Artificial intelligence

by Chris Woodford. Last updated: December 20, 2021.

Does the future still need us? That was the question computer scientist Bill Joy famously posed a couple of decades ago. If the things we read in the media about artificial intelligence can be believed, the answer's a definite "no." Smart computers already grade our exams, help us choose our next YouTube video, and decide what we can post on social media; and they'll soon be doing much more, from powering vast armies of robot soldiers to safely steering our cars and trucks. But just because machines like these do seemingly smart things, does it follow that they really are intelligent? And, even if they can be described that way, is their growing ability a sure signpost to a future of redundant, human stupidity? What exactly is artificial intelligence and why should we care?

Artwork: Can we imagine a world where computers can imagine a world without us? Composite photo by explainthatstuff.com incorporating dummy photo by Conrad Johnson courtesy of US Army and DVIDS.

Listen instead... or scroll to keep reading

What is artificial intelligence?

Some of the real (human) minds that have wrestled with the problem of artificial (machine) intelligence have offered deceptively simple definitions:

• John McCarthy, the US computer scientist who coined the term "artificial intelligence" back in the 1950s, said: "It is the science and engineering of making intelligent machines, especially intelligent computer programs." [1]
• Marvin Minsky, another leading scholar in the field, suggested AI is "the science of making machines do things that would require intelligence if done by men." [2]
• Alan Turing, one of the founding fathers of this field, essentially defined artificial intelligence as the ability of a machine to pass itself off as a completely convincing human. [3]

But what is intelligence?

Of course, to understand definitions like these, we need to reflect a little bit on what we mean by intelligence. When we say someone is "clever," typically we mean they're good at brainy, book-type stuff, although you can certainly be a clever football player, pastry chef, or just about anything else. If we say someone is "talented," we mean they have some sort of natural ability—a born head-start that makes them better than average.

So what does intelligence mean? Effectively, it's a talent for cleverness that you can choose to deploy however you want. You're not just clever at being a lawyer or a historian: you have a general aptitude for stuff that you can apply in different ways at different times. You're like an actor who can play many roles. Today you might be a world-class lawyer; tomorrow, you might turn your hand to learning classical piano or becoming a chess grand master. Like an actor, you have an overarching talent and a conscious, free-willed ability to play any part you choose. So intelligence is a kind of general-purpose thinking talent underpinned by qualities like a good memory, excellent language and reasoning skills, an ability to communicate with other people, often combined with creativity, emotional awareness, morality, self-awareness, and so on. A clever person might be good at just one thing (because they've been doing it a long time) and flounder when they try something else; an intelligent person has the potential to be good at anything and everything (or, at least, many different things), even when they've not been explicitly trained in those things.

Photo: Map-reading, fire-fighting, flying a plane, playing football—which of these things takes intelligence? They all do. We could build "clever" computers or robots to do all these things, but they still might not be what we consider "intelligent." Yet if we could build a machine that could do all these things, and many more, without any human intervention, would that be truly "intelligent"... or is there more to it than that? Photos by William Johnson, Christopher Quail, Trevor T. McBride, and Trevor Cokley, all courtesy of US Air Force.

But "general problem-solving ability" might seem like a vague and woolly definition of intelligence—and some people prefer to be more specific. So, in the field of psychology (the science of human behavior), you'll find more than a few people willing to define intelligence as the ability to pass intelligence (IQ) tests. That's a circular definition, but it's not as cynical as it sounds. An intelligence test is something specific, well known, and well studied. You could take a typical IQ test and break it up into all the different puzzles it contains. You could figure out step-by-step ways of solving each one and train a computer in how to do it. Give your computer an IQ test and it ought to score pretty highly, in theory making it an artificially intelligent machine, at least by Marvin Minsky's definition. Well, no. Because if you gave it a slightly different test it hadn't been trained for, it would probably have no idea what to do and fail, without any of the crushing self-awareness that a human would have in the same situation. Still, we might say it has some of the qualities of an intelligent machine or an intelligent human.

Strong AI

If we define intelligence this way, it means machines with artificial intelligence would have similar general-purpose problem-solving ability: they might behave as human-like brains in computer-like boxes. This way of conceiving things is sometimes called strong AI or artificial general intelligence (AGI) and it's what most people assume we mean when we talk about intelligent machines: if not machine replicas of humans, machines with most (or all of) the intellectual qualities of a thinking human being—and sometimes their emotional and creative qualities too.

