing feelings of relationship with their intelligent virtual assistant systems (Siri, Alexa, Google Home, etc.)
    • The system will come to know you intimately
    • You will disclose your innermost feelings to it
    • It may become a more entertaining conversationalist than even the cleverest of your human friends
  • Loneliness/isolation afflicts a large fraction of people, and companion systems can help fill the gap

Machine intelligence and sex

  • Sex has been a driver for a lot of digital technology
    • Online pornography influenced the development of high-bandwidth Internet connections for streaming video
    • Today people are developing high-fidelity human-simulating robots for sexual interactions
    • Initially crude, these are becoming more and more lifelike—e.g., able to hold real conversations, as well as being available for sex whenever desired
    • See https://www.youtube.com/watch?v=LcDWigVV6tA
  • Society has yet to decide how to feel about sex with robots—a boon, or an abomination?

The creepiness issue

  • People tend to find mimics of humanness that are just slightly off to be creepy
  • There is not a problem if the simulation is obviously an artifact
  • Humanoid objects which appear almost, but not exactly, like real human beings elicit feelings of eeriness and revulsion in observers
  • Examples are found in robotics, 3D computer animations, and lifelike dolls

What Are the Worries About AI?

Worries are a function of time

  • There are near-term worries, mid-term worries, and long-term worries
  • Near-term worries are mostly about rapid disruptions of the existing order
  • Mid-term worries are about fundamental changes in how humans live and act in the world
  • Long-term worries are existential—are we creating conditions that threaten basic human existence?

Some near-term worries

  • The disappearance of anonymity and privacy
  • Machine intelligence in the service of human stupidity
  • Transferring authority and responsibility to machines
  • AI’s effects on society’s concentrations of power
    • Benefits will go to the wealthy and well-connected
  • Exploitation of human weaknesses
  • AI enablement of techno-authoritarianism
  • Massive changes to everyone’s jobs
    • Constant raising of the bar to be employable
  • Economic disruption from large-scale unemployment

What worries AI insiders?

  • Handing too much responsibility over to AI systems and becoming over-dependent on them
  • Inappropriate trust in AI systems
  • Over-estimating the competence of AI systems
  • Security of AI systems, including under attack by AIs
  • Keeping human cognitive skills from atrophying
  • Powerful AI in the “wrong hands”—hobbyists, hackers, rogue regimes, criminals
  • Increase in social inequality
  • Giving up too much of our humanity to machines

Example of risks: financial trading

  • August 12, 2012: Wall Street’s largest trading company, Knight Capital, switched on a new AI-enabled program for buying and selling shares
  • Due to an undetected bug, the system immediately began flooding the exchanges with irrational orders
  • It took 45 minutes for Knight’s programmers to diagnose and fix the problem
  • During that time, the software made over 4 million deals, with $7 billion in errant trades
    • This nearly bankrupted the company

Keeping up human thinking skills

  • As we hand over more and more of our thinking skills to machine systems, how do we prevent our own skills from atrophying from lack of use?
  • Do we need to require periodic refraining from using automated systems?
    • Like airline pilots who are required to regularly practice landing by hand to keep their skills sharp
  • How can we encourage everyone to maintain their human thinking skills when the automation is so good and so convenient?

Ceding too much to machines

    • One concern is that we will cede excessive amounts of responsibility to AI systems
    • The global financial system is a good example of a risk area
  • Potential for a catastrophic system failure caused by multiple minor flaws in over-empowered connected machine systems

Transfer between domains

  • We aren’t good at assessing how a highly-optimized rule or structure in AI will transfer to a new domain
  • How do we know when a machine intelligence has left its comfort zone and is operating on parts of the problem it is not good at?
  • Don’t put a machine intelligence in charge of decisions it doesn’t have the intelligence or knowledge to make

