History of artificial intelligence: from Turing to ChatGPT

The history of artificial intelligence is about 75 years of bursts and stalls. It begins with Alan Turing asking in 1950 whether machines can think, and with a 1956 summer workshop at Dartmouth that gave the field its name. Two waves of hope ended in "AI winters" when promises outran what computers could do. A third wave, built on learning from data, began in 2012 and led to the chatbots of the 2020s.

In short

• AI has moved in cycles: big promises, a rush of funding, limits nobody had foreseen, then a cold spell called an AI winter.
• Two approaches traded places: writing rules by hand, and letting machines learn patterns from examples. Learning won from the 2010s.
• Modern AI took three things at once: large neural networks, huge amounts of data, and fast chips first built for video games.

How did the idea of thinking machines begin?

The ideas came before the computers that could test them. In 1943, Warren McCulloch and Walter Pitts described a simple mathematical model of a neuron that switches on when its inputs add up past a threshold, the seed of every later neural network. In 1950, Alan Turing published "Computing Machinery and Intelligence", which opened with the question "Can machines think?" and replaced it with a practical test, the imitation game now called the Turing test. Turing guessed that in about fifty years a computer would fool an average questioner often enough to make the question moot.

The field got its name in a funding proposal. On 31 August 1955, John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon asked for "a 2 month, 10 man study of artificial intelligence" at Dartmouth College the next summer. They wrote that "every aspect of learning or any other feature of intelligence" could in principle be described precisely enough for a machine to simulate it. The 1956 workshop solved little, but it gathered the people who would lead artificial intelligence research for decades.

What happened in the first boom and the first winter?

The first programs looked like fast progress. The Logic Theorist, shown in 1956, proved theorems from Whitehead and Russell's Principia Mathematica. Frank Rosenblatt's perceptron, from 1958, learned to sort simple images. Joseph Weizenbaum's ELIZA, from 1966, held typed conversations by turning a user's words back into questions, and some users confided in it. Leading researchers predicted human-level machines within a generation.

The problems were harder than they looked. Programs that worked on toy puzzles collapsed on real ones, because the number of possibilities exploded and computers were small and slow. In 1969, Minsky and Seymour Papert's book Perceptrons showed the limits of single-layer networks, and interest in neural networks faded. In Britain, a 1973 report by the mathematician James Lighthill judged that AI had not delivered on its promises, and government funding was cut. American funding also narrowed toward short-term military projects. The lean years that followed became known as the first AI winter.

What were expert systems, and why did a second winter follow?

In the 1980s, AI came back as a business. Expert systems captured the rules of human specialists as long chains of "if this, then that" and applied them to narrow jobs such as configuring computer orders or diagnosing infections. Companies built AI departments, and Japan launched its Fifth Generation computer project in 1982. In 1986, David Rumelhart, Geoffrey Hinton and Ronald Williams showed in Nature how networks with hidden layers could learn by backpropagation, which revived neural networks in research.

Expert systems proved brittle. Every rule had to be written and kept up by hand, they failed on cases their builders hadn't foreseen, and the specialized computers they ran on lost out to cheaper desktop machines. By the end of the decade, the market had shrunk and funding dried up again: the second AI winter.

How did learning from data take over?

From the 1990s, AI moved from writing rules to learning them. Machine learning uses statistics to find patterns in examples, and it improved steadily on tasks such as speech recognition and spam filtering. IBM's Deep Blue beat the world chess champion Garry Kasparov in 1997, though mostly by searching millions of positions a second rather than by learning. Meanwhile the web was producing data on a scale no lab could have collected.

The turning point came in 2012. A neural network called AlexNet, from Hinton's group in Toronto, won the ImageNet image-recognition contest with a top-5 error rate of 15.3%, against 26.2% for the next-best entry. It had 60 million adjustable numbers and was trained on graphics chips. Within a few years, nearly every leading lab had switched to deep learning, the use of networks with many layers. In 2016, DeepMind's AlphaGo beat Lee Sedol, one of the world's best players, at Go, a game long thought too subtle for computers.

How did we get from transformers to chatbots?

In 2017, researchers at Google published "Attention Is All You Need", which introduced the transformer, a network design that reads all the words of a passage at once and learns which ones matter to each other. It trained far faster on large amounts of text and became the basis of GPT and most models since. Through 2018 to 2020, language models grew from millions of adjustable numbers to hundreds of billions.

OpenAI released ChatGPT on 30 November 2022, and millions of people tried it within weeks. It made AI something ordinary people used every day, and set off a race among technology companies. In 2024, the Nobel Prize in Physics went to John Hopfield and Geoffrey Hinton for foundational work on neural networks, and half of the chemistry prize went to Demis Hassabis and John Jumper of Google DeepMind for predicting protein structures with AI, the other half to David Baker for designing new proteins.

Year	Event	Why it mattered
1943	McCulloch and Pitts model a neuron	First mathematical neural network
1950	Turing's "Computing Machinery and Intelligence"	Framed the question and a test
1956	Dartmouth workshop	Named the field
1958	Rosenblatt's perceptron	A machine that learns from examples
1966	ELIZA	Early chatbot, and early worry about people trusting one
1973	Lighthill report	Funding cuts; the first AI winter
1980s	Expert systems boom, then bust	The second AI winter
1986	Backpropagation paper	Multi-layer networks learn
1997	Deep Blue beats Kasparov	Brute-force search wins at chess
2012	AlexNet wins ImageNet	Start of the deep-learning era
2016	AlphaGo beats Lee Sedol	Learning wins at Go
2017	The transformer	The design behind today's language models
2022	ChatGPT	AI goes mass-market
2024	Nobel prizes for AI work	Science honors the field's methods

Why did progress come in bursts?

Each burst came when an idea met the computing power and data to make it work, and each winter came when promises outran both. Neural networks were proposed in the 1940s and 1950s and trained properly in the 1980s, but they only beat everything else in 2012, once chips were thousands of times faster and the web supplied millions of labeled images. Hand-written rules hit a different wall: the world has more exceptions than anyone can write down.

The pattern also has a human side. Bold predictions attract money, and disappointment takes it away. Herbert Simon, one of the field's founders, wrote in 1960, and repeated in 1965, that machines would be capable "within twenty years, of doing any work that a man can do." That forecast missed by decades, and its failure fed the backlash.

Still argued

Whether the current wave will run into its own winter is debated. Some researchers argue that bigger models, more data and more computing will keep paying off. Others argue that today's systems hit limits in reasoning and reliability that more scale won't fix, and that investment has outrun real returns. The history shows both patterns: ideas that took decades to pay off, and booms that ended in a bust.

Check yourself

• What caused the first and second AI winters, and what did they have in common?
• Why did neural networks, an idea from the 1940s, only start winning in 2012?
• What did the transformer change, and which products are built on it?

Sources

• Computing Machinery and Intelligence — A. M. Turing, Mind, 1950
• A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence — McCarthy, Minsky, Rochester and Shannon, 1955
• Learning representations by back-propagating errors — Rumelhart, Hinton and Williams, Nature, 1986
• ImageNet Classification with Deep Convolutional Neural Networks — Krizhevsky, Sutskever and Hinton, NeurIPS, 2012
• Attention Is All You Need — Vaswani et al., 2017
• The Nobel Prize in Physics 2024, press release — The Royal Swedish Academy of Sciences, 2024