The history of artificial intelligence is shorter than most people assume and stranger than the tidy timelines suggest. The field has a founding date, 1956, and a founding prediction that it would solve its central problems within a generation. It then spent decades cycling between bold promises, public disappointment and cut funding, before a mix of faster hardware, huge datasets and one clever architecture turned a long-running research program into a consumer product. This is the version without the hype in either direction.

Where did the idea of thinking machines come from?

People imagined mechanical minds long before computers existed, in myth and fiction. The technical story starts in 1936, when Alan Turing described an abstract device, now called a Turing machine, and showed that a single universal machine could carry out any computation that can be written as an algorithm. That result is the foundation of the modern computer, and it made the question “can a machine think?” feel like something with a mathematical shape.

In 1943, Warren McCulloch and Walter Pitts proposed a simple model of artificial neurons. Their model did not learn, but it showed that networks of such units could represent logical functions. Then in 1950 Turing published “Computing Machinery and Intelligence,” which proposed what became the Turing test: if a human questioner cannot reliably tell a machine from a person through conversation, the machine could be said to think. He also suggested that building a simple, childlike program and educating it might be easier than trying to write an adult mind from scratch. That idea reads as strikingly modern.

The Turing test is still cited constantly, but it was a thought experiment, not an engineering benchmark, and researchers have long argued about whether passing it would mean anything.

What happened at Dartmouth in 1956?

In summer 1956, a workshop at Dartmouth College in New Hampshire gave the field its name. John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon had submitted the proposal in September 1955. Its central conjecture was that every aspect of learning, or any other feature of intelligence, could in principle be described precisely enough for a machine to simulate it.

Attendance was patchy. Names associated with the event include Herbert Simon, Allen Newell, Arthur Samuel, John Nash and Oliver Selfridge, with only a few people staying for the whole run, and records differ on who was there when. The workshop was mostly open discussion and produced no single breakthrough. Its lasting effect was to establish “artificial intelligence” as a label and to give the early symbolic approach, the idea that reasoning can be programmed as explicit rules and logic, a flag to rally around. Newell and Simon showed a program called the Logic Theorist there.

Why did early AI promise so much?

The first results were genuinely impressive for their time. Programs proved theorems, played checkers and handled word problems in narrow settings. In 1958, Frank Rosenblatt described the perceptron, an early learning machine inspired by neurons. Funding came readily, especially from the US military and from governments that wanted machine translation and similar tools.

Researchers extrapolated. Succeeding in toy domains, they assumed real-world versions were a matter of scale. That assumption turned out to be badly wrong, because real problems contain endless exceptions and a combinatorial explosion of possibilities. Some leading figures of the time made confident predictions about human-level machines arriving within decades, and those predictions did not age well.

This is the pattern worth remembering. Optimism was not irrational given the early results, but it mistook progress on well-defined puzzles for progress on open-ended understanding.

What was the first AI winter?

Disappointment arrived in stages. In 1966, a US report criticized machine translation, and the National Research Council ended its support after spending roughly 20 million dollars. In 1969, Minsky and Papert’s book *Perceptrons* highlighted the limits of single-layer perceptrons, and research on the approach largely stopped. Nobody at that point knew how to train multilayer networks, which is the problem backpropagation later addressed.

Then, in 1973, Sir James Lighthill reported to the UK Parliament that AI had failed to achieve its grandiose objectives and had run into the combinatorial explosion. British support for undirected AI research was largely withdrawn, and in the United States DARPA shifted toward mission-oriented funding, cutting speech understanding research in 1974.

“AI winter” is the usual label for this period, roughly 1974 to 1980. A caveat: at least one historian has disputed that it was a true winter, pointing out that membership of the main professional interest group kept growing. As with many historical labels, the story is cleaner than the record.

How did expert systems revive the field?

The next wave replaced grand ambitions with commercial narrowness. Expert systems captured the knowledge of human specialists as thousands of if-then rules. In 1980, a system called R1, built at Carnegie Mellon for Digital Equipment Corporation, was completed to configure computer orders. Corporations took notice.

In 1981, Japan’s trade ministry set aside 850 million dollars for the Fifth Generation Computer project, which alarmed Western governments into spending more of their own. The AI industry grew from a few million dollars in 1980 to billions by 1988, and companies built specialized machines for running the Lisp language.

Here the history rhymes with the earlier cycle. Expert systems were brittle and could not learn. Maintaining a rule base of thousands of hand-written rules was costly, and the systems fell apart on cases their authors had not foreseen.

Why did the second winter hit?

In 1987, the market for specialized Lisp machines collapsed when cheaper desktop workstations matched their performance, and an industry worth around half a billion dollars was largely replaced within about a year. DARPA cancelled new spending under its Strategic Computing Initiative in 1988. By the early 1990s, expert systems had a reputation as costly and fragile, and the Fifth Generation project ended in 1992 without meeting its goals.

