A short history of AI: booms, winters and the ideas that finally worked

If you started paying attention to AI when ChatGPT arrived in late 2022, it can look like the technology came from nowhere. It didn't. The field is older than the moon landing, and it has been through at least two cycles of excitement and collapse before this one. Many of the ideas behind today's systems were first written down decades ago and then ignored, sometimes for twenty years.

This article walks through that history in five chapters, not as trivia but as a tool. Knowing which bets failed, which ones worked and why gives you a better way to read today's claims, both the breathless ones and the dismissive ones.

Chapter 1: The founding bet (1950–1970)

The story usually starts with Alan Turing. In 1950 he asked "Can machines think?" and, finding the question too vague to answer, replaced it with a practical test: could a machine hold a written conversation well enough that a judge couldn't reliably tell it apart from a person¹? That move, swapping an unanswerable question about minds for a measurable question about behaviour, still shapes how the field judges progress.

Five years later, four researchers (John McCarthy, Marvin Minsky, Nathaniel Rochester of IBM and Claude Shannon of Bell Labs) asked for funding for a summer workshop at Dartmouth College in New Hampshire. Their proposal, dated August 31, 1955, coined the term "artificial intelligence." It asked for "a 2 month, 10 man study" with a total budget of US$13,500, and said: "We think that a significant advance can be made in one or more of these problems if a carefully selected group of scientists work on it together for a summer"². The problems on their list included language, abstraction and self-improvement. Most of them are still open research areas today. That gap between a summer's ambition and a seventy-year reality is the first lesson of AI history: the difficulty of a problem is very hard to judge from the outside.

Two different approaches grew from this period. The first, which dominated for decades, was symbolic AI: write down knowledge and rules of reasoning explicitly, and let the computer manipulate them. McCarthy's programming language LISP, created in 1958, became the "lingua franca" of this style of AI⁶.

The second was learning from examples. In 1958 Frank Rosenblatt of the Cornell Aeronautical Laboratory published the perceptron, a simple artificial "neuron" that adjusted its internal settings, called weights, each time it got an example wrong³. A neural network is many such units connected in layers. Rosenblatt's version had only one layer of adjustable weights, which limited it to simple patterns, and for most of the next three decades the spotlight stayed on the symbolic approach. Neural networks did not go away, but they moved to the edges of the field.

Chapter 2: The first winter (the 1970s)

Early programs could prove simple theorems and solve toy puzzles, and researchers expected that scaling up would be a matter of time. It wasn't. In 1973 Britain's Science Research Council asked Sir James Lighthill, a Cambridge mathematician, to assess the field. His report was blunt: "In no part of the field have the discoveries made so far produced the major impact that was then promised"⁴.

Lighthill's main diagnosis was what he called the "combinatorial explosion": as a problem gets bigger, the number of possibilities a program must consider grows explosively, so methods that work on toy examples collapse on real ones⁴. A chess position has more possible continuations than any computer could check one by one, and everyday life is far messier than chess.

The years that followed are now called the first AI winter: a stretch when results fell short of promises and money and interest dried up. (The term is a historian's label rather than an exact date range.) Because much early funding came from governments, particularly the military⁶, public assessments like Lighthill's mattered a great deal.

Chapter 3: Expert systems, and the second winter (1980–1995)

AI came back in the 1980s with a more modest pitch. Instead of general intelligence, expert systems captured the know-how of specialists in a specific domain as hundreds or thousands of "if this, then that" rules.

The flagship example was R1, later called XCON, built for Digital Equipment Corporation to configure its VAX computer systems. By 1986 it had processed 80,000 orders at 95–98% accuracy and was reckoned to be saving the company US$25 million a year⁵. Companies built special-purpose "LISP machines" to run AI software, starting with Symbolics in 1981, and Japan's government invested heavily in a "Fifth Generation Computer" project that turned out to be largely unsuccessful⁶.

The boom didn't last. Expert systems were brittle: they knew only what someone had written down, failed badly on cases their authors hadn't foreseen, and became expensive to maintain as the rule sets grew. Even XCON needed costly rewriting⁵. By the late 1980s enthusiasm had faded, and a second AI winter set in.

Quietly, the learning approach was being repaired during this period. In 1986, David Rumelhart, Geoffrey Hinton and Ronald Williams showed in Nature how to train networks with several layers using a method called backpropagation. The method "repeatedly adjusts the weights of the connections in the network so as to minimize" the gap between the network's output and the right answer, and the internal "hidden" units "come to represent important features of the task domain"⁷. In plain terms: the network works out for itself which features matter, instead of a person writing them down. That is the core of modern deep learning, and it was published almost four decades ago.