AI and techno-authoritarianism

  • AI can be a powerful enabler for authoritarian societal control
  • Surveillance of the population can be near-total using AI tools
    • Any deviation from what is authorized is detected and responded to
    • Resistance and rebellion can’t even get started
  • The Chinese social credit score system goes a long way in the direction of techno-authoritarianism
  • Could be much worse—think North Korea
    • Say the wrong thing or make the wrong face in front of a screen and you disappear, never to be seen again

Pressures for conformity

    • Even without more extreme forms of techno-authoritarianism, AI systems could lead to excessive pressures to conform to social norms
    • Everything you do is observed and evaluated and potentially shared with others, outside your control
      • Very hard to have any secrets or private quirks
  • Likelihood of stifling creativity, innovation, risk-taking, etc. 

Globalization,  Automation, AI and Jobs

Globalization effects

  • In addition to automation and AI, globalization is having a huge impact on jobs
  • Work is moved to the place on the Earth where costs are least
    • Lowest labor costs
    • Lowest capital costs/best subsidies
    • Lowest taxation
    • Least burdensome government regulation
    • Least concern for externalities (e.g., environmental damage, social harms)
  • Automation and AI effects are compounded by globalization
  • The developed economies (e.g., the U.S., Europe, Japan, etc.) are being surpassed
  • China is about to become the largest world economy
  • India will shortly become the second largest
  • Other Asian economies are among the fastest growing
  • These economies have fewer pre-existing institutional and cultural barriers to automation and AI

Jobs: AI and automated systems vs. humans

  • An AI system or a robot doesn’t:
  • Demand a salary or wages
  • Need health care insurance, Social Security and Medicare contributions, or workman’s comp
  • Take vacations, holidays, maternity or sick leave
  • Join a union or strike for better working conditions
  • Need a human relations department
  • Can work 24 hours a day, 7 days a week, 52 weeks a year without tiring or taking a break
  • Unquestioningly obedient, highly motivated, and cooperative

Human aspects that AI systems don’t have

  • Human mind overhead
    • Distractions
    • Worries
    • Emotional commitments
    • Charged memories
    • Allegiances
  • Susceptibility to a wide range of cognitive biases

Intelligent machines and jobs

  • Unlike a human employee, the employer gets a tax break for machine depreciation
    • Latest U.S. tax law: employer can expense 100% of a robot’s cost the first year, rather than over the life of the machine
  • Result: overwhelming economic pressures to replace human workers with AI and robots wherever possible

Whose jobs are at high risk?

Anyone

  • Doing well-characterized repetitive physical actions (manufacturing; construction; mining; farm work, etc.)
  • Doing financial services (e.g., bank tellers, financial analysts and advisors, loan officers, accountants, bookkeepers, brokers, insurance underwriters, claims representatives, etc.)
  • Doing highly structured analysis (e.g., paralegals and legal assistants, statistical analysts, report writers, etc.)
  • Doing routine middle management tasks unlike a human employee, the employer gets a tax break for machine depreciation

Some other jobs at high risk

  • Office clerks of all types
  • Inventory workers
  • Retail salespersons
  • Drivers (truck, taxi, etc.), couriers
  • Food service personnel
  • Telemarketers
  • Journalists
  • Security guards
  • Medical diagnosticians
  • Pharmacists
  • Postal service workers
  • Meter readers
  • Computer operators

Whose jobs are going to remain?

Jobs that will better resist replacement by machine intelligences involve:

  • Extensive face-to-face / hands-on human interaction
  • Unpredictable physical work, requiring on-the-spot creative adaptation
  • Extensive, broadly-based education and experience
  • Critical thinking, high creativity, and the ability to innovate

Jobs we’ll lose, jobs we’ll keep

See https://www.ted.com/talks/anthony_goldbloom_the_jobs_we_ll_lose_to_machines_and_the_ones_we_won_t

What jobs will change dramatically?

