AI History at a Glance

Eight decades of artificial intelligence — and several thousand years of dreaming before that — in one page. By the end you will be able to tell the whole story in two minutes, and explain why it keeps swinging between boom and bust.

In the summer of 1956 a handful of researchers gathered at Dartmouth College in New Hampshire, convinced that “a significant advance” on making machines think could be made if “a carefully selected group of scientists” worked on it together for a summer. Seventy years later we carry assistants in our pockets that write essays, translate Bahasa Melayu into Tamil and pass exams — yet nobody would say the problem is solved. How did we get from that confident summer to here? Not in a straight line. The history of AI is a story of bold promises, painful collapses and quiet, stubborn progress, repeated several times over.

Why the history matters

It is tempting to treat AI as something that started with ChatGPT in November 2022. But almost every idea inside today’s systems is old. The artificial “neuron” dates from 1943. The idea of a machine that learns from experience was sketched by Alan Turing in 1948 and 1950. The method that trains modern neural networks, backpropagation, was popularised in 1986. Even the arguments we have today — can a machine really understand? Will it take our jobs? Is it all hype? — were rehearsed in the 1950s and 1960s.

Knowing the history helps you do two practical things. First, you can place any new claim in context: is this a genuine step, or a replay of an earlier promise? Second, you can see which ideas won, which lost, and why. That is often the best guide to what will last.

The whole arc, era by era

This guide’s framing Historians divide the story in different ways, and the boundaries below are approximate — a winter did not begin on a particular morning. This guide uses nine eras, the same as The Complete AI Timeline; other histories draw the lines differently. Use the chips to filter, and tap any entry for detail.

  1. Antiquity–1920Myths, automata and mechanical reasoningBronze Talos, golden attendants in Homer, the golem, al-Jazari’s machines (1206), Leibniz’s dream of settling arguments by calculation, and Čapek’s play R.U.R. (1920), which gave us the word “robot”. See Ancient Dreams of Thinking Machines.
  2. 1837–1854Babbage, Lovelace and BooleCharles Babbage designs a general-purpose mechanical computer; Ada Lovelace’s 1843 Notes describe programming it and ask whether it could ever “originate” anything; George Boole turns logic into algebra (1854). See Babbage, Lovelace and the Engine.
  3. 1936–1950Turing and the foundations of computingTuring defines what it means to compute (1936), breaks codes in the Second World War, and in 1950 proposes the imitation game. See Alan Turing.
  4. 1943–1950Neurons, feedback and informationMcCulloch and Pitts model the neuron as a logic switch (1943); Wiener founds cybernetics and Shannon founds information theory (both 1948); Hebb proposes how brains learn (1949). See Neurons, Feedback and Information.
  5. 1955–1956Dartmouth names the fieldJohn McCarthy coins “artificial intelligence” in a 1955 funding proposal; the 1956 summer workshop gathers the founders. See Dartmouth 1956.
  6. 1956–1974The golden yearsLogic Theorist, LISP, the perceptron, ELIZA and the robot Shakey. Generous government funding and confident predictions. See The Golden Years.
  7. c. 1974–1980The first AI winterThe UK’s Lighthill Report (1973) and US funding cuts follow unmet promises; the limits of the perceptron, set out by Minsky and Papert (1969), chill neural-network research. See The First AI Winter.
  8. 1980–1987The expert-systems boomRule-based expert systems such as XCON save companies money; Japan launches its Fifth Generation project (1982). See Expert Systems and the Second Winter.
  9. c. 1987–1993The second AI winterThe market for specialised Lisp machines collapses in 1987; expert systems prove brittle and costly to maintain. Meanwhile backpropagation revives neural networks. See Expert Systems and the Second Winter; for the revival of neural networks, Backprop and the Connectionist Revival.
  10. 1990s–2000sThe statistical turnProbability and machine learning replace hand-written rules; Deep Blue beats Kasparov (1997); the web supplies data at scale. See The Statistical Turn.
  11. 2012–2016The deep-learning revolutionAlexNet wins the ImageNet challenge (2012) on GPUs; AlphaGo beats Lee Sedol (2016). See The Deep Learning Revolution.
  12. 2017–2021The Transformer“Attention Is All You Need” (2017) introduces the Transformer; BERT and the first GPTs follow. See Attention Is All You Need.
  13. 2022–todayThe generative eraChatGPT (2022) brings large language models to hundreds of millions of people; image, audio and video generation and AI agents follow. See The Generative Era.