Chapter 4: Search, statistics and the deep learning breakthrough (1997–2016)

In May 1997, IBM's Deep Blue beat world chess champion Garry Kasparov 3.5–2.5 in a six-game match in New York, the first time a computer had defeated a reigning world champion under standard tournament conditions⁸. It was a landmark, but it did not resolve the old debate. Deep Blue was a purpose-built chess machine: it could play chess and nothing else.

The deeper shift came from learning. Through the 2000s, a small community kept working on neural networks while most of the field used other statistical methods. Canada played an outsized role: Hinton worked at the University of Toronto and Yoshua Bengio at the University of Montreal, and both later received computing's top honour for this work¹⁰.

The turning point arrived in 2012. Alex Krizhevsky, Ilya Sutskever and Hinton entered a deep neural network, later nicknamed AlexNet, into the ImageNet competition, a contest to label photos into 1,000 categories. With 60 million adjustable parameters, trained on graphics processors originally built for video games, it achieved a winning top-5 error rate of 15.3%, compared with 26.2% for the second-best entry⁹. ("Top-5 error" counts an answer wrong only if the correct label isn't among the model's five best guesses.) The gap was so large that within a few years most of computer vision switched to deep learning.

Three ingredients came together: algorithms that had existed since the 1980s, far larger labelled datasets, and cheap parallel computing power. None of them was enough alone.

In 2019 the Association for Computing Machinery gave Bengio, Hinton and Yann LeCun its Turing Award, often described as computing's Nobel Prize, "for conceptual and engineering breakthroughs that have made deep neural networks a critical component of computing"¹⁰.

Learning then cracked a problem search alone could not. Go has far too many possible positions for Deep Blue-style brute force. In 2016 DeepMind's AlphaGo, which combined deep neural networks with search and learned partly by playing against itself, won 99.8% of its games against other Go programs and beat the European champion 5 games to 0, "a feat previously thought to be at least a decade away"¹¹.

Chapter 5: Transformers, chatbots and the scale era (2017–today)

In 2017, a team at Google introduced the transformer, a neural-network design that lets a model weigh the relationships between all the words in a passage at once rather than reading strictly in order¹². Transformers were well suited to the parallel hardware that had powered AlexNet, which made it practical to train much larger models on much more text. Nearly every major language model today is built on this design.

On November 30, 2022, OpenAI released ChatGPT, a model from its GPT-3.5 series fine-tuned with reinforcement learning from human feedback (RLHF): people ranked the model's answers, and the model was trained to produce the kind of answers they preferred¹³. The underlying technology was not new to researchers, but a free, conversational interface put it in front of the public, and the current boom began.

The field's scientific standing changed too. In October 2024, the Nobel Prize in Physics went to John Hopfield and Hinton "for foundational discoveries and inventions that enable machine learning with artificial neural networks"¹⁴. The Chemistry prize the next day was shared by Demis Hassabis and John Jumper "for protein structure prediction" with the AI model AlphaFold2, alongside David Baker for protein design¹⁵. AlphaFold2 can predict the structure of virtually all of the roughly 200 million proteins researchers have identified¹⁵.

What drives the current era is mostly scale. Epoch AI, an independent research group, estimates that the computing power used to train frontier language models has grown about 5x per year since 2020, or roughly 10,000-fold in total (figures as of February 2026)¹⁷. Stanford's 2026 AI Index reports that global corporate AI investment passed US$581 billion in 2025, more than double the year before, and that industry now produces over 90% of notable AI models, up from just under half in 2015¹⁸. Canada, meanwhile, launched the world's first national AI strategy in 2017, built around three institutes: Amii in Edmonton, Mila in Montreal and the Vector Institute in Toronto¹⁹.

What the history does and doesn't tell us

Read as a whole, seventy years of AI history suggest three patterns, each with a caveat.

General learning methods have beaten hand-written knowledge. The researcher Richard Sutton calls this "the bitter lesson": "The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin"¹⁶. Expert systems lost to learned models, hand-built vision features lost to AlexNet, and chess-style search alone lost to AlphaGo's learned intuition. The caveat: this is an interpretation of history, not a law of nature, and it is debated. Whether scaling alone will keep delivering is one of the field's biggest open questions.

Predictions have repeatedly been wrong in both directions. The Dartmouth group expected major progress in a summer². Lighthill found that two decades of promises had not been met⁴. Yet AlphaGo arrived a decade earlier than many expected¹¹, and backpropagation, published in 1986, took a quarter-century to show its full value⁷⁹. The lesson is not "experts are always too optimistic." It's that timelines in this field are very uncertain.