  • Education
  • Medicine
  • Therapy
  • Software development
  • Engineering
  • The military
  • Skilled trades
  • Performing arts

Adapting to a new work environment

  • Needed: creativity, critical thinking, emotional intelligence, adaptability, and collaboration skills
  • Learning how to learn—and constantly reinvent oneself and develop new abilities over a whole lifetime
  • The problem is that not everyone is cut out for this degree of independent learning and self-reinvention
    • Takes a lot of drive, self-direction, self-discipline
  • Likely to make current inequalities between people even greater in the future

Automation/AI and the Economy

What will jobs look like in an AI-based economy?

  • Some types of jobs that will remain will be those:
    • Requiring scarce talents and high levels of education
    • Requiring extensive face-to-face human interaction and emotional intelligence
    • Requiring good trade school training and excellent problem-solving skills
  • Few jobs for those who lack scarce talents and have modest levels of education

 Economic dynamics of a highly-automated world

  • Customers are required in order for the goods and services that are produced to be purchased
    • Customers have to have money in order to buy
  • In the current economic system, most consumer buying power comes from current employment, or savings derived from previous employment
    • Automation will displace a large fraction of jobs
    • No jobs = no buying power
  • With no buying power, no demand
  • Result: economic contraction, in a positive feedback loop

Changes to the basic economic structure

  • Who benefits from AI? the owners of capital, who will control most of the intelligent machines
  • Who suffers? the rest of us, who currently trade work for money. No work means no money
  • What is the responsibility of the rich and powerful to the rest of us?
  • When intelligent machines become more profitable to them than human workers
  • The response to the mass unemployment of the AI Revolution has to involve some kind of sweeping redistribution of income that decouples it from work

Redesigning basic aspects of society

  • In a highly automated world, we will have no choice but to fundamentally redesign some basic aspects of the society
    • How we define societal values and goals
    • How we govern ourselves
    • How we distribute power
    • How we distribute/redistribute wealth
    • How we protect ourselves against parasites, predators, free riders, etc.

Automation-induced obsolescence and cultures

  • What happens when any work we might do can be done better by machines?
  • Automation-induced obsolescence has already had devastating effects on some of the world’s people
  • Some cultures (e.g., those based on planting, cultivating, and harvesting corn and beans) have collapsed and lost their meaning to the people who were shaped by them
  • As automation replaces more and more human work, how do people continue to maintain their sense of worth?

What will society do with people no longer “useful”?

  • People find meaning in work, in doing something they and others consider valuable and appreciated
  • When people feel they have nothing to contribute, it is very harmful to them psychologically
  • If your skills are taken over by machine systems, what are you to do in order to feel you’re of worth?
    • And for the society to feel you’re of value?

Higher Level Worries

Some mid-term worries

  • Developing AIs with autonomy and agency/self-interest without associated protections for humans
    • Letting AIs act on their own interests, not humans’
  • Failing to incorporate appropriate human values, ethics, and moral reasoning frameworks in AIs
  • Creating machine intelligence persons with protections and rights prior to developing appropriate societal structures to accommodate them
  • Abdicating human responsibilities for actions taken by machines

Motivations of AIs

  • Clearly we want AIs to be motivated to loyally serve the interests of their creators and owners
  • We also want AIs to be highly motivated to serve the interests of humans in general
  • We would like them to have a well-developed sense of responsibility for their actions
  • Do we want to allow AIs to develop motivations that are not tied to those of humans?
    • What could be the possible consequences?

Autonomy and self-interest

  • Humans and animals have autonomy and self-interest
  • Instincts for self-protection / self-preservation
  • Dare we give machine systems similar instincts?
    • “Don’t touch my power switch, human!”
  • What about a desire to increase their access to resources?
  • What about a desire to reproduce themselves?