Weak AI

The idea of making machines that can behave in clever ways to solve specific, limited problems (like playing chess, driving a car, operating a factory robot, or whatever) is called weak AI or artificial narrow intelligence (ANI). Most of the vaguely "intelligent" computers and robots that we come across at the moment fall into this category. A chess-playing computer like IBM's Deep Blue made world headlines when it defeated grand master Gary Kasparov back in 1997, but much of its cleverness was down to brute-force computation—and just imagine it trying to drive a car down the freeway. By the same token, self-driving Google cars look pretty impressive, but don't expect them to sit down across the chess board anytime soon.

What's the point of AI?

Strong versus weak, general versus narrow, human-like versus machine-like, intelligent versus merely clever—there are all sorts of ways of considering "degrees" of artificial intelligence. We can picture a scale ranging from simple, narrow artificial cleverness (clanking, preprogrammed, machine-like ability to carry out specific tasks, like a Roomba vacuum cleaner scrabbling round your living room without any awareness of what it's doing); through chess-playing computers, self-driving cars, and robots that can rescue people from disaster zones; right up to a supremely, generally artificially intelligent computer that's smart enough to reprogram itself, recognize its flaws and design better versions of itself, and even has some sort of consciousness of what it's doing.

Yet many would argue that this isn't a simple scale, however, and that AGI is entirely different from ANI (qualitatively different), and not just "turbocharged" ANI (quantitatively different). And if that's true, it definitely doesn't follow that our ultimate goal should be to create the most generally, artificially intelligent, humanlike computer we possibly can; that's where much of the debate about AI gets bogged down. Our first goal is to understand our goal. If we want a rescue robot that we can deploy in emergencies to pull people out of a burning building, it needs a selected range of useful behaviours. It probably doesn't need to play chess, drive cars, or speak Spanish, though the more things it can do, obviously the better it can improvise in unpredictable situations, just like humans.

One goal of AI is simply to do something that humans do as well as they do it, without getting hung up on trying to do everything that humans do better than they do them or doing it exactly as a human would. "Strong" is not automatically better than "weak," even if the use of those emotionally loaded words make it sound so. Weak is a misnomer. There's nothing "weak" about a world-class human chess player if all you want is someone who can play chess. A "weak" Roomba robot can make a very nice job of cleaning your carpet. You can take a very pragmatic approach to AI, in other words, without worrying excessively about wider philosophical questions, even if those things are of great interest to philosophers. In other words, you can program a computer to play chess without sitting down to define the word "intelligent" or worrying about just how intelligent your computer will turn out to be. The important thing is to be true to our own goals, whatever they might be.

One of the often-overlooked goals of AI—overlooked in the popular media, at least—is to shed light on how our own brains work through the scientific study of how machines could mimic them. Since the 1950s, psychology has undergone a "cognitive" revolution in which quite a bit of the mystery hidden in our heads has been unraveled by research that assumes brains process information in similar ways to computers. AI research sheds more light on cognitive psychology and neuropsychology in a similar way: by figuring out how machines could be intelligent, we learn more about how humans are already intelligent.

Sponsored links

Types of artificial intelligence

Since the dawn of the field in the 1950s, most real-world, AI computer programs have fallen into several broad types, which (for the sake of simplicity) I'm going to categorize into just three: heuristic search, expert systems, and machine learning. There are also hybrid systems that combine two or more of them. Let's consider these in turn.

Heuristic search

How do you build a computer that can play chess? If you play chess yourself, your strategy is probably to arrive at each move by considering every move you could possibly make, moves those might lead to, and so on, running your lithe, monkey mind along a tree of possibilities until you figure out the move most likely to win the game. If you have unlimited brainpower, you could theoretically consider every possible move and rank them accordingly. But, do the math, and you'll find the number of possible moves is well beyond the limits of memory and time. According to Marvin Minsky, writing about this problem back in 1966, we're talking something like 10¹²⁰ moves, whereas even a simpler game like checkers can come in at 10⁴⁰ possible moves. [7]

Photo: Chess-playing computers typically use heuristic search. Modern chess programs often work the same way as similar programs designed in the 1960s, but because today's machines are much more powerful, they can consider far more moves in the same amount of time. That's essentially why today's computers are better than yesterday's.

