The AI timeline: from 1950 to today, including the times it did not work

If you only read headlines, AI sounds like something that appeared recently. It did not. The word is older than most of the people who use it today, the technology has come close to being written off more than once, and between the big breakthroughs lie decades in which almost nothing of what was promised happened.

This article goes through the timeline in order. It does not try to impress with the number of dates. It tries to answer a single question: how did we get here, and what can we learn from the times things went the other way. If you want to start with what artificial intelligence actually is, from the ground up, that article comes first in the area.

1950: the question before the word

Five years before the term artificial intelligence existed, a British mathematician set out the problem.^{2, 15} In October 1950 Alan Turing published an article in the journal Mind that opened with the question of whether machines can think.^{2, 15}

Turing then did something clever. He noted that the question is too vague to be answered, and replaced it with a game.² In his imitation game, a judge has to decide which of two hidden conversation partners is the woman and which is the man.² The question Turing asks is what happens when a machine takes the place of one of them. If the machine makes the judge guess wrong just as often, the question of whether the machine "thinks" is less interesting than what it actually achieves. Today the set-up is called the Turing test.

The point to take away is about method. Turing replaced a question that could not be measured with one that could be tested. That is still what separates work from hope in AI.

1955 and 1956: the word gets a date

As far as is known, the term artificial intelligence was first used in a research proposal dated 31 August 1955.^{1, 5} Behind it were John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon.¹ They applied for funding for a two-month study with ten participants, in the summer of 1956 at Dartmouth College.¹

What makes the proposal worth reading seventy years later is how open it is about its own uncertainty. The starting point is called a conjecture, plainly: that every aspect of learning, or any other feature of intelligence, can in principle be described so precisely that a machine can be made to simulate it.¹

A conjecture. Two months. Ten people. That is how the field began.

1966: the first setback came quickly

Ten years later the bill came for the first big promise. It was about machine translation, which during the Cold War had an obvious and well-funded use.

In 1964 an advisory committee was set up at the US National Academy of Sciences to assess the situation.³ The report came in 1966 under the title "Language and Machines".³ The conclusion was brief and uncomfortable: there was no machine translation of general scientific text, and none was in sight for the foreseeable future.³

The committee also put a price on the hope. Over about ten years the government had spent around 20 million dollars on machine translation and related subjects, more than a whole year of translation work cost the government to buy.³

It is an early version of a pattern that comes back: a concrete task turns out to be harder than it looked, and the funder does the sums.

1969 to 1976: the money comes with conditions

In the 1960s American AI research had been funded generously and with few demands. That changed.

Conditions tighten in the US and Britain

In 1969 the Mansfield Amendment was passed, which banned any of the defence budget from being used for research without a direct and apparent relationship to a specific military function.¹⁷ For the research funder ARPA it meant mission-driven research instead of free basic research. Researchers now had to show that the work would lead to something useful, and fairly soon.

In 1973 the mathematician James Lighthill delivered a review of British AI research.^{7, 18} His verdict was that in no part of the field had the discoveries produced the major impact that was once promised.⁷ The report is usually given as the reason British support was withdrawn. The historian who recalculated the figures writes that it is unclear both how and whether the review triggered any wider funding freeze.^{4, 18} In practice the work continued at a few places. The two big groups were in Sussex and Edinburgh.¹⁸

The programme that met its goals

A year earlier, in 1971, DARPA had started a five-year programme for speech understanding worth 15 million dollars in total.^{19, 20} The best system met the goals when the programme was demonstrated in September 1976: it understood whole sentences from a vocabulary of a thousand words.^{19, 21} But the goals had been set so that the sentences had to follow a constructed grammar, and when the programme ended it was not renewed.¹⁹

The difference between meeting a goal and getting a product is the whole problem of AI history in a nutshell.

On the winter: the story is disputed

Here we need to be careful, because there is a standard story and it is not undisputed.

The standard story and the objection

The standard story says that the years 1974 to 1980 were AI's first winter, a period of collapsed interest and withdrawn money.⁴ That version is in most of the summaries you find online.

The historian Thomas Haigh has argued in Communications of the ACM that the first winter did not happen.⁴ His objection can be measured. When the Lighthill report was published in 1973, the ACM's AI group SIGART had 1,241 members, about twice as many as in 1969.⁴ During what is usually described as the darkest part of the winter, the group almost tripled: by mid-1978 it had 3,500 members.⁴ It grew faster than the ACM as a whole.⁴

What the figures say and do not say

So it was not the field that shrank. It was the largest laboratories and the most ambitious projects that were turned down. The number of people working on AI kept rising.

We do not take a position on what the period should be called. We note that the picture is disputed, and that anyone who copies the usual version without checking risks repeating something that is not true. That in itself is a lesson about how the history of technology is written.

1984: someone says it out loud

There is, however, a period that almost everyone agrees on, and the interesting thing is that it was predicted in public.

In August 1984 a panel debate was held at the AAAI annual conference in Austin, Texas.^{6, 22} The title was "The Dark Ages of AI". According to accounts at the time, the conference itself felt more like a trade fair: expert system companies were springing up everywhere, large companies were rushing to set up AI departments, and the money was flowing.^{5, 6}

Drew McDermott opened the panel by talking about a deep unease that overly high expectations would end in disaster.²² He sketched what he himself called an unlikely worst case: all the start-ups fail, the companies lose interest, it becomes impossible to get money for anything to do with AI, and everyone hastily renames their research projects as something else.²²

Roger Schank turned it around on the same occasion. He said he was worried not that expectations were too high but that they were too low: the AI community had forgotten that it was there to do science, and it was nowhere near any solution.^{6, 22}

It was 1984. The field warned itself, on stage, in the middle of the boom.

