The Turing Test at 76: did any AI pass it?
The claim
“An AI has passed the Turing Test.”
Verdict: Partly true
Partly true. In a 2025 study, GPT-4.5 was judged to be the human 73% of the time in five-minute, three-way chats, more often than the real people it was up against. The authors call it the first evidence that an AI passes a standard three-party Turing test. But the result depended on a prompt telling the model which person to play, the paper is a preprint, and Alan Turing's 1950 paper offered a prediction, not a rulebook.
Where does the Turing Test come from?
Turing opened his October 1950 paper in the journal Mind with a line that is still quoted: "I propose to consider the question, 'Can machines think?'" Then he replaced it with a game. Later in the paper he goes further and calls the original question one he believes "too meaningless to deserve discussion."
His imitation game starts with three people: a man, a woman and an interrogator who can't see them and has to work out which is which. Answers are typed, so that tone of voice can't help. Then Turing asks what happens when a machine takes one of the places. He mostly calls it a game; the word "test" shows up later, when he answers objections to the idea.
He also made a forecast. In about fifty years, he wrote, computers would play the game so well that an average interrogator would have no more than a 70% chance of making the right identification after five minutes of questioning. That is a prediction about the year 2000, not a pass mark, and the paper never names one.
What did the 2025 study do?
Cameron R. Jones and Benjamin K. Bergen posted the study on arXiv on March 31, 2025. They ran two randomized, pre-registered experiments: one with 126 undergraduate students and one with 158 people recruited online through the Prolific platform. In each game, an interrogator chatted for up to five minutes with two witnesses, one human and one AI, then decided which was the human. After exclusions, 1,023 games were analyzed.
Some AI witnesses got a "persona" prompt that told them to act like a young, introverted person who knows internet culture and uses slang. Others got only minimal instructions to convince the interrogator they were human.
How often each AI was judged to be the human
So did an AI pass?
By the standard the authors used, yes. GPT-4.5 with the persona prompt was chosen as human significantly more often than the actual humans. LLaMa with the same prompt was chosen about as often as the humans; the difference was not statistically significant. Turing's forecast was met by a wide margin: against GPT-4.5 with the persona, interrogators picked correctly in only 27% of games, far under his 70% ceiling.
Three things keep the verdict at "partly."
The prompt did much of the work. Without the persona, the same two models were judged human 36% and 38% of the time, and the authors say they did not robustly pass. What passed was a model plus careful instructions about who to be.
Five minutes is short. The authors note that longer conversations could be more demanding, and that the test does not mirror longer situations where someone might be fooled for real.
It measures seeming human, not intelligence. The authors present the result as one piece of evidence about these systems, to be read alongside others. The arXiv page lists a single version and no journal publication.
What gave the AI away?
Most interrogators made small talk. In 61% of games they asked about daily activities or personal details, and in half they probed opinions, emotions or humor. Only 19% asked the witness directly whether it was human, and 12% tested knowledge or reasoning.
The strategies that worked best were rare: saying something strange, or trying a "jailbreak" to make the model drop its role. Asking about the weather or about human experiences worked worst.
The oldest witness in the room
ELIZA, a pattern-matching program from 1966, was judged human in 23% of games, slightly more than GPT-4o with minimal instructions. Both were included as baselines and both scored well below chance. ELIZA has its own piece in our 3D museum of AI history.
The question Turing set aside never went away. Whether passing his game says anything about thinking is still argued in the philosophy of AI.
Ask your AI
Ask your AI to play Turing's imitation game with you for five minutes, then ask it what in the conversation would have given it away. Compare its answer with the strategies that actually worked for the judges in the 2025 study.
FAQ
Who invented the Turing Test?
Alan Turing described it as the imitation game in his paper "Computing Machinery and Intelligence," published in the journal Mind in October 1950.
How long does a Turing Test conversation last?
Turing's 1950 prediction referred to five minutes of questioning, and the 2025 study by Jones and Bergen used games of up to five minutes.
Did ChatGPT pass the Turing Test?
In the 2025 study, GPT-4.5 with a persona prompt was judged human 73% of the time. GPT-4o with minimal instructions was judged human 21% of the time, below chance.
Sources
Cite this
The Turing Test at 76: did any AI pass it?, New Mind Museum, November 17, 2026. https://newmindmuseum.com/philosophy/2026/11/turing-test-did-ai-pass/
New Mind Museum. (2026, November 17). The Turing Test at 76: did any AI pass it?. https://newmindmuseum.com/philosophy/2026/11/turing-test-did-ai-pass/
“The Turing Test at 76: did any AI pass it?.” New Mind Museum, 17 Nov. 2026, https://newmindmuseum.com/philosophy/2026/11/turing-test-did-ai-pass/.
@misc{new-mind-museum-2026-the-turing-test-at-76-did-any-ai-pass-it,
author = {{New Mind Museum}},
title = {The Turing Test at 76: did any AI pass it?},
year = {2026},
url = {https://newmindmuseum.com/philosophy/2026/11/turing-test-did-ai-pass/}
}Free to quote and cite with a link to this page.