Our brains can't process so much stuff—or anything like it. So instead of considering every possible move, you (or a computer) can use basic rules of thumb to narrow down the searches you make, turning an overwhelming problem into something your brain (or a processor chip) can reasonably handle. This is called a heuristic search. An obvious heuristic most of us apply in game situations, if only out of consideration for the people we're playing with, is to spend at least a certain amount of time looking for moves but not too much. Or, getting more complex, you might use strategies like trying to dominate the center of the board, recognizing certain key board patterns, or preserving important pieces like your bishops and queen at the expense of losing less-valuable pieces. The key point about heuristic search is that it settles on a good-enough solution in the time available rather than trying to find the one and only perfect outcome.

Heuristic search is great for logical board games like chess, checkers, Scrabble, and so on, which involve considering lots of very similar potential moves, but how useful is it in the real world? You can imagine an artificially intelligent app that searches through thousands of houses and flats for sale or rent to help you find the best one using a heuristic approach to narrow things down. Instead of showing you everything, it might show you homes like ones you've tended to look at before within a certain radius or price bracket. But what about more complex problems like offering legal advice, figuring out what's wrong with a broken-down car, or diagnosing illnesses?

Expert systems

You can probably see straight away that medical diagnosis doesn't necessarily lend itself to a simple heuristic search: if you're a critically ill patient, you're not looking for a good-enough diagnosis in the time available; you want the right diagnosis, however long it takes. Your doctor doesn't think "Well I'll consider the first 10 diseases that pop into my head and then just pick the most likely one."

Expert systems (sometimes called knowledge based systems or KBS) are computer programs designed to go beyond simple search and decision making using more detailed "if X then Y" analysis and reasoning. Typically they have a database of knowledge gleaned from studying real, human experts and a separate system that can reason by dipping into that knowledge. The limitation of expert systems, and it's not always a problem, is that a computer trained in one domain of knowledge (like legal advice or medical diagnosis) isn't any use to us in a different field. It's perfectly possible for humans to change careers and switch from being brain surgeons to corporate lawyers, but expert-system machines can't voluntarily do the same thing without swapping their databases. Indeed, some medical knowledge is so very complex, so expert, that even an expert system trained in one medical field (say, cancer) might be of limited use in another medical field (emergency medicine). While that's true of human doctors, the key difference is that human experts tend to recognize their own limitations and know when to ask for help; machine experts don't know when they're making incompetent decisions.

Machine learning

Machine learning is one of the buzzwords of cutting-edge AI, although it's actually a very old term that dates back to the 1950s. In practice, it often means training a neural network (very loosely, a hugely simplified computer model of a brain-like structure, made from layers of interconnected cells called "units") on millions, billions, or trillions of examples of something so it can quickly recognize or classify something it hasn't seen before. So, trained to look at millions of pictures of tables and chairs, it can tell you whether a photo it's never seen before shows a table or a chair—and that's how the automatic photo classification algorithms work on your phone. Machine learning explains how Google can filter out explicit adult images from your search results if you don't want your family to see them: algorithms trained on adult photos can spot tell-tale signs in other photos as people upload them. Machine learning also underpins the automated translation you can find on Google, Bing, and Skype. Programs like these are now so good that they can convert almost any language into a fairly decent (at least understandable) translation of any other language without "understanding" a word of either; they're great examples of Searle's Chinese room, except they're capable of operating in any tongue you like.

One of the characteristics of machine learning programs is—the clue is in the name—that they learn as they go along. So unlike a preprogrammed expert system, they get better and better at what they do the more they do it, just like a real human, to the point where they can do whatever you program them to do better than you can do it yourself.

A few more things are worth quickly noting. Although the terms "machine learning" and "neural network" sound like they stem from psychology, they're much more to do with complex math and statistics. And when we talk about "neural networks" being "brain-like," that's really just an analogy. There's not necessarily anything brain-like in a neural network (synapses that work like one-way streets and chemical neurotransmitters are two very obvious differences, to start with). What neural machine learning and human brains do have in common is that they both process large amounts of data in parallel (or "pseudo-parallel," in the case of neural networks, which are usually models of parallel, brain-like structures implemented on traditional, serial computers).

