The Turing Test Explained: Can a Machine Pass for Human?

In 1950 the mathematician Alan Turing asked whether machines could think, and then set the question aside as too vague to answer. He proposed a game instead: if a machine could hold a conversation that people could not tell apart from a human's, that would be a test worth taking seriously.

For decades, the Turing test was a distant goal and a favourite of science fiction. In the 2020s, large language models began to pass it. What the test measures, and what it does not, has never mattered more.

Who Alan Turing was

Alan Turing was born in London on 23 June 1912. In 1936 he published a paper that introduced what is now called the Turing machine, a simple mathematical model of a general-purpose computer. It became one of the foundations of computer science.

During the Second World War he was a leading codebreaker at Bletchley Park, pictured on this page, where he helped break German messages encrypted with the Enigma machine and designed the bombe, an electromechanical codebreaking machine. After the war he designed an early computer, the Automatic Computing Engine, and worked on computers at the University of Manchester.

The imitation game

In 1950 Turing published a paper called Computing Machinery and Intelligence. Instead of asking whether machines can think, he proposed replacing the question with a game. A human interrogator holds text conversations with two hidden participants, one a person and one a machine, and tries to work out which is which.

If the machine can imitate a human well enough to fool the interrogator, Turing suggested, we have as good a reason to call it intelligent as we have for anyone else. He predicted that within about fifty years, an average interrogator would have no more than a 70 per cent chance of making the right identification after five minutes.

Early attempts

In 1966 Joseph Weizenbaum's program ELIZA imitated a psychotherapist by turning a user's words back into questions. It was very simple, yet some people became convinced they were talking to a person. In 1972 Kenneth Colby's PARRY imitated a patient with paranoid schizophrenia, and psychiatrists identified it correctly only about 52 per cent of the time, close to guessing.

These programs did not understand anything. They showed, early on, that people can be fooled by surprisingly little, a theme that runs through the whole history of the test.

Prizes and claims

From 1991 the Loebner Prize held an annual Turing test competition. Winning programs often succeeded partly by imitating human typing errors. Its top prizes, for a program indistinguishable from a human, were never awarded, and the competition ended in 2019.

A chatbot called Eugene Goostman, posing as a 13-year-old Ukrainian boy, convinced 33 per cent of judges in one competition. Critics noted that its persona excused its odd answers, which says more about the test's weak points than about the program.

Passing at last

Large language models changed the picture. In March 2024 Stanford researchers reported that GPT-4 passed a rigorous version of the test. In a pre-registered study run in March 2025, GPT-4.5, given a humanlike persona, was judged to be the human 73 per cent of the time, more often than the real humans it was compared with.

In the same study ELIZA was judged human 23 per cent of the time and GPT-4o 21 per cent. By Turing's own standard, his prediction was right, though it took about seventy-five years rather than fifty.

What passing does not prove

Critics have long argued that the test measures how easily people are fooled, not whether a machine is intelligent. In 1980 the philosopher John Searle argued, with his Chinese room thought experiment, that a system could pass by manipulating symbols without understanding any of them.

The rise of chatbots has sharpened that debate. In 2022 a Google engineer claimed the company's LaMDA model was sentient; Google investigated and dismissed the claim. Passing a conversation test, most researchers agree, is a different thing from thinking or feeling.

Turing's fate

Turing did not live to see any of this. In 1952 he was prosecuted for a relationship with a man, then a crime in Britain, and given probation on condition that he undergo chemical castration. He died in 1954, aged 41; the inquest ruled his death a suicide.

In 2009 the British prime minister, Gordon Brown, publicly apologised for how he had been treated, and in 2013 he received a royal pardon. A 2017 law extended pardons to other men convicted under the same laws.

Testing claims about AI

The Turing test is a reminder that convincing is not the same as true. That applies well beyond chatbots: pictures and claims can also look real without being so, as our guide to how experts check whether a photo is fake shows.

When a headline says an AI has done something remarkable, it is worth asking what exactly was measured. Our piece on AI in archaeology applies the same question to another field.