What Is the Turing Test?


5 min read≈ 1,190 words3 questions · up to 3 points
What this is about
In his famous 1950 paper, Alan Turing wrote a sample conversation in which his imagined machine gets a simple sum wrong. You learn how his test works, what happened when real chatbots took it from 1966 to 2025, and why the strongest objection says that passing it proves less than it seems.
In a nutshell· skip ahead
The Turing test is a game Alan Turing proposed in 1950. A judge chats in writing with two hidden partners, one human and one machine, and tries to tell which is which. If the judge cannot reliably pick out the machine, it passes. The test measures how humanlike a machine's conversation seems, not whether the machine understands anything.
In 1950 the mathematician Alan Turing published a short dialogue between a questioner and a machine. The questioner asks it to add 34,957 and 70,764. The machine waits about thirty seconds and answers 105,621, although the right answer is 105,721. Why would one of the founders of computer science give his imagined machine a wrong answer?
How does Turing’s imitation game work?
The Turing test is a game in which a judge holds a written conversation with two hidden partners, a human and a machine, and tries to work out which one is the machine. If the judge cannot reliably tell them apart, the machine passes. Turing called it the imitation game, and he built it in two steps.
- In the first version there are three people: a man (A), a woman (B) and an interrogator (C) in a separate room. C knows the other two only as X and Y and must decide which is the man. A tries to make C choose wrongly, B tries to help, and all answers are typed, so that voices give nothing away.
- Then Turing replaces A with a machine. If the interrogator decides wrongly as often as when a man plays, the machine has done as well as a person at seeming human.
This explains the wrong sum. Later in the same paper Turing writes that a machine playing the game would make mistakes in arithmetic on purpose to confuse the interrogator, because perfect accuracy would give it away. So the test does not reward the machine for knowing the most. It rewards the machine for behaving the way a person would.
Turing then made a forecast:
I believe that in about fifty years’ time it will be possible to programme computers, with a storage capacity of about 10⁹, 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.
In plain words, Turing expected that by about 2000 a computer could fool an average judge in three conversations out of ten, each five minutes long. This was his guess about progress, not a pass mark.
Has any machine passed the Turing test?
Machines fooled people long before anyone ran a proper test. In 1966 Joseph Weizenbaum at MIT published ELIZA, a program that imitated a psychotherapist by turning the user’s own sentences into questions. It understood nothing, yet in his 1976 book Computer Power and Human Reason Weizenbaum reported that his secretary, after a few exchanges with it, asked him to leave the room. The habit of reading a mind into a program that only reflects your words is now called the ELIZA effect.
The first public contest, the Loebner Prize, took place in November 1991 at the Computer Museum in Boston. Each program could stick to one topic, and the winner, which offered “whimsical conversation”, was taken for a human by five judges. A human expert on Shakespeare, meanwhile, was classed as a computer three times, because at least one judge thought no real person could know so much about Shakespeare.
In June 2014 a chatbot called Eugene Goostman, presented as a 13-year-old boy from Ukraine, convinced a third of its judges in five-minute chats at the Royal Society in London. The organisers announced that the Turing test had been passed, but many researchers disagreed. They said that 30 percent was Turing’s forecast, not his criterion, and that a young non-native speaker has an excuse for every odd answer.
The clearest result so far comes from Cameron Jones and Benjamin Bergen at the University of California, San Diego, who ran Turing’s three-party version with five-minute chats in 2025. GPT-4.5, told to play an introverted young person who knows internet culture and uses slang, was picked as the human 73 percent of the time, more often than the real humans it was compared with. Without that persona its score fell to 36 percent. The study was published in PNAS in 2026.
What the Turing test teaches you
The test teaches you to separate two questions that feel the same: does this sound like a person, and is there real understanding behind it? Fluent text answers only the first.
- Chats and messages. A support chat or a scam message can come from a machine that sounds entirely human. Instead of asking whether it feels human, ask what you could check that a machine would get wrong.
- Judging written work. A smooth essay no longer shows that a person wrote it or understood it, which is why many teachers now ask students to explain their work aloud.
- Your own first impressions. Judges in these tests decide quickly from style and tone. Thinking, Fast and Slow by Daniel Kahneman: Summary and Review explains why such fast judgments feel certain even when they are wrong.
What does the Turing test not measure?
Many people think that a machine that passes the test has been shown to think. Turing claimed less. He called the question “Can machines think?” too meaningless to discuss and replaced it with the game, so the test measures how well a machine imitates human conversation, not what goes on inside it.