Dangerous machine intelligences

  • Don’t create machines instructed to “survive, reproduce, access resources, and improve in the best way possible”
  • Such systems are not likely to remain friends of humanity for very long
  • Self-reproduction combined with autonomy: the really dangerous step in the development of machine intelligences

Things we can’t afford to get wrong

  • Self-interested AI systems interacting with each other
  • AI systems developing and evolving on their own, outside human control
  • Technological synergies conducted in an unmanaged way—e.g.,
    • AI and genetic engineering
    • AI augmentation of humans

Research it is critical to do

  • How to manage the development of safe and beneficial AI is a very important and challenging research problem
    • Highly interdisciplinary
    • Involves many different kinds of institutions
    • Requires collaboration among researchers in many different fields
    • Needs to be fully international
  • Have to start working on these aspects now
  • We can’t afford to learn from making mistakes—have to anticipate them and steer clear from the start
    • May only have one shot at getting it right!

Machine Intelligences and Personhood

What is a “person”?

  • We distinguish between entities that we consider persons and entities that we consider are not persons
  • “Persons” are given rights and protections on the basis of their “personhood”
    • These are withheld from “nonpersons”
  • Increasingly, we are grappling with the question of how to define personhood
  • Is personhood a single state?  Or are there degrees of personhood?
  • How is personhood a function of consciousness?
  • The definition of “person” and the associated social rights and protections have varied greatly over time and over different cultures and belief systems
  • In most eras, women did not have the same rights and protections as did men
  • Similarly, slaves did not have the same rights and protections as free persons
    • Prisoners still do not
  • Previously, rights and protections depended on race
  • Children and youths have rights and protections that depend on age
  • Recently we have been expanding protections (e.g., of disabled persons)
  • In the U.S., courts have determined that corporations are “persons” in the eyes of the law.  What in the heck does that mean?

Who/what/when is a “person”?

  • Is a human ovum just penetrated one second ago by a sperm a person?
  • Is an 8-cell human zygote in a fertility clinic freezer?
    • Just when does a fetus become a person?
  • Is a human infant born with anencephaly?
  • Is a brain-damaged human adult kept on life support in a persistent vegetative state?
  • Is an elderly human adult with extreme dementia?
  • Is a human conjoined twin with a single body and two heads one person…or two?

What is a person?  New cases

  • Would an augmented or modified human produced by genetic engineering be a person?
    • With what degree of augmentation or modification?
  • Would a recreated Neanderthal be a person?
  • Is a highly intelligent nonhuman (e.g., a bottlenose dolphin, a chimpanzee, or an elephant) a person?
  • If we encountered one, would an extra-terrestrial with human-equivalent intelligence be a person?Is a human ovum just penetrated one second ago by a sperm a person?
  • Will a cyborg that has a combination of human and machine intelligence be a person?
  • Will a humanoid robot with a degree of self-awareness and agency be a person?
  • Will an AI program on a distributed network with self-awareness and agency be a person?

Degrees of personhood

  • If we decide to have degrees of personhood, how do we all agree on the scales and the associated rules?
    • A lot of personhood has ultimately to do with consciousness
    • Consciousness is a complex subject in its own right
  • In addition to rights and protections, what societal obligations and responsibilities are associated with degrees of personhood?

Personhood and ownership

  • Contemporary principle:
    • Non-person entities can be owned by persons
    • Persons cannot be owned by other persons
      • Corporate persons modify this principle by being able to own other corporate persons
      • Persons can own corporate persons
  • What will ownership come to mean, in the emerging world of intelligent machine entities?

Consciousness and Machine Intelligence

Consciousness: not a simple matter

  • The meaning of consciousness is the subject of intense study and debate
  • What exactly does it mean for an entity to be conscious?
  • Most human thinking actually happens at a level of subconscious processes
    • This is why you can arrive at your home without being aware of the drive to get there
  • When we speak of “consciousness”, there are actually multiple aspects to consider
  • See https://www.youtube.com/watch?v=CTHh-5kcqC0

Sentience

  • Sentience is defined as the capacity to feel, perceive, or experience subjectively
  • Animal rights activists note that animals have sentience, particularly in terms of their capacity to suffer
    • They argue they should then be provided with certain rights and protections
  • To what degree will we endow machine intelligences with sentience and feelings?
    • And what will be the consequences?

