Turing predicted 70 percent by 2000. GPT-4.5 scored 73 in 2025.

A parlor game, played by teleprinter

In 2025, two cognitive scientists ran GPT-4.5 through a text-only imitation game and reported that interrogators picked it as the human 73 percent of the time, more often than they picked the real human in the same game.¹ The finding entered the peer-reviewed record the next year, and it entered the press right away: one headline called it "the first AI model to pass an authentic Turing test."²

The phrase comes from a paper almost none of those headlines link to. Alan Turing published "Computing Machinery and Intelligence" in the journal Mind in October 1950, and it does not open by asking whether machines can think. He calls that question "too meaningless to deserve discussion" and replaces it with a game.³

Turing borrowed the game from a parlor game about telling a man from a woman. In the original, a man and a woman sit in separate rooms and answer a third player's written questions. The man tries to pass as the woman, and the interrogator has to catch him. Turing's version keeps that shape and changes one part of the cast: "It is played with three people, a man (A), a woman (B), and an interrogator (C) who may be of either sex."³ Communication runs only by written message. Turing specifies a teleprinter, the typewriter-like machine that let two rooms exchange typed text, because the whole test depends on judging someone by what they write and nothing else: no face, no voice, no handwriting. He then asks what happens if a machine takes A's place: "We now ask the question, 'What will happen when a machine takes the part of A in this game?' ... These questions replace our original, 'Can machines think?'"³ The unanswerable question about thinking becomes an answerable one about behavior.

The one number in the paper

Turing makes exactly one quantitative prediction, in a section called "Contrary Views on the Main Question," and he states it in one sentence: "I believe that in about fifty years' time it will be possible, to programme computers, with a storage capacity of about 10^9, to make them play the imitation game so well that an average interrogator will not have more than 70 per cent chance of making the right identification after five minutes of questioning."³ Fifty years from 1950 is the year 2000. Ten to the ninth is a billion binary digits. Turing separately estimates the human brain holds somewhere between ten billion and a thousand trillion (10^10 to 10^15) of the same units, and says he'd be surprised if the game needed more than his billion-digit machine.³ The bet has three conditions: a machine holding a sliver of a brain's raw storage, an ordinary (not expert) questioner, and five minutes on the clock. Under those conditions, the interrogator still gets it right at best 70 times in 100, wrong at least three times in ten.

Turing frames all four parts as a claim about language, not a threshold for consciousness: what he's betting on is that by 2000, ordinary usage of the words "machine" and "think" will have shifted enough that calling a machine intelligent won't need defending.³ Passing the bet was never supposed to settle whether a machine thinks. That question, Turing had already set aside as unanswerable.

The objection Turing saw coming

Turing spends the rest of the paper working through nine objections to his own claim, and one of them still gets made about today's models. Ada Lovelace had written, a century earlier, that "The Analytical Engine has no pretensions to originate anything. It can do whatever we know how to order it to perform."³ A machine, on that reading, can only do what its programmer explicitly told it to. Turing's reply is one line: "Machines take me by surprise with great frequency."³ His point is that a program can produce behavior its author never anticipated, which is not the same as behavior its author never permitted, and the Lovelace objection runs the two together.

The persona prompt behind 73 percent

The 2025 finding is now the peer-reviewed version of the claim, published in the Proceedings of the National Academy of Sciences in 2026 under the title "Large language models pass a standard three-party Turing test."⁴ Cameron Jones and Benjamin Bergen ran the closest thing to Turing's own game that anyone has tried on a modern model: an interrogator holds two simultaneous five-minute text conversations, one with a human witness and one with an AI witness, then says which is which.¹ Across two pre-registered studies (126 undergraduates at UC San Diego and 158 paid workers recruited through Prolific), the researchers ran 1,021 games in total. GPT-4.5, prompted to adopt a persona, was picked as the human 73 percent of the time.¹ Llama-3.1-405B, given the same prompt, reached 56 percent. Two baselines, GPT-4o with no special prompting and the 1960s chat program ELIZA, both landed under a coin flip: 21 and 23 percent.¹

Turing's own bet has an arithmetic flip side. If an average interrogator would be right no more than 70 percent of the time, the machine only had to fool that interrogator at least 30 percent of the time to clear his bar.³ Read that way, GPT-4.5 clears it without any coaching: with no persona prompt at all, it was still picked as human in 36 percent of games, already past the 30 percent floor Turing's own number implies.¹ The 73 percent figure is a different achievement, and it depends on a script. To get there, the researchers gave GPT-4.5 a written persona telling it to answer as a young, introverted person who knows internet culture; strip that prompt out and the same model drops from 73 percent to 36.¹ Turing's forecast described a computer "with a storage capacity of about 10^9": one machine, unspecified, tested on its own terms. The 73 percent headline describes one commercial model, coached.

Its own authors call it a deception test

The three-party design is also new. In an earlier two-party version of the same test, run in 2024 by the same two researchers, an interrogator held one five-minute conversation with a single witness, human or machine, and judged which it was, with no side-by-side comparison.⁵ GPT-4 was judged human 54 percent of the time that year. Real humans were judged human 67 percent of the time. The machine lost. The researchers called it "the first robust empirical demonstration that any artificial system passes an interactive 2-player Turing test,"⁵ a narrower claim than the one that made headlines in 2025.

The Stanford Encyclopedia of Philosophy's entry on the test keeps the same two questions separate that Turing did: whether a machine can win the game, and whether winning would prove it thinks. Success, it notes, is at most a reason to suspect thought, never proof of it.⁶

That is not a new complaint about the test. Three decades earlier, Patrick Hayes and Kenneth Ford made a related argument: a program built to win Turing's game has to be built to deceive, not to think. In their words, "To pass the species test we must make not an artificial intelligence, but an artificial con artist."⁷ They called the setup circular, too: it claims to measure "human conversational competence," but the only definition of that competence on offer is the ability to pass the test.⁷ Turing's own answer to Lovelace, that a program can surprise its author, doesn't touch this objection. A script can produce a fluent, surprising conversation built from nothing more than its most probable next word, a mechanic covered in an earlier lesson, without any of it being evidence of what is happening behind the words.