The strongest objection comes from the philosopher John Searle. In his 1980 paper “Minds, Brains, and Programs” he imagined himself locked in a room with a rulebook for handling Chinese characters. People outside pass in questions in Chinese, and by following the rules Searle passes out answers good enough to convince them, although he does not understand a word of Chinese. A computer running a program does the same, Searle argued, so producing the right symbols is not understanding them. Critics answer that the whole system of person, rules and paper might understand even if Searle does not, and the debate is still open.
The second limit is that every result depends on the judges and the setup. Judges in 1991 took an expert for a machine, and in the 2025 study the San Diego students spotted the machine more often than online participants. Short chats, a single topic or a persona with a built-in excuse all make passing easier, so a pass tells you about one setup and one group of judges.
Learn more
The test has a clear originator. Alan Turing, born in London in 1912, had described the theoretical computer in 1936 and broken German codes at Bletchley Park during the Second World War. He wrote “Computing Machinery and Intelligence” while working on one of the first computers at the University of Manchester, and he died in 1954. Some philosophers see a forerunner in René Descartes, who argued in 1637 that a machine could never use words to answer sensibly whatever is said to it.
Turing’s forecast assumed a memory of about 10⁹ bits, roughly 125 megabytes, and computers had far more storage than that by 2000. What Is Moore’s Law? explains why hardware grew so predictably while conversation took until the 2020s. For a world in which computers know people better than they know themselves, read “Homo Deus” by Yuval Noah Harari: Summary and Review.
Quiz
3 questions · up to 3 points
Open a question and pick one answer. Only your first answer counts: +1 if it is right, −1 if it is wrong.
In the standard version of Turing's imitation game, what does the judge do?
Show answer
The judge chats in writing with a hidden human and a hidden machine and decides which one is the machine. Turing's game has three players. The judge sees only typed answers from two hidden partners and must say which is which.
A customer chats with a support service for ten minutes and is sure the helper was a person. Later the company says it was an AI system. What does the Turing test idea say about this?
Show answer
The system was humanlike enough to fool this customer in this setting, which shows nothing certain about whether it understood anything. Passing depends on the judge, the length of the chat and the topic. A Turing-style result tells you about humanlike behaviour, not about understanding.
Many people think that a machine that passes the Turing test must be intelligent. Which objection does the article present as the strongest?
Show answer
Searle's Chinese Room: following rules for symbols can produce the right answers without any understanding of what they mean. Searle imagined himself answering Chinese questions with a rulebook while not understanding a word of Chinese. Good answers alone do not show understanding.
Your points are saved only in this browser for now. Accounts that keep them for good are on the way.
Continue learning: What Is Moore's Law? →The takeaway
The Turing test, proposed by Alan Turing in 1950, asks whether a judge chatting in writing with a hidden human and a hidden machine can tell which is which. Simple programs like ELIZA already fooled some people, and in a 2025 study GPT-4.5 with a human persona was picked as the human 73 percent of the time. The test measures how humanlike a conversation seems to particular judges, not understanding, as Searle's Chinese Room argument shows. When a fluent text or chat convinces you, ask what it would take to tell the difference.
Key quotes
“The original question, 'Can machines think?' I believe to be too meaningless to deserve discussion.”
References
- Computing Machinery and Intelligence. A. M. Turing, Mind 59 (236), 433–460, 1950
Turing's original paper: the imitation game, the sample conversation, his prediction for the year 2000 and his replies to nine objections.
- The Turing Test. Graham Oppy and David Dowe, Stanford Encyclopedia of Philosophy, 2021
A careful survey of what Turing claimed, how his prediction should be read, and the main objections to the test.
- The Chinese Room Argument. David Cole, Stanford Encyclopedia of Philosophy, 2024
Searle's 1980 thought experiment, the systems reply and the long debate about whether running a program can produce understanding.
- Lessons from a Restricted Turing Test. Stuart M. Shieber, Communications of the ACM 37 (6), 70–78, 1994
An eyewitness critique of the first Loebner Prize contest in 1991, including the human expert whom judges took for a computer.
- Large Language Models Pass the Turing Test. Cameron R. Jones and Benjamin K. Bergen, arXiv preprint 2503.23674; published in PNAS in 2026, 2025
The three-party study in which GPT-4.5, told to play an introverted young person, was picked as the human more often than the real humans.