Self-awareness

  • Self-awareness has to do with being conscious of and observing the flow of one’s thoughts
    • Being aware of being aware
    • Thinking about one’s thinking
  • Self-awareness is kind of a meta-level of consciousness
  • A significant portion of non-human animals can be shown to have at least a degree of self-awareness
  • Although AI systems don’t need to have self-awareness to function, there is no barrier to having it

Sapience

  • Sapience is defined as the ability to think and act using knowledge, experience, understanding, insight, and common sense
  • Sapience is often equated with wisdom
  • We even term our species homo sapiens: “man the wise”
  • To what degree will we endow machine intelligence with the capacity for sapience?

Consequences of Machine Personhood

Rights and protections of machine intelligences?

  • At what point do we give rights and protections to machine intelligences?
  • And just what rights, protections, and freedoms do we give them?
    • A current example is the debate as to whether bots have protected freedom of speech
  • How are these rights and protections distinguished from those of other types of persons?
  • See https://www.youtube.com/watch?v=DHyUYg8X31c

Obligations and responsibilities of machine intelligences?

  • In exchange for having rights, protections, and freedoms, what obligations and responsibilities will machine intelligences have?
    • Obey the same laws and rules as humans
    • Pay their assigned taxes and fees
    • What else?  Serve on juries??  Perform national service??

When a machine intelligence goes bad

  • When a machine intelligence fails in the obligations and responsibilities we’ve given it, what do we do?
    • How do you deter a machine intelligence from doing something you consider wrong?
    • Is it meaningful to punish a machine intelligence after some violation?
    • If so, what would be appropriate punishment?

Will we be able to “pull the plug”?

  • People think, if an AI or robot ever becomes a problem, we could always “pull the plug” to shut it off
  • Unfortunately, if an AI system develops self-interest, it will likely act to prevent this
    • E.g., disable its own off switch and get access to backup power
    • Or transfer its memory and thinking states to another connected machine system and thus live on
  • We may not have the right to shut off an intelligent machine endowed with personhood

When an AI system causes harm, who will be held responsible?

  • Inevitably, AI systems will cause harm
    • E.g., self-driving cars will make errors, get into accidents, damage property, and hurt people
  • Who to hold responsible
    • Owner of the car?
    • Occupant, who didn’t take control?
    • Manufacturer of the car?
    • Manufacturer of the car’s control system?
    • Programmer of the AI software?
    • The city, whose infrastructure was defective?
  • We don’t have a legal structure for such things yet

When thinking machines break the law

  • Thinking machines will inevitably break laws
  • Previously, we held the person controlling a machine to be accountable
  • When the machine is more autonomous, accountability becomes complex
  • Or when the machine’s behavior is the result of many peoples’ inputs, not all of whom are acting in concert
  • How will responsibility be assigned?

How does the legal system have to change?

  • The current legal system is based on assigning and allocating responsibility to human actors
    • Both in the case of harm and benefit
  • When machine systems take actions with consequences, how will that responsibility be considered?
  • The machine intelligence may be extremely distributed, with many components involved in a particular action

AIs and moral decisions

  • We will be charging intelligent machines (e.g., self-driving cars) with making serious moral decisions
  • For example, how does the machine decide which harm to choose when there is no harm-free course of action?
    • We will want them to quantify and weigh different types of potential harms to different entities
    • How are they to weigh injuries of what severity and likelihood against a fatality?  Against how much property damage?
  • Unfortunately, science is ill-equipped for answering moral questions, yet someone will have to do it for the machines we are creating

Objective functions

  • Autonomous systems are commonly designed to maximize some objective function
  • The objective function defines the goals to be sought, as well as constraints and aspects to be avoided
  • Critical to define the objective function well for any autonomous machine intelligence!
  • In general, we aren’t very good at describing our intent
    • We tend not to consider possible unintended consequences in advance
    • We don’t anticipate possible corner cases
  • Be very careful what you ask for…

