Turing Test: Rules, Famous Attempts and the 2025 Result

In brief

The Turing test is Alan Turing's 1950 imitation game: a judge chats by text with a person and a machine and must say which is which. In a 2025 UC San Diego study, GPT-4.5 with a persona prompt was picked as the human 73% of the time, so one version has been passed, as of November 2026.

What is the Turing test?

The Turing test asks whether a machine can pass for a person in a written conversation. Alan Turing proposed it in 1950 in his paper "Computing Machinery and Intelligence," published in Mind, and called it the imitation game. Here is the test, the main attempts and the open question in one table.

Two matching wooden screens on a black plinth with a small rainbow glass speech bubble resting on a silver chess pawn, beside a typewriter holding blank paper, in a charcoal studio under white rim light.
DateSystem or eventWhat happenedSource
1950Turing's paper in MindProposes the imitation game in place of the question whether machines can thinkStanford Encyclopedia of Philosophy
1966ELIZAWeizenbaum's keyword program plays a Rogerian therapist and fools some peopleWikipedia
1972PARRYPsychiatrists picked the right transcripts about 52% of the time, which is chanceWikipedia
Nov 1991First Loebner Prize contestA yearly chatbot contest that ran until 2019 and never awarded its gold or silver prizeWikipedia
7 Jun 2014Eugene Goostman33% of 30 judges at the Royal Society thought it was humanWikipedia
31 Mar 2025GPT-4.5 with a persona promptPicked as the human in 73% of five minute chatsarXiv 2503.23674

Stanford HAI states the pass rule in one sentence: if the evaluator can't reliably tell the two apart, the machine passes. The rest of this page covers how the game worked, who tried it and whether the 2025 result settles anything. For where the test sits in the longer story, see when was AI invented.

How does the Turing test work?

The original game has three players and a written channel. Wikipedia describes it as a party game: an interrogator, called C, questions a man and a woman in another room through written notes and tries to say which is which. The man tries to deceive C and the woman tries to help.

Then Turing swaps a player. The Stanford Encyclopedia of Philosophy quotes him asking, "What will happen when a machine takes the part of A in this game?" The machine now plays the deceiver, and the judge has to pick the computer.

Most pages teach a simpler form. A human judge reads a text transcript of a chat with a human and a machine, and the machine passes if the judge can't reliably tell them apart. The three party form is harder for the machine, because the judge can compare the two answers to each question.

Turing also floated a jury version. Wikipedia says he proposed a 1952 BBC broadcast with a panel of judges. Scholars label the forms differently: the original imitation game compares a man and a woman, and the standard Turing test compares a person and a machine. Turing never said whether the judge should know a machine is involved.

Three rules decide the result. The channel is text only, so voice and looks are out. The conversation is timed, and Turing named five minutes. The judge knows one of the two is a machine, which Wikipedia lists as a design question of its own.

Has AI passed the Turing test?

Yes for one version, no for the strictest reading, and the honest answer depends on the study. Cameron Jones and Benjamin Bergen of UC San Diego posted the result on arXiv on 31 March 2025 under the title Large Language Models Pass the Turing Test. They ran two pre registered, randomized, controlled tests of a three party game and compared four systems.

GPT-4.5 with a persona was picked as the human 73% of the time
GPT-4.5 with persona73 %LLaMa-3.1 with persona56 %ELIZA23 %GPT-4o no persona21 %Chance line50 %

Fifty percent is chance. GPT-4.5 with a persona prompt scored 73%, so judges chose it as the human more often than they chose the real person beside it. LLaMa-3.1-405B with a persona scored 56%, which the authors report is not significantly different from the humans it faced. ELIZA scored 23% and GPT-4o with no persona scored 21%, both significantly below chance.

The scale was real. The paper reports 1,023 games across 126 UC San Diego undergraduates and 158 Prolific workers, with a median of 8 messages per game and an average length of 4.2 minutes inside a five minute limit. Without the persona prompt, GPT-4.5 scored 36% and LLaMa-3.1 scored 38%, both under chance.

Two years earlier the same pair tested GPT-4 in a two player game. Their 2023 paper, Does GPT-4 pass the Turing test?, reports that the best GPT-4 prompt passed 49.7% of games, ELIZA 22% and human participants 66%. The 2025 jump came with a newer model and a three party game.

The authors don't say the models think. They describe the result as a test of humanlikeness, not a direct test of intelligence, and they say social and emotional style drove many verdicts. Treat the paper as a preprint: I could not confirm a journal version for this page.

What did Turing actually predict?

Turing predicted that in about fifty years an average interrogator would have no more than a 70% chance of a right identification after five minutes of questioning. The Stanford Encyclopedia of Philosophy quotes the passage. He was forecasting, and he wasn't setting a pass mark.