Artwork: How would an AI computer/robot go about climbing a tree? Left: It could use heuristic search to consider a number of the most likely paths along the branches; Middle: It could use an expert system database of knowledge about trees and climbing, and some IF... THEN... rules to arrive at the best route by reasoning; Right: Using machine learning, perhaps it could analyze thousands of photos of people climbing trees to figure out a good route when presented with a new photo of a tree?

Hybrid systems

I've adopted a fairly arbitrary, three-way classification here purely for the purposes of an easy-to-understand explanation. But you can classify AI however you wish. Heuristic search and expert systems are examples of what some call symbolic AI or Good Old-Fashioned AI (GOFAI), and work in a classically, cognitive fashion like the "this-leads-to-that" flowchart diagrams we draw to explain simple computer programs. Machine learning, on the other hand, proceeds in a more parallel, brain-like, "connectionist" fashion, without obvious serial logic. Symbolic AI systems work things out by serial, logic reasoning using a limited amount of data, while connectionist systems use parallel processing on massive amounts of data.

Artwork: Serial versus parallel processing. In traditional symbolic AI (such as an expert system), one step proceeds logically after another; neural-network machine learning uses parallel processing (although, confusingly, it's usually modeled on traditional computers that work through serial processing!

If you're trying to build a self-driving car or a robot that can rescue people from buildings, you're going to be using elements of the three previous types of AI in different ways, at different times. So hybrid systems are increasingly interesting to researchers who might once have worked exclusively with either symbolic, serial AI or parallel, connectionist, machine-learning systems. DeepMind's Atari-game playing system is one very recent example of a hybrid system that works partly through symbolic AI and partly through a neural network.

Self-driving cars rely on machine learning to interpret images of the streets they're driving down in real time. We've all see those Google captcha tests that get us to prove we're not robots by classifying fire hydrants, bridges, bicycles, and taxis. That exercise is part of Google's effort to train machine learning systems to recognize different objects that their cars will be able to tap into later. But driving isn't just about recognizing objects; you also need to recognize and analyze situations. For example, if you're driving along and you see a round red object rolling out into the road, you might think "Oh look, there's a ball", but what you should really think is "That ball probably belongs to a child and if the ball's rolled onto the road, a child may be right behind it, so I need to slow down and be prepared to stop". This is more like expert system decision making. A self-driving car will probably always have human occupants and a backup human driver to get it out of trouble, so it doesn't need to be completely autonomous in the same way as a robot soldier, which might need to extricate itself from a wider variety of unexpected situations. So, revisiting our original idea of artificial intelligence as a spectrum between narrow and general, we can see that the more autonomous and general purpose a computer or machine, the wider the range of different AI tactics or techniques it's likely to need to draw on. And it will also need the ability to figure out which type of thinking to use in different situations.

One key problem—for all types of AI—is representing some aspect of the world in a way that a computer system can understand and process. At a simple level, if you're using a neural network to recognize faces, how exactly do you "translate" holistic faces into discrete bits of data that the network can work with? And how do you convert the computer network's output into a form that makes sense to humans?

What is AI used for in the real world?

When John McCarthy died in 2011, an obituary in the British Independent newspaper noted how he'd once observed that a major breakthrough in the field could come in anything from "five to 500 years." McCarthy set his sights on the very distant horizon of strong AI and despaired at researchers who satisfied themselves with narrower goals. As he wrote in 2006:

"I have to admit dissatisfaction with the lack of ambition displayed by most of my fellow AI researchers....For example, the language used by the Deep Blue program that defeated world chess champion Garry Kasparov cannot be used to express "I am a chess program, but consider many more irrelevant moves than a human does." and draw conclusions from it. The designers of the program did not see a need for this capability." [8]

Just because we don't have computers that are smart enough to go head-to-head with humans at anything and everything, it doesn't mean we've made no progress with artificial intelligence because, as we've seen already, strong AI was not always the goal. Look around the modern world and you'll see endless applications of less spectacular (but still very impressive) artificial intelligence, including things like Alexa, Siri, and machine-driven customer agents, automated stock trading, "Things you might like" recommendations on Amazon and eBay, online advertising software that shows you ads based on who you are and what you're likely to buy, vacuum cleaning Roomba robots that use increasingly intelligent tactics to clean more efficiently, and much more.