1987: the warning comes true

The commercial AI of the 1980s was called expert systems: programs that captured an expert's knowledge in sets of rules. The best known, XCON, was built at Carnegie Mellon for a computer maker, which said that together with a related system it saved more than 40 million dollars a year.⁵ It was a strong sales story, and it pulled a whole industry along with it.

From 1987 the market for the specialised hardware that AI programs ran on gave way, and over the following years most of the makers disappeared or lost their lead.²³ The reason was not that AI had failed in any deep sense. Ordinary workstations had become just as powerful and much cheaper, driven by a new generation of processors.²³

The expert systems themselves turned out to have a more basic problem. Systems meant to automate expert knowledge required companies to hire new experts to maintain them.⁵ In 1989 the computer maker mentioned had 59 technicians assigned to look after the infrastructure and rule base of its internal expert systems.⁵ Few companies could afford that investment.

The downturn that followed lasted into the 2000s.⁵ The sources do not agree on the year it ended, and we do not claim a year.

1997, 2012, 2016: three dates that changed the picture

May 1997. The reigning world chess champion lost a match to a computer. The deciding sixth game was played in New York on 11 May; the champion resigned after 19 moves and the match ended 3.5 to 2.5.^{24, 25} It was the first time a reigning world champion was beaten by a machine in a match played under standard tournament time controls.²⁵

2012. A neural network won the ILSVRC image recognition competition with a top-5 error of 15.3%.⁹ The second-best entry was at 26.2%.⁹ The model had 60 million parameters, was trained on 1.2 million images and was built on an efficient implementation for graphics processors.⁹

That margin is the turning point of the whole timeline. The gap to the runner-up was too large to be dismissed, and the method behind it was not a new theory but more computing power and more data applied to an old idea.

March 2016. In Seoul a five-game match of the board game Go was played between one of the world's top players and a computer program.²⁶ The program won 4 to 1.²⁶ The human took the fourth game.

2017 onwards: the architecture behind what you use today

In June 2017 an article titled "Attention Is All You Need" was published.¹⁰ It proposed an architecture built entirely on attention mechanisms, without the recurrence that had been standard until then.¹⁰ The decisive property was practical: it could be parallelised much more and therefore needed much less training time.¹⁰

Almost all the text tools the public began using from 2022 onwards are built on that architecture.

That the technology also delivers results outside competitions and chat windows became clear in 2024, when half of the Nobel Prize in Chemistry went to David Baker and the other half jointly to Demis Hassabis and John Jumper, the latter for predicting the structure of proteins with AlphaFold2.¹¹

2024 to 2026: the rules catch up

The latest part of the timeline is not about models but about laws.

The EU AI Act entered into force on 1 August 2024 and applies in stages.^{12, 27, 28} The bans on certain uses apply from 2 February 2025, the rules for general-purpose AI models from 2 August 2025, and the general date of application was 2 August 2026.^{27, 28} From that date the transparency rules apply: you must be told when you are interacting with an AI system, and AI-generated content must be labelled.¹³ The labelling requirement was, however, given a grace period until December 2026 for services already on the market.^{12, 13}

The same summer something happened that says more about the maturity of the technology than any model release. On 8 July a simplification act was signed, published on 24 July and in force from 27 July 2026.^{12, 28} It pushed back the heaviest requirements, those for high-risk AI systems, to 2 December 2027 and 2 August 2028 respectively.¹² The reason given in the act is that the technical standards and national supervision were not ready in time.¹²

In Sweden, on 4 June 2026, the government gave five agencies the task of acting as national competent authorities: the Swedish Post and Telecom Authority, the Swedish Authority for Privacy Protection, the Swedish Financial Supervisory Authority, the Swedish Medical Products Agency and Swedac.^{14, 29} The task runs until 31 December 2026, pending supplementary Swedish legislation.^{14, 29}

What the timeline is good for

Seventy years of history give three things worth taking away.

The word is older than the technology. The term turned seventy in 2025.¹ What is new is not the idea but that it works well enough to build on.

The field has promised too much before, and has known it. The 1984 warning came from inside, from the stage, while the money was still flowing.²² Anyone who wants to understand technology cycles has more to learn from 1984 than from any product launch.

The breakthroughs usually came when capacity caught up with an old idea. But capacity was not the whole answer. As early as 1955 the proposal wrote that the main obstacle was not the capacity of the machines but the inability to write programs that made use of it, and in 1966 the committee explained the failure of machine translation by a semantic barrier and not by computers being too slow.^{1, 3} That is why the timeline speeds up where it does.

The next article in this area moves on from when to how.

What to take away

The term artificial intelligence was first written down in a research proposal dated 31 August 1955.¹ The turning point is 2012: a top-5 error of 15.3% against the runner-up's 26.2%.⁹ The strength lies in the pattern rather than in single years. The breakthroughs came when computation and data caught up with old ideas, and capacity was never the whole answer: in 1955 the proposal pointed to the programs, and in 1966 the committee pointed to a semantic barrier.^{1, 3} The field has also warned itself in the middle of the boom, on stage in 1984.²² Follow three dated things at the Commission and the government: the labelling requirement's grace period in December 2026, the new high-risk dates of 2 December 2027 and 2 August 2028, and whether the five Swedish agencies' task, which runs until 31 December 2026, is replaced by Swedish legislation.^{12, 14}

Education, not advice.