Skip to content
Book a session
Menu
Back to Activities

Seventy years in one scroll.

AI from 1950 to ChatGPT, each date checked against its original document. Eight times along the way someone sounded sure. Call whether it held up before you see what the record shows.

1950 to 2022

  1. 1950October

    Turing swaps the question

    Alan Turing asks “Can machines think?”, calls the question “too meaningless to deserve discussion”, and swaps in a guessing game played through typed answers.

    In his game a questioner tries to tell a man from a woman by their typed replies. Turing then asks what happens when a machine plays the man. He never called it the Turing test. His 70 per cent figure was a prediction for about the year 2000, and he set no pass mark.

    Source: Turing, “Computing Machinery and Intelligence”, Mind, 1950

    “passed for the very first time”

    The University of Reading in June 2014, announcing that a chatbot had passed the Turing test by convincing 33% of its judges.

    Didn’t hold up.

    Reading’s rule was that a machine passes if it’s taken for a human more than 30% of the time. That rule isn’t in Turing’s paper. He set no pass mark at all, and the only percentage in it is a prediction.

    Sources: University of Reading press release, 8 June 2014; Turing, “Computing Machinery and Intelligence”, Mind, 1950

  2. 1955August

    The field gets its name

    Four researchers ask for money for a summer study at Dartmouth College, and their proposal calls the subject “artificial intelligence”.

    John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon dated it 31 August 1955. McCarthy later said the term was chosen “to nail the flag to the mast”. He half-remembered hearing it before, and a 1951 science-fiction story had already given a metal giant “an artificial intelligence equal to the human”.

    Sources: The Dartmouth proposal, 31 August 1955 (McCarthy’s transcription); McCarthy, “The Dartmouth Workshop—as planned and as it happened”, 2006; Startling Stories, January 1951, page 120

    “a significant advance can be made … if a carefully selected group of scientists work on it together for a summer”

    The Dartmouth proposal, August 1955, on what one summer could do.

    Didn’t hold up.

    McCarthy later wrote that “there was no real joint work” that summer, and no report was written. In 2006 he wrote that his hope for a breakthrough “was not realized”.

    Sources: McCarthy, “The Dartmouth Conference”, draft, 1987; McCarthy, “The Dartmouth Workshop—as planned and as it happened”, 2006

  3. 1958July

    A Navy machine “learns by doing”

    The New York Times reports a Navy-backed computer called the Perceptron, designed by the psychologist Frank Rosenblatt, that “learns by doing”.

    Rosenblatt’s own magazine article that summer said “we are now about to witness the birth of such a machine”. His journal paper that year was careful about what a perceptron couldn’t do.

    Sources: The New York Times, 8 July 1958; Rosenblatt, “The perceptron”, Psychological Review, 1958

    “walk, talk, see, write, reproduce itself and be conscious of its existence”

    What the Navy expected the Perceptron to do, in a wire story in The New York Times, 8 July 1958.

    Didn’t hold up.

    In 1961 Rosenblatt wrote that the press had covered it with “the exuberance and sense of discretion of a pack of happy bloodhounds”. He called the perceptron “first and foremost a brain model”.

    Source: Rosenblatt, Principles of Neurodynamics, preface, 1961

  4. 1966January

    ELIZA talks back

    Joseph Weizenbaum publishes ELIZA, a program at MIT that holds a typed conversation while playing a kind of therapist.

    It spotted key words in what you typed and rebuilt your own words into a reply. Weizenbaum wrote that some people were “very hard to convince” it wasn’t human. Ten years later he wrote that his secretary, who had watched him work on it for months, asked him to leave the room after a few exchanges with it.

    Sources: Weizenbaum, “ELIZA”, Communications of the ACM, 1966; Weizenbaum, Computer Power and Human Reason, introduction, 1976

  5. 1973

    A British review finds disappointment

    Sir James Lighthill reviews AI for Britain’s Science Research Council and finds that no part of the field has made the impact once promised.

    “In no part of the field”, he wrote, had the discoveries “produced the major impact that was then promised”. The council published his report in 1973 with replies giving other points of view. Later histories say British confidence in AI, the funding body’s included, collapsed for about a decade.

    Sources: Lighthill, “Artificial Intelligence: A General Survey”, 1973; Howe, “Artificial Intelligence at Edinburgh University”, 2007

  6. 1980January

    An expert system goes to work

    Digital Equipment Corporation starts using R1, also called XCON, a program built from rules that works out how to put customers’ computer orders together.

    It checked each order for missing parts and drew diagrams for the technicians. It grew from under 800 rules in 1980 to 3,303 by 1984. Its builders wrote that a system like it “will never have all the knowledge it needs”, so “it will always make mistakes”.

    Sources: McDermott, “R1: an Expert in the Computer Systems Domain”, 1980; Bachant and McDermott, “R1 Revisited”, AI Magazine, 1984

  7. 1980September

    Searle’s Chinese Room

    The philosopher John Searle argues that following a program, however convincing the answers, isn’t the same as understanding.

    He imagines himself locked in a room, matching Chinese symbols by rules written in English. To people outside, his answers can’t be told apart from a native speaker’s. Yet he doesn’t understand “a word of the Chinese stories”.