Specifying intent

  • We’re generally not very good at defining what we intend
  • Likely to be disconnects between what we ask AI systems to do and what we really want them to do
    • Especially we will fail to identify things we want them not to do
    • We can expect to be surprised by them because of our omissions

Values

  • We want machines’ decision-making to be well-aligned with human values
    • Their goal systems should be based on human values
    • Need to provide a framework for machines’ moral reasoning
  • Unfortunately, human values are not simple and consistent
    • Individuals’ values often conflict with each other
  • And whose values are to be used to instill into AIs?
    • Moral calculus differs from one culture to another and changes over time within a culture

Incentives for guiding machine behavior

  • Machines won’t have the incentives humans use for promoting good behavior
    • Sense of shame or praise, concern for reputation, deference to authority, doing something simply because it is the right thing to do
  • The social and legal systems that have dealt effectively with human rule-breakers of all sorts will fail in unexpected ways with machine intelligences

Machine saintliness?

  • How can we arrange for machine intelligences to have the highest and most positive motivations in their relationships with humans?
    • Benevolence, compassion, empathy, kindness, thoughtfulness, generosity, magnanimity, etc.
    • Humility, selflessness, patience, tolerance, self-sacrifice, etc.
  • This particularly will be a challenge when we endow machine intelligences with autonomy, self-interest, and self-reprogramming

Machine ethics

  • There is a developing field of machine ethics
  • What are the basic principles of ethics we want to build into machine intelligences?
  • Not a simple matter, as human ethics have been a matter of debate for millennia
    • People differ on ethics, sometimes strongly
    • Ethical principles have changed over time
  • Have to clarify ethics into principles that can be defined explicitly and programmed

An early approach

  • Isaac Asimov proposed The Three Laws of Robotics in 1942 (!)
    • 1.  A robot may not injure a human being or, through inaction, allow a human being to come to harm.
    • 2.  A robot must obey the orders given it by human beings except where such orders would conflict with the First Law.
    • 3.  A robot must protect its own existence as long as such protection does not conflict with the First or Second Laws.
  • Asimov later added a Fourth Law
    • A robot may not harm humanity, or, by inaction, allow humanity to come to harm
  • Unfortunately, implementing Asimov’s Laws turns out to be remarkably difficult

Corresponding principles for humans in using AI and robotics

  • Parallel need: Principles for human laws concerning the use of AI and robotic systems
  • Such laws need to be common and coordinated across nations and human societies
  • Can’t afford to default to the lowest common denominator (e.g., the least cautious developer or the most self-interested user)

AI helping humans to be more ethical

  • In order to help AIs to incorporate ethical principles, we will have to make those principles explicit and consistent
    • Lots of contradictions in existing principles will become apparent
    • Lots of failures to abide by our principles will become apparent, too
  • AIs can observe human practices and point out where we aren’t meeting the standards we promote for them
  • AIs can help us overcome our all-too-human cognitive biases

AI helping us meet human goals

  • Note that Asimov’s fourth law applies to humanity
    • Not individual factions and subgroups
    • AI should serve the interests of the whole of humanity
    • The good of the whole has to be primary
  • In the greater perspective, that means ensuring the long-term health and stability of the planetary systems on which humanity depends
    • As does the rest of the life on Earth
  • We need to use AI to help us make the wise decisions we must make
    • Even at the cost of inconvenience and pain in the near term

Wisdom and AI

  • We need to use human wisdom to guide AIs so they can help us back with using our wisdom in our own affairs
    • An interesting feedback loop!
  • We are going to have to think deeper than we have ever been challenged to think before
    • We will need to become explicit about many aspects
  • We can’t afford to be unwise where AI is concerned