Why bother with artificial intelligence?

Photo: Roomba vacuum cleaning robots use increasingly sophisticated models to clean your home. We'd hardly call them intelligent, but why would we want to?

In a world packed with intelligent humans, why bother developing intelligent machines? Is it a sign of our own intelligence that we recognize our potential stupidity as a limitation we can overcome? Or yet another sign of human arrogance—that we somehow always think we can do better than nature? Is the argument between "supporters" and "opponents" of machine intelligence two different sides of that arrogance: that humans are special because we think we can develop machines better than ourselves but can't... or special because we really can't? Is the quest for artificial intelligence rather like the quest for the perfect move in a long and difficult game of chess... something we feel that's out there somewhere, always just beyond reach? Or is it more like a heuristic search in a board game, where quick, practical, good-enough solutions to limited problems are better than waiting around for perfection?

AI raises all kinds of philosophical questions, but it raises social problems too, like whether smart factory robots and machine-learning computer systems will put "intelligent" people out of work. There are ethical problems as well. For example, people are already asking difficult questions about the legal responsibilities of self-driving cars. If you get mown down by a car like this on a crosswalk, is the car somehow to blame? Is the passive "driver" sitting with their hands on a dummy steering wheel to blame for not intervening? Or is the designer to blame even if they're on the other side of the world? If robot soldiers kill civilians by accident, does the buck stop with the general who gave the order or the engineers who built and programmed the machines?

One of the big problems we're already seeing is that algorithms using machine-learning can come up with decisions that we have no way to understand or challenge. We know the human assumptions on which the algorithms are based, but there's no easy way of seeing why a neural network trained on billions of disparate items of data will, for example, wrongly flag a world-famous war photo as pornography or scandalously miscategorize a photo of three black teenagers. In the second case, after Google's algorithms tagged people as gorillas, Google didn't even bother trying to fix the problem: it simply stopped its algorithms labeling any photo as a gorilla, chimpanzee, or monkey. The AI problem was too hard to solve, so they solved an easier problem instead.

What next for artificial intelligence?

While pragmatic computer scientists get on with building ever cleverer machines, philosophers continue to wrestle with endless variations on essentially the same stale question: whether ingenious (or brute-force) computational "cleverness" (the cake of AI) can truly replicate human "intelligence" if it can't replicate quintessential, more subtle human qualities like understanding, empathy, morality, emotion, creativity, free-will, and consciousness (the all-important icing). Perhaps understanding the difference between computational "cleverness" and human "intelligence" is the real Turing test?

Arriving where we started, let's ask again: does the future really need us? Arguably, the biggest limitation of today's relatively weak AI systems is that they don't recognize their own limitations: for that, they still need us. Until artificially intelligent machines are smart enough to understand how really stupid they can be, perhaps they pose no ultimate threat to us humans.

In the longer term, is humankind at risk from ultra-intelligent AGI machines—or is that just science fiction nonsense, as skeptics like Hubert Dreyfus were arguing over a half century ago? Are such ideas a dangerous distraction from a far more plausible threat: how many livelihoods are at risk as machine-learning-type algorithms become increasingly "clever" at doing jobs we once regarded as absolutely human? The jury is still divided. Where pessimists like Bill Joy have warned that AI is a Pandora's box, optimists, such as AI "prophet" Ray Kurzweil, look to the singularity—effectively, where machines surpass human intelligence—and a bold, rosy future where "stupid," intractable human problems like war and poverty are deftly swept aside by brainy machines. Meanwhile, pragmatic robot scientists such as MIT's Rodney Brooks argue that machine intelligence is simply the latest human technology that helps people overcome their all-too-human limitations: "We will become a merger between flesh and machines. We will have the best that machineness has to offer but we will also have our bioheritage to augment whatever level of machine technology we have so far developed." [10]

So the question isn't really whether the future still needs us; it's "What kind of future do we want?"—and how can we use technologies like AI to bring it about?