The arithmetic is simple. If the judge is right 70% of the time at most, the machine fools the judge at least 30% of the time. That is where the 30% figure in headlines comes from. Fifty years from 1950 is 2000, and the 2014 Goostman claim, the first widely reported pass, came 14 years after that date.

He added a second forecast about talk, not about machines. By the end of the century, he wrote, general educated opinion would have changed so much that people could speak of machines thinking without being contradicted. That one is harder to score.

Which programs tried the Turing test before 2025?

Four efforts before 2022 matter, and each shows a different weakness of the test.

  1. ELIZA, 1966. Joseph Weizenbaum's program reflected a user's words back as a Rogerian therapist would. It did no understanding, and Wikipedia says it fooled some people. The 2025 study kept ELIZA as a baseline and it scored 23%.
  2. PARRY, 1972. Kenneth Colby's program simulated a person with paranoid schizophrenia. Psychiatrists picked the right transcripts about 52% of the time, which is chance, according to Wikipedia.
  3. The Loebner Prize, 1991 to 2019. Wikipedia says the first contest was in November 1991, the silver and gold prizes were never won, and the last contest was in 2019. The Stanford Encyclopedia says entries were caught by not very subtle questioning.
  4. Eugene Goostman, 2014. The program posed as a 13 year old boy from Odesa with limited English. Organizer Kevin Warwick of the University of Reading ran the event at the Royal Society on 7 June 2014, and 33% of 30 judges thought it was human.

A fifth episode shows the opposite problem. In June 2022 Google engineer Blake Lemoine claimed its LaMDA chatbot was sentient, and Google dismissed the claim, according to Wikipedia. Sounding human in chat isn't evidence of a mind.

Critics weren't convinced. Wikipedia reports that Gary Marcus called Goostman a cleverly coded piece of software, and Murray Shanahan of Imperial College London said the pass claim devalued real AI research. The Stanford Encyclopedia adds that the trial was small and its result not reliably projectible. A 13 year old with limited English lowers the bar for odd answers.

What do three real cases show about the Turing test?

Three cases show what a pass can and can't mean.

Case 1, ELIZA in 1966. A program with no model of the world got people to confide in it, because humans fill gaps with meaning. That habit is called the ELIZA effect, and the 2025 authors say it partly explains some verdicts. A pass can be a fact about the judge.

Case 2, Eugene Goostman in 2014. The result was 33% of 30 judges, one trial, a persona built to excuse errors. The headline said a machine passed. The data said ten people out of thirty were fooled in short chats.

Case 3, GPT-4.5 in 2025. The same model scored 73% with a persona prompt and 36% without it. The pass depended on the persona instruction. That is a stronger result than 2014 and still not a general one. For why models also state falsehoods with confidence, see what is an AI hallucination.

Why do critics say passing the Turing test proves little?

Critics say the test measures imitation, not understanding. The best known objection is John Searle's Chinese Room, from his 1980 paper Minds, Brains, and Programs. Wikipedia summarizes it: a program that swaps symbols by rule, as ELIZA does, can produce fluent replies without grasping any of them.

The Stanford Encyclopedia lists more. One is chauvinism: the test only recognizes intelligence that talks like us. Another is the Blockhead idea from Ned Block, that a giant look up table could pass with no thought at all. A third is that a short test could be passed by luck, which longer and repeated tests make unlikely.

Wikipedia also lists weaknesses of its own: naive judges, a language only focus, and the case of silence, where a machine says nothing and so can't be caught out. The Jones and Bergen paper agrees on the central point. They describe their result as humanlikeness. A pass means the judges couldn't tell, and it leaves the question of mind open. AGI vs AI covers the claim most people actually care about.

Is CAPTCHA a Turing test?

Yes, in reverse. A CAPTCHA asks a computer to judge whether the visitor is a person, so the roles flip. Wikipedia calls this a reverse Turing test and says CAPTCHAs are based on the Turing test.

It also shows how fragile the idea is. Wikipedia reports that Google engineers built a solver for distorted text puzzles with 99.8% accuracy in 2014, which pushed sites toward invisible checks such as reCAPTCHA v3.

What does the Turing test do well?

Wikipedia lists three strengths. The test is simple, because it gives a measurable result where thinking has no agreed definition. It is broad, because a judge can raise almost any topic. It also rewards emotional and aesthetic skill, since Turing's own sample dialogues ask for poetry and feeling.

It has a social cost too. Economist Erik Brynjolfsson calls a race to imitate humans the Turing Trap, because a machine that copies workers can replace them instead of helping them, as Wikipedia summarizes his argument.

What tests replaced the Turing test?

No single test replaced it, but several narrower ones took its place in research. Wikipedia lists them, with dates where it gives one.

