The History of the Turing Test

By Victor Hoang · Updated Oct 22, 2026

In 1950 nobody agreed on what "think" meant. So how could anyone settle whether a machine thinks? Turing's answer was to stop defining the word and start running a procedure.

The Turing Test is Alan Turing's proposal, printed in the journal Mind in October 1950, to replace the question "Can machines think?" with a conversation game that a machine either wins or loses. The deeper idea is older: in 1637 René Descartes named language as the first of two things a machine could never do well enough to pass for a person. The throughline of the whole story is that every generation of machines has beaten the last generation's definition of "passing". The argument has shifted each time from whether the test is possible to what it actually measures.

Three quarters of a century later, in March 2025, researchers at the University of California San Diego reported that interrogators picked a language model as the human more often than they picked the real human. This article follows the test from its first sketch to that result. For the wider field around it, see The History of AI and the AI timeline.

Timeline at a Glance

The Question of the Talking Machine (1637-1949)

Philosophers draw a line at language

In Part V of the Discourse on the Method, published in 1637, Descartes considers a machine made to look and move like a man. He grants that such a machine could imitate our actions, and then gives "two most certain tests" for telling it from a real person, per John Veitch's English translation. The first is language. A machine might utter words, even cry out when touched in a particular place, but it could never arrange them variously "so as appositely to reply to what is said in its presence." The second test is that such machines would fail at some tasks because they act from the arrangement of their organs, not from knowledge. The Stanford Encyclopedia of Philosophy dates the Discourse to 1637.

Be careful with the lineage. Descartes drew a line a machine could never cross. He did not propose an experiment, and nobody then called it a test of intelligence. The Stanford Encyclopedia entry on the Turing Test notes that Copeland finds an anticipation of it in the 1668 writings of the Cartesian Cordemoy, and that David Abramson has presented archival evidence that Turing knew of Descartes' language test when he wrote his 1950 paper. So the link from Descartes to Turing is documented, even if it is not a direct quotation.

A second ancestor was an objection, not a test. In the paper Turing quotes a memoir on Babbage's Analytical Engine by Lady Lovelace, who wrote that the machine "has no pretensions to originate anything". Turing cites her memoir as dated 1842 and treats it as the sixth of his objections. And a third came from his own time. Professor Geoffrey Jefferson's Lister Oration of 1949, which Turing quotes, held that only when a machine could "write a sonnet or compose a concerto" because of thoughts and emotions felt, and not by the chance fall of symbols, would he accept that machine equals brain.

None of this was a test yet. It was a standing argument: machines can calculate, but they can't converse, originate or feel. What it needed was someone to turn "can't" into something you could measure. For the broader roots of that argument, see the origin of artificial intelligence.

For teams today: the oldest form of the test is still the most practical. Before judging any system, write down the specific thing you believe it can't do, and then try it.

The Paper Years (1950-1965)

Turing replaces a vague question with a game

The setting is a typed conversation between rooms. In "Computing Machinery and Intelligence," published in Mind in October 1950, volume LIX, pages 433-460, Turing opens by proposing to consider the question "Can machines think?" and then explains why he won't define "machine" and "think" by common usage, which would lead to something like a Gallup poll. Instead he describes a game. A man (A), a woman (B) and an interrogator (C) sit in separate rooms. The interrogator, who may be of either sex, has to work out which is the man and which the woman. In the version that matters here, a machine takes the part of A, and Turing asks whether the interrogator will decide wrongly as often as when the game is played between a man and a woman. To keep voices out of it, the answers should be typed, ideally over a teleprinter between the two rooms.

The sample exchange in the paper is the best picture of what he had in mind. The interrogator asks the machine to write a sonnet on the Forth Bridge. It answers that it never could write poetry. Asked to add 34957 to 70764, it pauses about 30 seconds and gives 105621, which is 100 short of the true sum, 105,721. A machine in this game does not have to be flawless. It has to be believable.

He also made a dated prediction. Turing wrote that in about fifty years it would be possible to program computers with a storage capacity of about 10 to the power 9 to play the game so well that an average interrogator would have no more than a 70 per cent chance of making the right identification after five minutes of questioning. He added that he considered the original question "too meaningless to deserve discussion". That prediction is the number people later turn into a pass mark: a 30% chance of fooling the judge.