A few threads run through all of it. One is the argument between two broad schools. Symbolic AI held that intelligence is the manipulation of symbols by explicit rules — write down the knowledge and the logic, and the machine will reason. The connectionist or neural school held that intelligence emerges from many simple units learning from examples. Symbolic AI dominated from the 1950s to the 1980s; learning from data has dominated since the 1990s, and neural networks since about 2012. Both schools grew from the same 1940s roots, and today’s systems still borrow from each.

Summers and winters: the pattern

Researchers themselves borrowed the language of the seasons. An “AI summer” is a period of excitement, money and big predictions; an AI winter is the cold spell after the predictions fail, when funding dries up and the very words “artificial intelligence” become something to avoid in a grant application. The term “AI winter” was used publicly at a 1984 meeting of the American Association for Artificial Intelligence, where researchers warned that the expert-systems boom was heading for a fall — which it then did.

The cycle tends to turn the same way each time. Step through it:

HYPEOne turn

Two lessons hide in this loop. The first is that the breakthroughs were real. Winters did not happen because the work was worthless, but because it was oversold. The second is that much of the lasting progress happened during the winters, when nobody was watching. Backpropagation was refined while “AI” was out of fashion; the statistical methods that power search engines and spam filters grew up in the 1990s under other names.

There is also a deeper pattern that the computer scientist Richard Sutton called “the bitter lesson” (2019): over the long run, general methods that make better use of more computation — search and learning — have beaten approaches that try to build in human knowledge by hand. The Lessons from History page takes this up.

Is this another summer?

Since 2012, and especially since 2022, AI has been in the longest and warmest summer it has known. Investment, computing power and public use are far beyond anything in earlier booms As of 2026. That raises the obvious question.

Debated Will the generative-AI boom end in another winter?

Other views
  • “This time is different”: unlike the 1970s and 1980s, today’s systems are already used daily by very large numbers of people and businesses, so demand does not rest on promises alone. Even a financial correction would not erase the working technology.
  • “A correction is coming”: sceptics point to spending that has run far ahead of revenue, to persistent errors such as hallucinations, and to bold forecasts about human-level AI that echo 1960s predictions. On this view, a cooling of investment — if not a full winter — is likely.
  • “Wrong question”: some historians argue the seasons metaphor flatters the past. Funding fell, but research never stopped, and the word “winter” mainly describes the mood of funders and the press.

Speculative Which view proves right will depend on whether current systems keep improving at the rate of the last few years, and whether that improvement turns into lasting economic value. No one knows yet.

Your map of the history section

Each page in this section tells one chapter in more depth. Tap a step to see what it covers; the links underneath take you there.

Quick check

In which year was the term “artificial intelligence” first used, in a funding proposal for a summer workshop?

McCarthy, Minsky, Rochester and Shannon wrote the proposal in August 1955; the workshop itself ran in the summer of 1956.

What usually triggered an AI winter?

The achievements were real but narrow. The collapse came when inflated expectations met hard limits, and funders pulled back.

Which approach dominated AI from the 1950s to the 1980s?

“Good old-fashioned AI” built intelligence from hand-written symbols and rules. Learning from data took over from the 1990s.

Which 2012 event is usually taken as the start of the deep-learning era?

Deep Blue was 1997 and used search, not deep learning; ChatGPT came in 2022. AlexNet’s 2012 win showed what deep networks on GPUs could do.

// Pause and reflect

  1. Pick one bold prediction about AI you have heard recently. Which step of the summer–winter cycle does it belong to?
  2. If much lasting progress happened during the winters, what does that suggest about how we should judge a field by its public mood?
  3. Which do you find more convincing as a description of today’s systems: “rules written by people” or “patterns learned from data”? Why might the answer be “both”?

Your notes are saved privately in this browser.

What happened next

The rest of this section tells the story in order, starting long before electricity: with bronze giants, clay men and clockwork writers. Begin with Ancient Dreams of Thinking Machines.

Key terms on this page

Sources & further reading