Notably, the work that mattered most was continuing quietly. Backpropagation, popularized in a 1986 paper, was being used to train neural networks, and in 1989 the LeNet network learned to read handwritten zip codes. Statistical machine learning methods were maturing and being applied to practical problems.

What were the milestones of the 1990s and 2010s?

Public attention returned through games. In May 1997, IBM’s Deep Blue beat world chess champion Garry Kasparov in a six-game match, 3½ to 2½, after losing to him 4 to 2 in 1996. Deep Blue relied on special-purpose chess hardware searching about 200 million positions per second. It was a triumph of brute-force search and human-written evaluation, not of learning, and it did not generalize to anything but chess.

The real hinge came in 2012. A neural network called AlexNet, built by Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton at the University of Toronto, won the ImageNet image-recognition contest with a top-5 error rate of 15.3 percent, more than 10 percentage points better than the runner-up. It was trained on two consumer graphics cards. Yann LeCun called it an unequivocal turning point in the history of computer vision.

Why then? A common explanation is that three ingredients finally converged: much larger labeled datasets, graphics processors suited to the arithmetic, and training know-how built up over the previous decades. That explanation is widely accepted, though people disagree about how much weight to give each ingredient.

Then Go fell. In March 2016 in Seoul, DeepMind’s AlphaGo beat Lee Sedol 4 to 1. Go had long been considered harder for computers than chess, and before 2015 the best programs only reached strong amateur level. Unlike Deep Blue, AlphaGo learned much of its skill from data and self-play.

What are transformers and why did they change everything?

On 12 June 2017, a paper titled “Attention Is All You Need” introduced the transformer architecture. Its authors worked at Google. The key move was to drop the step-by-step processing that earlier language models used and let every word in a passage weigh its relevance to every other word at once. That made training easy to run in parallel on large clusters of chips, which meant models could be scaled up dramatically.

BERT, an encoder-only transformer, followed in 2018, and Google began using it for search queries in October 2019. OpenAI’s GPT series of decoder-only transformers also began in 2018. Late in 2022, ChatGPT, built on a fine-tuned GPT-3.5, reached the public, and the current boom in large language models took off. The 2020s have brought rapid scaling, many competing models and a surge of investment, and also a surge of argument about safety, copyright, energy use and law.

Is the story over, or just another cycle?

Honest answer: nobody knows, and be skeptical of anyone who does. The optimistic reading is that this time is different because the systems are actually useful and widely used, not just promising demos. The cautious reading points to the history above and notes that funding enthusiasm and technical reality have diverged before.

What can be said is that the pattern of the past seventy years is a ratchet with setbacks. Each wave left behind lasting tools, whether theorem provers, rule engines, statistical classifiers or deep networks, even when the grander claims collapsed. Whether current methods reach something like general human-level intelligence is an open scientific question, and it is a hypothesis, not a result.

Common questions

Who invented artificial intelligence?

No single person did. Turing laid the theoretical groundwork and posed the question in 1950, and John McCarthy introduced the term “artificial intelligence” in the proposal for the 1956 Dartmouth workshop. Many others built the early programs and ideas.

Did Turing’s test ever get passed?

There is no universally accepted answer. The test as Turing described it is informal, and different experiments have used different rules. Modern chatbots can hold convincing conversations, but researchers disagree on whether that counts as passing in any meaningful sense.

What is an AI winter?

A period of reduced funding and interest after hopes outran results. Two are usually named, one in the mid-1970s and one from the late 1980s into the 1990s. Some historians argue the idea is overstated, since work continued in both periods.

Is ChatGPT the first real AI?

No. AI as a research field is about seven decades old, and earlier systems such as chess programs, spam filters and speech recognizers used AI methods. ChatGPT was notable because it put a general-purpose language system in front of the public, not because AI began there.

What made deep learning take off around 2012?

Large datasets, fast graphics processors and refined training methods arrived together. AlexNet’s win at ImageNet that year became the visible turning point.

What to remember

AI has a documented founding in 1956, two well-recorded periods of collapsing enthusiasm, and a rapid rise since 2012 driven by neural networks, data and graphics hardware. The transformer, introduced in 2017, is the architecture behind today’s large language models.

The recurring lesson is about forecasting. Predictions of imminent human-level machines have been wrong before, and so have predictions that certain tasks, such as beating the best humans at Go, would never be done. Treat both extremes with caution.

Further reading

A well-known history of the field written by someone who knew many of its founders. See it on Amazon (affiliate link)

As an Amazon Associate I earn from qualifying purchases. A commission may be earned if you buy through this link, at no extra cost to you.

Sources