The paper is also a defense. The Stanford Encyclopedia lists nine objections Turing answers, including the theological objection, the mathematical objection, the argument from consciousness and Lady Lovelace's. His answer to Lovelace is that the evidence available to her did not encourage her to believe such machines could originate things, and that machines of that kind might have the property in a sense. Late in the paper he asks why not simulate a child's mind instead of the adult's and then educate it, the idea of a learning machine.

Then he took the argument to radio. His first BBC Third Programme talk, "Can Digital Computers Think?", aired in May 1951, according to Diane Proudfoot in IEEE Spectrum. His second was a symposium recorded on 10 January 1952 and broadcast on 14 January, with Max Newman, Geoffrey Jefferson and Richard Braithwaite. The King's College Turing Archive catalogues the typescript of the discussion, "Can automatic calculating machines be said to think?", and lists transmissions on 14 and 23 January 1952. Proudfoot reports that in the BBC broadcasts Turing estimated at least 100 years would pass before a machine could hold its own in unrestricted questioning. Proudfoot uses that estimate to challenge the popular claim that he expected a machine to pass around the year 2000, since the 1950 prediction was only about a 70 per cent ceiling on correct identifications after five minutes.

Notice what is missing. The paper specifies a game, not a program. No machine could play it in 1950, and for the next decade the test lived in philosophy journals. That is why this era is called the paper years: the idea had been written down and widely read, but nothing had been built to take it. The field that would try to build such a program, artificial intelligence, was named a few years later at the 1956 Dartmouth workshop.

For teams today: Turing's real contribution was to trade an argument about definitions for a procedure. When a debate about "does the tool really understand?" stalls, define a test everyone can run and score.

The Contested Test (1966-1989)

Programs fool judges, and philosophers question the judging

By the mid-1960s there were programs that could converse in text, and the first thing they showed was how easy it was to be fooled. Joseph Weizenbaum's ELIZA, described in 1966, is the standard example. The Stanford Encyclopedia puts it plainly: quite unintelligent programs such as ELIZA can seem to ordinary observers to be loci of intelligence for quite extended periods of time. The story of that program, its therapist script and Weizenbaum's alarm at how people treated it is told in the first AI chatbots. The point for this history is narrower. ELIZA was the first evidence that a Turing-style judgment says as much about the judge as about the machine.

The first controlled experiment in the Turing spirit came in 1972. As Stuart Shieber recounts in his 1994 analysis of the Loebner contest, Kenneth Colby and colleagues published "Turing-like indistinguishability tests for the validation of a computer simulation of paranoid processes" in Artificial Intelligence. Psychiatrists were given transcripts of electronically mediated dialogues with PARRY, Colby's program that played a paranoid patient, and with actual paranoid patients, and were asked to pick the simulation. They did no better than chance. Shieber adds the caution: this has been credited to the fact that unresponsiveness and non sequitur are typical behaviors of paranoids. The program's persona lowered the bar. The same trick returns throughout this story.

Then came the philosophical objection that outlived all the programs. In 1980 John Searle published "Minds, Brains, and Programs" in Behavioral and Brain Sciences, and the Stanford Encyclopedia summarizes the Chinese Room this way. Searle imagines himself alone in a room following a computer program for responding to Chinese characters slipped under the door. He understands nothing of Chinese, yet by following the program he sends back appropriate strings of characters, and people outside suppose a Chinese speaker is inside. The man in the room is doing what the rule-following programs of symbolic AI did: manipulating symbols he does not understand. The narrow conclusion, in the entry's words, is that programming may make a computer appear to understand language but could not produce real understanding, and hence the Turing Test is inadequate. The argument drew many replies. The encyclopedia says it has probably been the most widely discussed philosophical argument in cognitive science since the Turing Test itself.

A different attack came from Ned Block in 1981. The Stanford Encyclopedia describes his Blockhead, a creature controlled by a look-up tree holding a programmed response for every possible input, as a counterexample to the claim that passing the test is sufficient for intelligence. Turing and Newman had answered a version of it in the BBC broadcasts: Proudfoot reports they argued that brute-force matching against a database would take an impractically long time.