Booms have ended when promises outran real-world value. Both winters followed a gap between what was promised and what worked outside the lab⁴⁵. Today's situation differs in important ways: the systems are in daily use by many millions of people, and the money now comes mostly from private investment rather than government grants¹⁸. But the same question still applies to any specific AI product: does it deliver measurable value in real conditions, or only in a demo?

One limit on this account: histories of AI are written mostly by its practitioners and tend to be told as a march toward the current approach. Ideas that lost, and work outside North America and Britain, get less attention than they deserve.

What this means for you

• When you hear "this changes everything," ask what the system does in real use, at what cost and error rate. Both AI winters began when demos didn't survive contact with messy, real-world cases.
• When you hear "this is just hype," remember that ideas dismissed for decades (neural networks, backpropagation) turned out to work once data and computing power caught up. Dismissal has a poor track record too.
• When you read a prediction with a date on it, treat it as a guess. History gives almost no support to confident timelines in either direction.
• When a product relies on hand-written rules, expect it to be predictable but brittle. When it relies on learning, expect it to be flexible but harder to explain. Know which one you're buying.

Sources

1. Turing, A. M. (1950). Computing machinery and intelligence. Mind, 59(236), 433–460. doi.org/10.1093/mind/LIX.236.433
2. McCarthy, J., Minsky, M. L., Rochester, N., & Shannon, C. E. (1955, August 31). A proposal for the Dartmouth summer research project on artificial intelligence. www-formal.stanford.edu/jmc/history/dartmouth/dartmouth.html
3. Rosenblatt, F. (1958). The perceptron: A probabilistic model for information storage and organization in the brain. Psychological Review, 65(6), 386–408. homepages.math.uic.edu (PDF)
4. Lighthill, J. (1973). Artificial intelligence: A general survey. In Artificial Intelligence: A Paper Symposium. Science Research Council. aiai.ed.ac.uk/events/lighthill1973/lighthill.pdf
5. Huyck, C. (n.d.). Configuration with R1/XCon (1978). Middlesex University course notes (BIS2040). cwa.mdx.ac.uk/bis2040/lect6ES2/xCon.html
6. Computer History Museum. (n.d.). Solving the riddle of AI. Revolution: The First 2000 Years of Computing. computerhistory.org/revolution/artificial-intelligence-robotics/13/290
7. Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1986). Learning representations by back-propagating errors. Nature, 323(6088), 533–536. doi.org/10.1038/323533a0
8. IBM. (n.d.). Deep Blue. IBM History. ibm.com/history/deep-blue
9. Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). ImageNet classification with deep convolutional neural networks. Advances in Neural Information Processing Systems 25. papers.nips.cc/paper_files/paper/2012/hash/c399862d3b9d6b76c8436e924a68c45b-Abstract.html
10. Association for Computing Machinery. (2019, March 27). ACM announces 2018 Turing Award recipients. acm.org/articles/bulletins/2019/march/turing-award-2018
11. Silver, D., Huang, A., Maddison, C. J., et al. (2016). Mastering the game of Go with deep neural networks and tree search. Nature, 529(7587), 484–489. doi.org/10.1038/nature16961
12. Vaswani, A., et al. (2017). Attention is all you need. Advances in Neural Information Processing Systems 30. arxiv.org/abs/1706.03762
13. OpenAI. (2022, November 30). Introducing ChatGPT. openai.com/index/chatgpt
14. The Royal Swedish Academy of Sciences. (2024, October 8). The Nobel Prize in Physics 2024 (press release). nobelprize.org/prizes/physics/2024/press-release
15. Nobel Prize Outreach. (2024). The Nobel Prize in Chemistry 2024: Summary. nobelprize.org/prizes/chemistry/2024/summary, with Utrecht University (2024, October 9), Nobel Prize in Chemistry 2024 for AlphaFold2 and protein design. uu.nl/en/news/nobel-prize-in-chemistry-2024-for-alphafold2-and-protein-design
16. Sutton, R. (2019, March 13). The bitter lesson. incompleteideas.net/IncIdeas/BitterLesson.html
17. Epoch AI. (2026). Trends in artificial intelligence (updated February 5, 2026, accessed October 16, 2026). epoch.ai/trends
18. Stanford Institute for Human-Centered AI. (2026). The 2026 AI Index Report, as summarized in Smith, M. S. (2026, April 13), "12 graphs that explain the state of AI in 2026," IEEE Spectrum. spectrum.ieee.org/state-of-ai-index-2026
19. CIFAR. (2025). Pan-Canadian Artificial Intelligence Strategy. cifar.ca/ai