AI timeline: A brief history of artificial intelligence

Early days

• 1637: Wondering about the possibility of machines that can imitate people, French philosopher René Descartes feels certain we will always be able to tell the difference. He anticipates the Turing test by over 300 years when he writes: "If there were machines which bore a resemblance to our body and imitated our actions as far as it was morally possible to do so, we should always have two very certain tests by which to recognize that, for all that, they were not real men." Broadly speaking, he argued that 1) machines cannot respond to everything you might say to them or 2) anticipate and cope with every situation they might meet. [11]
• 1737: Jacques de Vaucanson, a French artist and inventor, builds delightful automata, including a mechanical flute player and a "digesting duck" (a realistic eating, drinking, walking model of a bird).
• 1763: Thomas Bayes develops "Bayesian" inference ( a type of probabilistic reasoning), which will form an important part in artificial intelligence programs and algorithms in centuries to come.
• ~1770: Pierre Jaquet-Droz dazzles emperors and kings with intricate, mechanical, animated dolls that do astonishing things using simple programmable memories.
• 1842: Charles Babbage and Ada Lovelace develop mechanical, gear-driven computers that can be programmed and reprogrammed.
• 1910: Spanish inventor Leonardo Torres y Quevedo develops El Ajedrecista, an early chess-playing machine.

Cybernetics

• 1940s: Physiologist William Grey Walter builds autonomous mechanical "tortoises" that can find their way around using photocells and simple electronics.
• 1943: Physiologists Warren McCulloch and Walter Pitts build simple, algorithmic models of brain cells that can learn things, so developing the first primitive neural networks.
• 1943: Cambridge psychologist Kenneth Craik outlines the concept of internal "mental models" in a book called The Nature of Explanation. The idea of machines that can use models of the world to solve problems of various kinds proves highly influential.
• 1948: Cybernetics (simply, developing biologically inspired machines that use sensing, learning, adaptation, feedback, and so on) is officially born with the publication of Norbert Weiner's classic book Cybernetics: Or Control and Communication in the Animal and the Machine.

The rise of intelligent machines

• 1949: Edmund Berkeley's book Giant Brains: Or Machines That Think popularizes the idea of computers as electronic brains.
• 1950: Alan Turing sets out the "Imitation Game" (now called the Turing test) in a hugely influential paper called "Computing Machinery and Intelligence."
• 1950s: Sci-fi writer Isaac Asimov proposes three laws of robotics to help keep robots in check.
• 1950s: British scientist Oliver Selfridge develops a computational model of pattern recognition called Pandemonium, which influences neural networks and machine learning.
• 1955: Allen Newell explores heuristic-search approaches to playing chess in The Chess Machine: An Example of Dealing with a Complex Task by Adaptation.
• 1956: John McCarthy coins the term "artificial intelligence" during a groundbreaking workshop at Dartmouth College in Hanover, New Hampshire.
• 1958: Computer scientist John von Neumann coins the term "singularity."
• 1959: Arthur Samuel coins the expression "machine learning" to describe computer programs that gradually get better at games than the people who program them.
• 1961: Mortimer Taube's polemic Computers and Common Sense: The Myth of Thinking Machines accuses AI scientists of "writing science fiction to titillate the public" and wasting public money.
• 1962: John McCarthy establishes Stanford University as a major research center in AI with the formation of Stanford Artificial Intelligence Laboratory (SAIL).
• 1965: Joseph Weizenbaum develops ELIZA, the first "chatbot" program that can hold a vaguely human conversation.
• 1965: Stanford's Edward Feigenbaum develops the first expert system, DENDRAL, designed to identify unknown molecules by reasoning from various input data using its pre-learned, expert knowledge of chemistry.
• 1967: Frank Rosenblatt develops the Mark I Perceptron, a brain-like computer based on a neural network design.
• 1968: Interest in neural networks stalls following publication of Perceptrons, a seminal book by Marvin Minsky and Seymour Papert, which attacks Rosenblatt's work.
• 1968: Tom Evans develops an AI program called ANALOGY that can solve geometric problems commonly used in IQ tests.
• 1968: Stanley Kubrick and Arthur C. Clarke's dystopian 2001: A Space Odyssey offers a scary vision of what happens when an artificially intelligent computer, HAL 9000, goes out of control.
• 1970: Terry Winograd's SHRDLU program learns to understand a simple world made of blocks using ordinary, "natural language."
• 1972: Hubert Dreyfus refutes the whole idea of intelligent machines in a provocative book What Computers Can't Do: A Critique of Artificial Reason (revisited in 1992 as What Computers Still Can't Do).
• 1970s: MYCIN expert system is developed to diagnose and suggest treatments for blood diseases.
• 1980: Philosopher John Searle outlines the Chinese room, his classic objection to "strong AI."

