Turing test: The question that started it all

The event and the specific claim

In October 1950 the philosophy journal Mind published a paper by Alan Turing that opened with a deceptively simple sentence: "I propose to consider the question, 'Can machines think?'" Turing then did something more radical than answer it. He declared the question itself "too meaningless to deserve discussion" and replaced it with a concrete procedure he called the imitation game. That substitution, argued in Computing Machinery and Intelligence, is the event: not the construction of a thinking machine, and not a proof that one is possible, but the reframing of an unanswerable metaphysical dispute into a description of an observable test.

The narrow documented claim is modest and precise. Turing proposed an evaluation, described how it would run over a teleprinter, predicted a specific level of future performance, and rebutted nine anticipated objections. He built nothing. The broad historical judgment, that this paper set the terms for how the field would later argue about machine intelligence, comes from reading the decades of work that treated his game as a reference point. Keeping those two levels of claim distinct is the whole task of placing the event correctly.

From the universal machine to a question about thinking

Turing did not arrive at the imitation game as a philosopher unfamiliar with machines. Fourteen years earlier he had defined, in On Computable Numbers, with an Application to the Entscheidungsproblem, the abstract device now called the Turing machine, and with it the idea of a universal machine that could simulate any other by reading a description of it from its tape. That result established a foundation the 1950 paper leans on quietly: a single general-purpose machine can, in principle, imitate the behaviour of any special-purpose one, so the interesting question is not which physical mechanism computes but what its behaviour can be made to look like.

The constraint Turing was addressing was therefore not hardware. It was language. Words like "think" and "machine" carried too much freight to support a clean argument, and any survey of ordinary usage, Turing noted with some sarcasm, would be settled by a Gallup poll rather than by evidence. The Stanford Encyclopedia of Philosophy account of Alan Turing traces how his mathematical logic and his wartime engineering fed into this move to sidestep definition in favour of a test. Rather than define thought, he proposed to fix a situation in which a candidate for thought could be examined behaviourally, and to let the situation carry the argument.

How the imitation game actually works

The mechanism is best read exactly as Turing wrote it, because popular summaries flatten it. The game begins with three human participants: a man (A), a woman (B), and an interrogator (C) of either sex in a separate room. Communication runs only through typewritten messages, so voice and appearance are removed. The interrogator's job is to determine which of the two unseen players is the man and which is the woman; the man tries to cause a wrong identification, and the woman tries to help the interrogator toward the right one. Only then does Turing introduce the pivot: "What will happen when a machine takes the part of A in this game?"

That original three-party, gendered framing has produced a lasting interpretive dispute, catalogued in the Stanford Encyclopedia of Philosophy entry on the Turing test. On one reading the machine must imitate a woman as convincingly as a man does; on the more common "standard interpretation" the machine simply plays against a human while the interrogator tries to tell computer from person. What both share, and what is genuinely load-bearing, is the deliberate isolation of conversational response from every other property. The test asks what can be inferred from language alone, and says nothing about how the responses are produced.

The numbers Turing actually put on the table

Turing did not leave the test purely qualitative. He offered a dated, quantified prediction, and the discipline of the report is to treat it as a forecast rather than an achievement. In the same Mind paper he wrote that in about fifty years' time it would be possible to program computers with a storage capacity of about 10^9 binary digits (on the order of a hundred megabytes) to play the imitation game so well that an average interrogator would have no more than a 70 per cent chance of making the correct identification after five minutes of questioning.

Every clause there is a specification, not a scoreboard. The horizon is roughly the year 2000; the storage figure is an engineering guess about capacity; the 70 per cent is a threshold for a five-minute interaction, not a claim that a machine would be 70 per cent as intelligent as a person. Turing was careful to add that he expected the original question, "Can machines think?", to become one people could discuss without contradiction by the end of the century, because usage and educated opinion would have shifted, not because the metaphysics would have been resolved. The prediction is about words and expectations at least as much as about machines.

