A date at the top of the first page
Under the title of "On Computable Numbers, with an Application to the Entscheidungsproblem" there are two dates in square brackets: received 28 May 1936, read 12 November 1936.1 The second is the day Alan Turing's paper was read to the London Mathematical Society, which published it in its Proceedings, volume 42. Turing was 24 and was not even in London at the time: in September he had sailed for Princeton University, in the United States.2
Ninety years on, the paper is remembered for one idea, the universal machine, and in September the London Mathematical Society and the National Museum of Computing held a three-day conference on it at Bletchley Park.3 It is often confused, though, with another piece of Turing's work that came out fourteen years later, the one with the test that bears his name. They are two different questions and it pays to keep them apart, not least because the second one is back in the news.
A clerk, a sheet of squared paper, a single machine
In 1936 the word "computer" meant a person, and that is exactly where Turing starts: "Computing is normally done by writing certain symbols on paper."4 He pictures someone working on a sheet divided into squares, looking at a few symbols at a time and deciding the next step according to what they see and their "state of mind". Then he strips it to the bone: a tape divided into squares, a head that reads and writes one symbol at a time, a finite table of instructions.
The aim was to answer a problem set by David Hilbert's school: is there a mechanical procedure that can decide, for any statement of logic, whether it can be proved? Turing's answer is no: "the Hilbertian Entscheidungsproblem can have no solution".4 In the spring he had found out that Alonzo Church, at Princeton, had got there first by another method, and in an appendix added on 28 August 1936 he showed that the two routes led to the same place.2
The part that changed the world is section 6, which opens like this: "It is possible to invent a single machine which can be used to compute any computable sequence."4 A single machine that, given a description of another machine on its tape, behaves like that machine. Instructions become data, like the numbers they work on. It is why the same phone today works as a calculator, a camera and someone to talk to, just by changing the program.
In 1936, however, it was all on paper. The first program held in an electronic memory ran at the University of Manchester on 21 June 1948, looking for the highest factor of a number.5 Four years later a UNIVAC called Eisenhower's victory live on air. Since 2021 a table from the 1936 paper has been printed on the Bank of England's £50 note, the one with Turing's portrait.6
From the paper machine to the chat
- May 1936The London Mathematical Society receives Turing's paper
- Nov 1936The paper is read to the Society on 12 November
- Jun 1948In Manchester the first program runs from an electronic memory
- Oct 1950Mind publishes the paper with the imitation game
- Jan 1952On the BBC Turing speaks of at least a hundred years for a test with no questions barred
- Jun 2014Eugene Goostman convinces one judge in three in five minutes
- May 2026In PNAS GPT-4.5 is picked as the human in 73% of games
- Oct 2026With no time limit 36 interrogators out of 37 spot GPT-4-Turbo
The test arrives fourteen years later
The 1936 paper does not contain a single line about intelligence or conversation. The question "Can machines think?" opens another text, "Computing Machinery and Intelligence", published in October 1950 by the philosophy journal Mind.7 Turing immediately judges it "too meaningless to deserve discussion" and proposes a game in its place. An interrogator writes to two hidden players and has to work out who is who. In the first version they are a man and a woman; then a machine takes the place of one of them.