    Source: Searle, “Minds, brains, and programs”, 1980

  8. 1988March

    The AI business hits winter

    The New York Times reports “a major retrenchment” in the AI industry. Companies had bet on special AI computers that cost too much and didn’t fit the machines their customers already had.

    Source: The New York Times, 4 March 1988

    “expectations about AI are too high”

    Drew McDermott, warning a US AI conference in 1984. He said the crash he feared was being called the “AI Winter” by some.

    Held up.

    Four years later the Times reported the retrenchment. A co-founder of the AI company Teknowledge told it, “People believed their own hype.”

    Sources: “The Dark Ages of AI”, AAAI-84 panel, AI Magazine, 1985; The New York Times, 4 March 1988

  9. 1997May

    Deep Blue beats Kasparov

    IBM’s Deep Blue beats the world chess champion, Garry Kasparov, 3.5 to 2.5 over six games in New York.

    Kasparov had beaten an earlier version 4–2 the year before. IBM puts the win down to brute force: 200 million positions a second. John McCarthy, who named the field, later wrote that chess programs “do not know that they are chess programs”.

    Sources: IBM, “Deep Blue”; The New York Times, 12 May 1997; McCarthy, “The Dartmouth Workshop—as planned and as it happened”, 2006

  10. 2012October

    AlexNet wins at ImageNet

    A University of Toronto team’s neural network, a program that learns from examples, wins the ImageNet photo-recognition contest by a wide margin.

    ImageNet is millions of labelled photos, checked by people paid through Amazon Mechanical Turk. Allowed five guesses a photo, the Toronto network missed 15.3% of the test photos, and the next-best team missed 26.2%. It trained for five to six days on two graphics cards.

    Sources: Krizhevsky, Sutskever and Hinton, NIPS 2012; ImageNet, “ILSVRC2012 results”

    “a winning top-5 test error rate of 15.3%”

    The AlexNet paper, 2012, on the score that won the contest.

    Held up.

    The win holds, with a footnote. The 15.3% entry used extra training data from ImageNet’s Fall 2011 release. On the contest’s own data alone it scored 16.4%, still far ahead of the next team’s 26.2%.

    Source: ImageNet, “ILSVRC2012 results”

  11. “a feat previously thought to be at least a decade away”

    How DeepMind’s Nature paper, in January 2016, described beating a Go professional on a full-sized board.

    Didn’t hold up.

    It had already happened, against Fan Hui in October 2015. Then AlphaGo beat Lee Sedol 4–1. DeepMind’s chief executive wrote, “To everyone’s surprise, including ours”.

    Sources: Silver et al., Nature, January 2016; Hassabis, “What we learned in Seoul with AlphaGo”, 2016

  12. “We stress-tested Tay under a variety of conditions”

    Microsoft, on how it prepared Tay before the launch.

    Didn’t hold up.

    Within Tay’s first 24 hours, Microsoft said, “a coordinated attack by a subset of people exploited a vulnerability in Tay”. It called that “a critical oversight”, apologised and took Tay offline.

    Source: Microsoft, “Learning from Tay’s introduction”, 25 March 2016

  13. 2017June

    Attention Is All You Need

    Eight researchers, most of them at Google, describe the Transformer: a design for translation software built entirely on a mechanism called attention.

    Attention lets each word in a sentence take account of all the others. The authors reported better translations for a fraction of the training cost of earlier designs. OpenAI’s 2023 report on GPT-4 calls it “a Transformer-based model”.

    Sources: Vaswani et al., “Attention Is All You Need”, 2017; OpenAI, “GPT-4 Technical Report”, 2023

  14. 2022November

    ChatGPT opens to the public

    OpenAI releases ChatGPT as a free “research preview”, a chatbot it said could answer follow-up questions and admit its mistakes.

    OpenAI’s launch post warned that it “sometimes writes plausible-sounding but incorrect or nonsensical answers”.

    Source: OpenAI, “ChatGPT”, 30 November 2022 (archived copy)

    “responded that it had supplied ‘real’ authorities”

    What ChatGPT told a New York lawyer who asked “Is Varghese a real case”, as the court recorded it in 2023.

    Didn’t hold up.

    The court found “non-existent judicial opinions with fake quotes and citations”, and in June 2023 it fined two lawyers and their firm. The lawyer said he had thought ChatGPT was “like a super search engine”.

    Source: Mata v. Avianca, opinion and order on sanctions, 22 June 2023

Six of the eight didn’t hold up.

The two that held were a warning about hype, and a famous score that needed a footnote.

What it shows.

Six of the eight claims on this line didn’t hold up. Each one sounded sure when it was made. What gave it away was checking it later against what actually happened.

That’s the habit the library is built on. Before you rely on a claim about AI, find where it came from and check it against the source.

Every date and quote here comes from the original document, and each one is linked. Where a line rests on later histories rather than a record from the time, it says so.

More on the Where it came from shelf, and you can check an answer’s references yourself in Check this answer.

Want your team checking the claims they rely on?

Every session teaches from checked material like this, on your team’s own work.

Book a session