Modern AI

• 1980s: Neural networks become hugely popular again thanks largely to the "parallel distributed processing" or "connectionist" models of James McClelland, David Rumelhart, Ronald Williams, and Geoff Hinton.
• 1980s: John Laird, Alan Newell and others develop SOAR, a set of "building blocks" that could be used to create AGI.
• 1987: Digital Equipment Corporation (DEC) develops R1, an expert system to help design its VAX computers.
• 1988: IBM researcher Peter Brown pioneers automated, "statistical" language translation using machine learning.
• 1989: Roger Penrose, a British mathematician, argues that artificial intelligence cannot be replicated by a conventional, Turing-style computer in a bestselling book named The Emperor's New Mind.
• 1990: US computer scientist Ray Kurzweil popularizes the singularity—effectively the point at which artificial intelligence overtakes the human kind.
• 1990: Robot pioneer Rodney Brooks rejects symbolic AI and embraces pragmatic, "bottom-up," "situated" AI in a provocative paper and book called Elephants Don't Play Chess.
• 1995: Inspired by ELIZA, a chatbot called ALICE (Artificial Linguistic Internet Computer Entity) sets new standards for machine conversation, though it still cannot pass the Turing test. Some years later, it spawns the Spike Jonze film Her.
• 1997: IBM's Deep Blue supercomputer beats Russian grandmaster Gary Kasparov in a New York chess tournament.
• 2002: Rodney Brooks' iRobot releases the first Roomba robot vacuum cleaner.
• 2009: Apple Computer buys Siri, developed from CALO (Cognitive Assistant that Learns and Organizes) at Stanford Research Institute.
• 2011: IBM's Watson program trounces its human opponents in the TV quiz show Jeopardy!
• 2012: AI researchers Jeff Dean and Andrew Ng show that a trained neural network develops a preference for cat pictures.
• 2012/3: A series of polls of AI researchers suggests a 50 percent chance that strong AI (artificial general intelligence) will be developed around 2040–2050.
• 2014: A chatbot named Eugene Goostman, developed by Vladimir Veselov, (arguably) passes the Turing test.
• 2015: Conversational neural networks make the news. Baidu's Minwa uses the technique to classify images more accurately than people, while Google's DeepDream develops the spooky human knack of seeing images that aren't really there—faces in clouds and so on.
• 2016: Google's DeepMind beats top player Lee Seedol at the board game Go. Seedol subsequently retires arguing that AI "cannot be defeated."
• 2020: DeepMind succeeds at the problem of protein folding, potentially speeding the development of novel medical drugs.

Don't want to read our articles? Try listening instead

If you'd rather listen to our articles than read them, please subscribe to our new podcast on Apple Podcasts, Spotify, Audible, Amazon, Podchaser, or your favorite podcast app, or listen below:

Find out more

On this website

• How computers work
• History of computers
• Internet and brain
• Neural networks
• Psychology
• Robots

On other sites

• Automaton: AI, robotics, and related topics from IEEE Spectrum.
• Professor John McCarthy: Father of AI: An archive of John McCarthy's papers and other writings.

Articles

Popular science and news

• Why is AI so dumb?: A special report in IEEE Spectrum's October 2021 issue featured several highly critical and thought-provoking articles:
  • The Turbulent Past and Uncertain Future of Artificial Intelligence by Eliza Strickland.
  • How deep learning works by Samuel K. Moore et al.
  • How DeepMind Is Reinventing the Robot by Tom Chivers.
  • 7 Revealing Ways AIs Fail by Charles Q. Choi.
  • A Human in the Loop by Rodney Brooks.
  • Deep Learning's Diminishing Returns by Neil C. Thompson et al.
• IBM Develops a New Chip That Functions Like a Brain by John Markoff. The New York Times. August 7, 2014. IBM's experimental TrueNorth chip uses a neural network architecture.
• Why the Future Doesn't Need Us by Bill Joy. Wired. April 1, 2000. Bill Joy anticipates human redundancy.
• Scientific American publishes great articles about AI, usually written by leading lights in the field, roughly once a decade. Take a look at Artificial intelligence by Marvin Minsky. Scientific American, September 1966; Artificial intelligence by David L. Waltz. Scientific American, October 1982.