The learning machine Turing also proposed

The imitation game is the famous part of the paper, but its final pages contain a second idea that later research arguably vindicated more directly than the test itself. Turing asked how one might ever build a machine good enough to play, and rejected the notion of hand-programming an adult mind. "Instead of trying to produce a programme to simulate the adult mind," he wrote in the same 1950 essay, "why not rather try to produce one which simulates the child's?" Such a child machine would then be subjected to a course of education, its behaviour shaped by rewards and punishments and by exposure to experience.

This is a recognizable sketch of learning from data rather than from explicit instruction, and Turing discussed random elements, teaching, and the modification of the machine's own rules. The interpretive caution is not to overclaim: he did not specify a training algorithm, a loss, or a network, and drawing a straight line from these paragraphs to modern machine learning would flatter a passage that is genuinely programmatic. What is fair to say is that the paper separated two problems that are still separate today, namely how to evaluate a candidate intelligence and how to bring one into being, and that Turing bet the second would be solved by education rather than by enumeration.

Nine objections, answered in advance

A large share of the paper is devoted to rebutting objections Turing expected, and the list reads as a preemptive map of a debate that has recurred ever since. He addressed the theological claim that thinking requires a soul; the "heads in the sand" wish that machines could not think because the consequences would be dreadful; and the mathematical objection drawn from Godel-style incompleteness, which he argued applied to any formal system, human reasoning included, and so did not single out machines. He took up Geoffrey Jefferson's argument from consciousness, that a machine could not truly feel, replying that its strict form leads to solipsism about other people too.

The most cited of these is Lady Lovelace's objection, her nineteenth-century remark that the Analytical Engine "has no pretensions to originate anything" and can only do what it is ordered to do. Turing's reply, preserved in the primary text and analysed in the Turing test entry, was that machines constantly surprised him, and that "originate" smuggles in an untested assumption. Notably, he found the argument from extrasensory perception the hardest to dismiss, a reminder that the paper is a document of 1950 and not a timeless oracle.

The competing visions of the era

Turing's operational, behaviour-first proposal was one option among several for what a science of machine intelligence might even be about. Six years later the founding statement of the field took a different tack. The Proposal for the Dartmouth Summer Research Project on Artificial Intelligence, drafted by John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon, conjectured that "every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it." Where Turing proposed to judge a black box by its conversation, the Dartmouth program set out to build the mechanisms of thought explicitly.

That constructive ambition hardened into the symbol-processing tradition. In their Turing Award lecture on computer science as empirical inquiry, Allen Newell and Herbert Simon advanced the physical symbol system hypothesis: that "a physical symbol system has the necessary and sufficient means for general intelligent action." This is a much stronger commitment than Turing's. It says what intelligence is made of, not merely how to test for it. Reading the 1950 paper against Dartmouth and Newell and Simon shows that the imitation game was consciously agnostic about internal architecture, which was both its strength as a criterion and, to its critics, its central evasion.

ELIZA, the Chinese Room, and the attack on the test

The first serious blow to naive confidence in the test came from a working program. In 1966 Joseph Weizenbaum described ELIZA, a pattern-matching script whose best-known persona, DOCTOR, mimicked a Rogerian psychotherapist by reflecting users' statements back as questions. ELIZA understood nothing, yet people attributed feelings and comprehension to it, a reaction so strong that Weizenbaum grew alarmed at how little was required to elicit it. The lesson was uncomfortable for the imitation game: surface plausibility over a short exchange is easier to manufacture, and easier to be fooled by, than Turing's framing might suggest.

The deeper philosophical challenge came from John Searle. His 1980 paper Minds, Brains, and Programs introduced the Chinese Room, in which a person who knows no Chinese follows a rule book to manipulate Chinese symbols and produces answers indistinguishable from a native speaker's, all without understanding a word. Searle's target, examined at length in the Chinese Room entry, was precisely the inference the test invites: that behaving as if one understands is the same as understanding. His conclusion, that syntax is not sufficient for semantics, does not refute the test as a test, but it denies that passing it settles the question of genuine thought.