TestWho and whenWhat it checks
Total Turing testStevan HarnadAdds vision and physical action to the text chat
Winograd schema challengeLevesque, Davis and Morgenstern, 10 June 2012Resolving what a pronoun refers to in a sentence
Lovelace testNamed for Ada LovelaceWhether a machine originates something new
Ebert testRoger Ebert, 2011Whether a synthetic voice can make people laugh
Social Turing game, Human or NotAI21 Labs, 2023Played more than 10 million times by more than 2 million people; 32% could not tell humans from machines
Scientific discovery testDavid Eagleman, 2023Whether a system can do scientific discovery

The AI21 figures are Wikipedia's report of the company's game, not an independent study. For the live version of the game, turingtest.live lets you chat with a person and an AI and guess which is which. It is a game and states no results, and the site credits Cameron Jones, the study's first author.

What questions should you ask an AI about the Turing test?

Paste these into Claude, ChatGPT, Gemini or Whizi. Each asks for a source so you can check the answer.

  1. Rules: "Explain Turing's three player imitation game in six sentences, name the journal, volume and year of the paper, and tell me what changed when the machine took the place of player A."
  1. Prediction: "What exactly did Turing predict about a 70% chance of a right identification and five minutes? Give the page or a link and say whether the prediction came true."
  1. Study: "Summarize the 2025 Jones and Bergen paper Large Language Models Pass the Turing Test. List each model's win rate and say what the persona prompt did. Link the arXiv page."
  1. Critique: "Argue the Chinese Room against the 2025 result in one paragraph, then give the best reply to it, with a source for each side."
  1. Disagreement: "Where do experts disagree on whether AI has passed the Turing test? Give two views, the evidence for each and a link for both."
  1. Design: "Write a stricter Turing test that fixes three weaknesses of the five minute version. For each, name the weakness and the source that describes it."

What mistakes ruin an answer about the Turing test?

Four mistakes recur, and each has a fix.

  1. Saying a machine passed in 2014 without the numbers. The fix: say 33% of 30 judges, one trial, a 13 year old persona.
  2. Treating the 30% figure as Turing's pass mark. The fix: call it his forecast of what an average interrogator would score.
  3. Calling a pass proof of thinking. The fix: use the authors' own word, humanlikeness, and cite the Chinese Room as the standing objection.
  4. Trusting a model's quote of the 1950 paper. The fix: open the paper in Mind and find the sentence, because a chat model can produce a plausible line that Turing never wrote. How to fact check AI answers has the routine.

One app for putting the same question to three models

Whizi puts Claude, GPT and Gemini in one app. Pro carries all three, and its first week is $1.99. The free plan gives 10 messages a day on Whizi AI, Auto and Claude Haiku, with no card.

Side by side Compare on Pro runs one prompt through two models, which is a small version of the game: read both answers and guess which one invented a date. Web search is on by default in a chat, so you can ask for links. For the paper itself, arXiv and Mind are faster than any chat app, and the turingtest.live game is free.

Where does the Turing test fall short?

Its limits are in the rules, the judges and the sources. A five minute text chat checks style, not reasoning. The 2025 authors say judges leaned on social and emotional cues, and the persona prompt mattered. Judges also vary: UC San Diego undergraduates and Prolific workers gave different win rates for the same model, 69% against 76% for GPT-4.5 with a persona.

The sources have gaps too. Wikipedia's page is a secondary source and its own text dates the first Loebner contest to 1991 in one place and 1992 in another. The Stanford Encyclopedia entry predates the 2025 study. Read the paper, then the criticism, and hold both. What is an LLM explains why a model can sound human without a mind behind it.

Workflow checklist
  • Say which version: three party or a single chat
  • Quote the conversation length and the number of judges
  • Call the 30% figure Turing's forecast, not a pass mark
  • Check whether the machine needed a persona prompt
  • Cite the paper in Mind for anything Turing said
Common questions

Frequently asked questions

Has AI passed the Turing test?

One version, yes. In a 2025 UC San Diego study, GPT-4.5 with a persona prompt was picked as the human in 73% of five minute chats, above the 50% chance line. Without the persona it scored 36%, and the authors call the result humanlikeness, not proof of intelligence.

What is the Turing test in simple words?

A judge chats by text with a person and a machine and has to say which is which. If the judge can't reliably tell, the machine passes. Alan Turing proposed it in 1950 as the imitation game, to replace the question whether machines can think.

Who passed the Turing test first?

The Eugene Goostman chatbot is the first widely reported pass, on 7 June 2014, when 33% of 30 judges thought it was human. Critics said the trial was small and its 13 year old persona excused mistakes. In 2023 the same authors as the 2025 study found GPT-4 at best 49.7%, below the human 66%.

Is the Turing test still used?

Rarely as a goal. The Loebner Prize ended in 2019, and researchers use narrower tests such as the Winograd schema challenge, from 2012. The test survives as a public benchmark and a game, and the 2025 paper shows it can still be run properly.