By the end of the 1980s the test had an odd status. No program passed it in a way anyone trusted, and philosophers had split on whether passing would mean anything. It was an argument that kept running without a way to run the test itself.

For teams today: a convincing demo is not evidence of competence. The ELIZA and PARRY episodes say to test with skeptical, qualified judges and to ask whether the system's persona is hiding the weak spots.

The Prize Years (1991-2014)

Contests multiply, and claims of a pass get contested

On November 8, 1991, an eclectic group of academics, business people, press and passers-by filled two floors of the Computer Museum in Boston for a tournament billed as the first actual administration of the Turing test. Shieber's report says it was the first attempt on the newly constituted Loebner Prize. Ten judges rated ten agents, six of them programs, and the committee restricted each program to one topic. Joseph Weintraub's program, on whimsical conversation, ranked highest and won the $1,500 prize. Five judges ranked it as human. A second contest followed on December 15, 1992, with Weintraub's program taking first prize again, this time under the topic "men vs. women". Shieber's account is also covered in more detail in our history of AI chatbots.

Shieber's verdict is the most useful piece of criticism from this period. He compared the contest to the Kremer prize for human-powered flight, established in 1959 and won in 1977, and argued that it worked because its goal was clear. The Loebner contest, he argued, lacked that clarity of purpose, and its restrictions let entrants win with tricks. He observed that verbatim repetition of earlier statements, answers keyed to trigger words and similar tricks of the ELIZA trade were everywhere in the transcripts. The Stanford Encyclopedia reports a related view: that the Loebner entries were easily detected by fairly simple questioning, and that Shieber called the competition an embarrassment. It adds that entrants seemed aimed at the best-competitor prize more than at passing the test.

The test also acquired a critic in the research community. In 1995 Patrick Hayes and Kenneth Ford argued that the imitation game was a poor research goal for AI. The Stanford Encyclopedia lists their claims, including that building a machine to pass is so hard no one could seriously adopt it as a goal, and it rejects several of them, including the charge that the test is a "null effect experiment". The debate left the same unresolved question: is it a measure, a goal or a distraction?

On June 7, 2014, the University of Reading held Turing Test 2014 at the Royal Society in London. According to Reading's release, republished by ScienceDaily, a program called Eugene Goostman, which simulates a 13-year-old boy and was developed in Saint Petersburg by Vladimir Veselov and Eugene Demchenko, convinced 33% of the judges that it was human in five-minute keyboard conversations. The organizers' threshold was more than 30%, and Kevin Warwick of Reading called the result a milestone. He also acknowledged that some would say the test had already been passed elsewhere. The developers said the teenage character made it reasonable for the program not to know everything.

Experts were not persuaded. The Stanford Encyclopedia notes that similar one-off results had occurred before: PC Therapist in 1991 with 50% of judges fooled, and Cleverbot in 2011 with a higher rate. In all three cases the trial was very small and the result was not reliably projectible, and in none was there strong ground for holding that an average interrogator had no more than a 70% chance of a right determination after five minutes. In other words, a number crossed a threshold, but the pass was a headline, not a replicated finding.

This era solved the question of how to run the test in public. It created a different one: a contest rewards whatever wins the contest. Persona, topic limits and short conversations all made passing easier without making the machine more capable.

For teams today: treat benchmark wins skeptically when the setup can be tuned. Ask who set the topic, how long the test ran, and whether the result has been repeated by someone else.

The Comeback (2020-present)

Large language models make the old test hard to dismiss

After OpenAI's GPT-3 arrived in 2020, the Stanford Encyclopedia could still say that it was not close to passing: it generated fiction, code and news articles, but neither perceived nor acted. It also records David Chalmers' speculation that GPT-3 "suggests a potential mindless path to artificial general intelligence", a question taken up in our guide to what AGI means. The test came back with the large language models that followed.