The empirical revival and what recent studies actually claim

For most of its life the imitation game was a thought experiment, not an experiment. That changed as large language models made controlled trials feasible. In 2023 Cameron Jones and Benjamin Bergen ran a public online study reported as Does GPT-4 Pass the Turing Test?, in which the best GPT-4 prompt was judged human in 49.7 per cent of games, above ELIZA at 22 per cent and GPT-3.5 at 20 per cent, but below the 66 per cent rate at which real humans were identified as human. Interrogators, the authors found, leaned on linguistic style and socioemotional cues rather than on factual competence.

A stronger result followed in 2025. In Large Language Models Pass the Turing Test, the same researchers ran a preregistered, controlled three-party test and reported that GPT-4.5, when prompted to adopt a humanlike persona, was judged to be the human 73 per cent of the time, more often than the actual human participants were, while a Llama-3.1 model reached 56 per cent and unpersonated baselines fell well below chance. The precise and limited claim is that, in a five-minute three-party setup, some systems can now be selected as human at least as often as people are. That is a fact about a specific protocol, a specific prompt, and a short interaction, and it says nothing about understanding, reliability, or consciousness.

The honest limits of a conversational test

Turing himself hedged more than his popularizers do. He did not claim that passing the game proves a machine thinks, only that a machine able to play it well would make the original question uninteresting to press. The distinction matters because a great deal follows from conflating the two. A short text exchange rewards the imitation of human style, including its evasions, typos, and refusals to answer, and the 2025 study's reliance on a persona prompt underlines that success there is partly a matter of performing humanity rather than demonstrating capability.

The critiques compound the caution. ELIZA showed that people over-attribute mind from thin evidence, so a positive result partly measures the interrogator. Searle argued that even flawless behaviour leaves the question of understanding open. And nothing in the test speaks to competence outside conversation: a system that convincingly chats for five minutes may still be unreliable at arithmetic, prone to fabrication, or useless at any task it was not coaxed toward. Passing a conversational test establishes indistinguishability under stated conditions, which is a real and measurable property, and is not the same as establishing intelligence, safety, or truthfulness.

How to place the event in AI history

The most defensible reading treats Turing's 1950 paper as a conceptual turning point rather than a technical one. It invented no algorithm and reported no experiment; its contribution was to convert an argument about definitions into a proposal about evidence, and to do so with enough clarity that the field could adopt, attack, and refine it for seventy-five years. The narrow claim is that Turing described a behavioural test and predicted a dated, quantified level of performance. The broad claim, that he framed the debate, is supported not by the paper alone but by the long line of work, from Dartmouth and the symbolists to ELIZA, Searle, and the recent trials, that keeps returning to his terms.

A reader navigating the timeline can use three questions to move outward from this event. Which earlier idea made it possible? The universal machine of 1936. Which rival program rejected its agnosticism about mechanism? The symbol-processing agenda of Dartmouth and of Newell and Simon. And which later result tested its prediction, and with what caveats? The controlled studies that now report systems passing under specific protocols, while leaving Searle's question exactly where he left it. The event matters because it gave a young field a shared, falsifiable question, not because it answered whether machines can think.

References

• Turing, A. M. (1950). Computing Machinery and Intelligence.
• Turing, A. M. (1936). On Computable Numbers, with an Application to the Entscheidungsproblem.
• Stanford Encyclopedia of Philosophy (2021). The Turing Test.
• Stanford Encyclopedia of Philosophy (2013). Alan Turing.
• Stanford Encyclopedia of Philosophy (2020). The Chinese Room Argument.
• Searle, J. R. (1980). Minds, Brains, and Programs.
• Weizenbaum, J. (1966). ELIZA: A Computer Program for the Study of Natural Language Communication Between Man and Machine.
• McCarthy, Minsky, Rochester and Shannon (2006). A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence.
• Newell, A. and Simon, H. A. (1976). Computer Science as Empirical Inquiry: Symbols and Search.
• Jones, C. and Bergen, B. (2023). Does GPT-4 Pass the Turing Test?.
• Jones, C. and Bergen, B. (2025). Large Language Models Pass the Turing Test.