The most quoted sentence of that paper is a prediction. Turing writes that he believes that "in about fifty years' time", so around 2000, it will be possible to program computers to play so well that an average interrogator will have no more than a 70% chance of identifying them after five minutes of questioning.7 That is where the five minutes come from, and the threshold that almost every later trial adopted. But in January 1952, in a discussion on BBC radio, Turing himself said that a machine able to hold out with no questions barred would take at least a hundred years.8
What happened in 2025 and 2026
On 31 March 2025 Cameron Jones and Benjamin Bergen of the University of California San Diego released two three-party tests built like Turing's: an interrogator, a person and a machine, five minutes of chat, then the verdict.9 A total of 284 people took part, university students and volunteers recruited online, playing more than a thousand games. The study was published in PNAS on 19 May 2026, with an additional fifteen-minute replication.10
GPT-4.5, told to play a slightly introverted young person, into internet culture and given to slang, was picked as the real human in 73% of games, more often than the actual humans beside it. LLaMa-3.1 with the same persona reached 56%. Without those instructions GPT-4.5 fell to 36%.9
At the bottom of the table there is a surprise. ELIZA, the program Joseph Weizenbaum published in 1966, which simply turns the other person's sentences around according to a few rules,11 was taken for a human in 23% of cases, more than GPT-4o without a persona, which stopped at 21%. According to the authors, ELIZA convinced some interrogators precisely because it did not behave the way we expect an AI to: it was rude and sarcastic.9 Weizenbaum had already noticed how easily people trusted it: soon after starting to use it, his secretary asked him to leave the room so she could talk to the program in private.12
How often the machine was picked as "the human"
The way the interrogators decided explains a lot. In 61% of games they made small talk, in 50% they probed emotional and social qualities, and only in 12% did they test knowledge and reasoning. The most frequent reasons concerned writing style and the flow of the conversation, and many were plain gut feelings.9 Bergen says it outright: the test increasingly measures humanlikeness. Jones describes it as a game in which, for the models, the task is to lie.10
It is not the first time a persona has done most of the work. On 7 June 2014 a program called Eugene Goostman convinced about one judge in three by posing as a 13-year-old Ukrainian boy with schoolboy English, and critics pointed out that his age excused every odd answer.13
The San Diego result also has a counterweight. Diane Proudfoot, Sharon Temtsin and Christoph Bartneck of the University of Canterbury in New Zealand tried to run the game the way Turing described it, with no time limit. Their study came out in Philosophy of Science on 14 October 2026: 36 interrogators out of 37 identified GPT-4-Turbo, 33 needed more than five minutes, and games lasted 14 minutes on average.14 The authors warn that they tested a single model, older than GPT-4.5, with a single prompt. In September two Finnish researchers reported instead that ChatGPT 5.2 passed the test in Finnish, a language with little presence in training data, and that many participants were guided by colloquial Finnish.15
What I think
I think the imitation game, as we play it, mostly measures us: what we expect from a person and what we expect from a machine. In five minutes of pleasantries we judge by style, by rhythm, by an awkward joke at the right moment. GPT-4.5 wins when someone tells it who to pretend to be, ELIZA beats GPT-4o because it is rude. Neither result says much about the ability to understand or to reason.
This is not a criticism of Turing. He wanted to replace an impossible question with an experiment, and in 1952 he was the first to think it would take a long time. We are the ones who turned a 1950 prediction into a finishing line, and then were amazed to have crossed it. On another point Turing saw clearly and early: in 1947 he said that what was wanted was "a machine that can learn from experience".16 That is exactly what today's machines do. The San Diego interrogators show the other half of the job.
Machines can learn from examples and help us explore questions. But every idea leaves a footprint, and the task of choosing the right questions remains ours.
Thirteenth law of innovation, from Un perché alla volta (translated)
In the study the strategies that worked best were saying something strange and trying to get the model to break out of its instructions, but few people used them.9 The right questions were there, and almost nobody asked them.
Of Turing's two ideas, the 1936 one is still, for me, the more important. The universal machine explains why a single object in your pocket can do almost anything, and the same paper proved that there are problems no machine of that kind can solve, whatever the program. Turing described the computer and its limits in the same thirty-six pages. The 1950 test was a way of reasoning about thought, and taken literally today it mostly tells us how easy we are to convince.
What changes for people with children, students or customers
The practical consequence is simple: "seems like a person" is no longer a clue. The profile messaging a teenager in an online game, the customer service chat, the new contact on social media may all be a model with a well-written persona. The watermarks I wrote about in the piece on SynthID help with images, video and text already produced, but in a live conversation there is no file to check. What remains are habits, and the studies say which ones work.
- TimeDo not decide in five minutesIn the New Zealand study 33 interrogators out of 37 needed more than five minutes, and in the end the machine was almost always identified.
- QuestionsGet past the pleasantriesIn San Diego the most effective strategies were saying something strange and asking the other side to drop the role. Small talk helps the machine.
- ContextAsk what only someone who is there would knowA local, shared fact: what happened in class this morning, who was at the meeting. A model has to make it up.
- ChannelCheck elsewhereA phone call, a meeting, a contact you already know. It is the same method that works for cloned voices on the phone.
With younger children I would put it plainly: someone who writes well and answers at once is not necessarily a person, nor necessarily who they say they are. When in doubt, talk to an adult before replying. For teachers in Italy, the 1950 essay is available in Italian in Diego Marconi's translation for Einaudi, under the title "Macchine calcolatrici e intelligenza".17 The fifty-year prediction is there, with all its conditions: an average interrogator, five minutes, 70%. Reading it in class makes it possible to see what was passed in 2025 and what was not.