Selected scholarly papers

• [PDF] Computing Machinery and Intelligence by Alan Turing, Mind 49, pp.433–460.
• The logic theory machine—A complex information processing system by A. Newell and H. Simon, A. Newell and H. Simon in IRE Transactions on Information Theory, Vol. 2, No. 3, pp. 61–79, September 1956.
• [PDF] DENDRAL: a case study of the first expert system for scientific hypothesis formation by Robert Lindsay et al, Artificial Intelligence 61 (1993) pp.209–261.
• [PDF] Elephants Don't Play Chess by Rodney A. Brooks, Robotics and Autonomous Systems 6 (1990) pp.3–15.
• The Anatomy of A.L.I.C.E. by Richard S. Wallace in Epstein R., Roberts G., Beber G. (eds) Parsing the Turing Test. Springer, Dordrecht.

Books

For older readers

• Human Compatible: Artificial Intelligence and the Problem of Control by Stuart Russell. Penguin, 2019. How much risk does AI pose to the future of humanity?
• Artificial Intelligence: A Very Short Introduction by Margaret A. Boden. Oxford, 2018. I'd describe this as more summary than introduction, since it assumes quite a lot of knowledge on the part of the reader.
• Artificial Intelligence: A Modern Approach by Stuart Russell and Peter Norvig. Pearson, 2016.
• The Emotion Machine: Commonsense Thinking, Artificial Intelligence, and the Future of the Human Mind by Marvin Minsky. Simon & Schuster, 2007.
• Flesh and Machines: How Robots Will Change Us by Rodney A. Brooks. Vintage, 2003.
• The Computer and the Mind by Philip Johnson-Laird. Fontana, 1993. A great introduction to cognitive science—the meeting point of computer science and experimental psychology. This wonderful book covers computational theories of the mind, including the key concepts of cognitive science, and also looks at the question of how robots could be taught to behave in human-like ways.
• The Age of Intelligent Machines by Raymond Kurzweil. Viking, 1992. The past, present, and potential future of artificial intelligence as seen by one of its most provocative supporters.
• Artificial intelligence by Patrick Henry Winston. Addison Wesley, 1984. The all-time classic introduction remains relevant today.

For younger readers

• Artificial Intelligence and Entertainment by Tammy Enz. Capstone, 2019. A 32-page intro for ages 8–10 produced in school library format.
• Artificial Intelligence: Thinking Machines and Smart Robots with Science Activities for Kids by Angie Smibert. Nomad, 2018. A fun, clearly written, and accessible introduction with interactive features like QR codes and a DIY Turing test. Aimed at ages 10+.
• Cutting-Edge: Artificial Intelligence by Anna Leigh. Lerner, 2018. Another 32-page guide for ages 8–10, also in classic school library format.

References

1. ↑   What is AI? Basic Questions by John McCarthy.
2. ↑   Semantic Information Processing by Marvin Minsky. MIT Press, 1968, p.v.
3. ↑   [PDF] Computing Machinery and Intelligence by Alan Turing, Mind 49, pp.433–460.
4. ↑   [PDF] Computing Machinery and Intelligence by Alan Turing, Mind 49, pp.433–460.
5. ↑    Searle, J., 1980, "Minds, Brains and Programs," Behavioral and Brain Sciences, 3: pp/417–57. For a broader discussion, see The Chinese Room Argument by David Cole. Stanford Encyclopedia of Philosophy, 2004/2020.
6. ↑   Robot: The Future of Flesh and Machines by Rodney Brooks. Penguin, 2002, pp.179–180.
7. ↑   Artificial intelligence by Marvin Minsky. Scientific American, September 1966.
8. ↑   The Philosophy of AI and the AI of Philosophy by John McCarthy, Stanford University, June 2006.
9. ↑   Artificial Intelligence: A Very Short Introduction by Margaret A. Boden. Oxford, 2018, p.129.
10. ↑   Robot: The Future of Flesh and Machines by Rodney A. Brooks. Allen Lane/Penguin, 2002., p.x.
11. ↑   Discourse on Method and The Meditations by René Descartes. Penguin, 1968, p.73.