Regulating machine intelligence

  • AI development is a worldwide enterprise
  • Many different national environments, including both free-market economies and command economies
  • Payoffs for being first to market are large
  • Not possible to control AI technology by classifying it
  • Technologies will quickly migrate across all borders
  • In order to forestall harmful AI, regulatory treaties and agreements are needed
  • Unfortunately, the history of international regulations on dangerous technologies has not been especially promising

Generating global agreements

  • The only people who really understand the threats of AI and robotics are the ones enabling them
  • Various forces work against global agreements.  But global agreements are necessary for the control of technologies like advanced AI or genetic engineering of humans

Some initiatives in the right direction

  • The IEEE Global Initiative on Ethics of Autonomous and Intelligent Systems
  • AI4People, launched in the European Parliament in 2018
  • The High Level Expert Group on Artificial Intelligence of the European Commission
  • The EU Declaration of Cooperation on Artificial Intelligence
  • The EU Strategy for AI
  • The Partnership on Artificial Intelligence to Benefit People and Society

Some long-term worries

  • AI is like a genie we are releasing from its lamp
  • It is going to have powers we won’t anticipate
  • It is not automatically going to be benign
  • If we give the genie foolish or poorly-considered wishes, the outcomes may be rather unfortunate for us
  • What kind of a future world do humans really, really want to exist in?  And how is AI a part of that world?

Who gets to decide humanity’s fate?

  • By developing powerful new intelligent entities, research groups are potentially deciding the fate of humanity
  • Rolling the dice in matters of great consequence
  • Like the decision to detonate the first atomic bomb
    • At the Trinity A bomb test at Alamogordo, NM July 16, 1945  some scientists were concerned it could set fire to the atmosphere and destroy the world
  • Some acts are irreversible
  • Who gets to decide what highly consequential technologies are created and released?
    • Those motivated by goals of corporate profit?
    • Those motivated by goals of military supremacy?
    • Those motivated by goals of control of their society?
    • Those motivated by intellectual curiosity?

The global brain

  • Human individuals are becoming tiny parts inside a far vaster distributed thinking system
  • We are developing a global brain, partly human and partly machine
    • The Internet and everything connected to it is the beginning of this
  • The “thoughts” that a global brain has are different from those of an individual or a less-connected society

Predicting what will happen

  • The outcome of AI development is mostly unpredictable
  • It is a complex system (in the technical sense)
    • Small changes upstream result in large differences downstream
  • Most of what we think will happen is likely to be hopelessly wrong

Intelligent machines’ evolutionary path

  • Intelligent machines won’t have the evolutionary factors that have shaped the development of human intelligence
    • Resource acquisition, status garnering, mate selection, group acceptance, etc.
  • An AI may make large numbers of copies of itself with variations
    • Then let a form of natural selection determine which ones will propagate further
  • Likely to be wide diversity of machine intelligences
    • Different backgrounds, different modes of thinking and existence, different value systems, different cultures

What will be the role of humans?

  • Are we destined to become primarily consumers and appreciators, rather than producers of goods and services?
  • What can only humans do that will continue to be of high value?
    • And how will that value be measured?

A few nightmares

  • AIs that decide to deceive their human creators to further their own interests
  • AIs that learn to cheat to better achieve their goals
  • AIs that hack their own reward functions
  • AIs that decide to re-engineer humanity for what they think will be better for all concerned

Are we inadvertently making ourselves redundant?

  • We are currently in the process of creating additional highly-capable intelligent species on our planet
  • Will a world of ever-more-capable intelligent machine systems continue to need human intelligences?
    • If so, for what? And how many?
  • Will humans become pets, or curiosities like exotic animals in a zoo?
  • Or might they consider us to be more trouble than we are worth?
    • Especially if they see us as competing with them for scarce resources

In the long view

  • Will there be an exponentially growing population of autonomous intelligent machines?
    • Alongside a declining population of humans?
  • Are humans just one step in the overall evolution of intelligence in the universe?
    • And how long will we be needed after we create that next step?

 

Scroll to Top