The best-documented evidence comes from Cameron Jones and Benjamin Bergen of UC San Diego. In an October 2023 preprint, Does GPT-4 pass the Turing test?, they reported that the best GPT-4 prompt was judged human in 49.7% of games, against 22% for ELIZA, 20% for GPT-3.5 and 66% for human participants. It was not a pass, but it was far above the older programs.

Their 2025 preprint, Large Language Models Pass the Turing Test, submitted March 31, 2025, used the original three-party format. In two pre-registered studies, one with UC San Diego undergraduates and one with Prolific participants, an interrogator held five-minute text conversations with two witnesses at once, one human and one AI, with each participant completing 8 rounds. When prompted to adopt a humanlike persona, GPT-4.5 had a win rate of 73% (69% with undergraduates, 76% with Prolific participants), and LLaMa-3.1-405B had 56%. Without the persona prompt, the numbers fell to 36% and 38%. ELIZA scored 23% and GPT-4o 21%. The authors used ELIZA as a manipulation check: if interrogators were guessing at random, ELIZA would not have been picked this rarely. The paper is a preprint, so treat the figures as reported results.

The authors are careful about what it means. They write that the test is a measure of substitutability, whether a system can stand in for a real person without the other party noticing, and not a direct test of intelligence. The persona prompt matters too: the same model scored 73% with one and 36% without. Machines that chat like people are the subject of what an AI chatbot is, and the underlying skills belong to natural language processing. The result says nothing about whether the models understand anything, which is the question Searle raised in 1980.

What this era solved is the 75-year-old question of whether a machine could be indistinguishable from a person in short text chats. The answer, in a controlled study, was yes under favorable conditions. The new problem is a practical one: if text can no longer separate people from models, trust has to come from somewhere else.

For teams today: assume customers, candidates and employees can't reliably tell a model's message from a person's, and decide where you will disclose that a machine is talking.

Key People

Key Facts

• Turing's paper appeared in Mind, volume LIX, issue 236, pages 433-460, October 1950.
• Turing predicted that in about fifty years an average interrogator would have no more than a 70 per cent chance of a right identification after five minutes.
• In the BBC broadcasts of 1951 and 1952, Turing estimated that at least 100 years would pass before a machine could hold its own in unrestricted questioning.
• In 1972, psychiatrists did no better than chance at picking the simulated paranoid patient from the real ones.
• Eugene Goostman convinced 33% of judges on June 7, 2014, against a threshold of more than 30%.
• In 2025, GPT-4.5 with a persona prompt was judged human 73% of the time, and 36% without one.

The Pattern Behind the History

The same four moves repeat across the five eras. We call the sequence the Pass-Then-Redefine Loop. It's our synthesis of the history above, not a research finding.

1. Define intelligence by behavior. Descartes used language, Turing used a game, and Jones and Bergen used a three-party chat.
2. Win the narrow version. ELIZA fooled observers, PARRY beat psychiatrists at chance, Weintraub's program won on whimsical conversation, and Goostman cleared 30% with a teenage persona.
3. Contest what the win means. Searle said passing is not understanding, Shieber said the contest had no purpose, and the encyclopedia said the Goostman trial was too small.
4. Redefine the test. Each rebuttal produced a tougher test: unrestricted topics, a three-party format, a control like ELIZA, and finally the claim that it measures humanlikeness and substitutability, not intelligence.

What It Teaches Teams Today

• Turn the argument into a procedure. Turing's best move was to stop defining and start testing. Define the 20 tasks that matter and score the system on them.
• Beware short, friendly tests. Five minutes, one topic and a forgiving persona made earlier passes easy.
• Include a baseline. Jones and Bergen used ELIZA to prove the judges could tell. Compare any AI against a simple tool and a human.
• Watch the prompt. The same model scored 73% with a persona and 36% without. Measure the setup you will actually deploy.
• Separate "sounds human" from "is right". Fluency is the thing the test measures, and correctness is a different question.
• Decide your disclosure policy. If text can't separate people from models, say when a customer is talking to one.

The Turing Test began as a way to stop arguing about definitions, and it has spent seventy-five years teaching the opposite lesson: whatever you measure, the machine will eventually match, and the real question moves. Read any new AI result by asking what it measured, who judged it, and how it would look under a harder version.