Turing Test
The Turing test is a proposed way of deciding whether a machine can be said to think by examining its conversational behavior instead of its internals: a human judge exchanges text messages with an unseen counterpart and tries to tell whether it is a person or a machine, and if the judge cannot reliably tell, the machine has passed. It was introduced by Alan Turing in his 1950 paper "Computing Machinery and Intelligence," published in the journal Mind, and it remains the best-known thought experiment in the field of artificial intelligence, even though almost no serious research program has ever been organized around passing it.
Turing deliberately swapped an unanswerable question for an operational one. The paper opens by calling the question "Can machines think?" too ill-defined to discuss, and replaces it with the "imitation game": an interrogator in a separate room puts written questions to two hidden players, a man and a woman, and must decide which is which, with the man trying to deceive. Turing then asks what happens when a machine takes the deceiver's role. Typed text matters because it strips away voice, appearance and physical presence, so that only linguistic competence is judged. Turing also predicted that by about the year 2000 a machine could fool an average interrogator about 30 percent of the time after five minutes of questioning, a modest, probabilistic target that is often misquoted as a pass-or-fail threshold he never set.
The test is easy to game, which is the central lesson of its history. Joseph Weizenbaum's ELIZA (1966), a script of keyword-triggered reflection rules that imitated a Rogerian therapist, made some users feel understood despite understanding nothing, a reaction now called the ELIZA effect. The Loebner Prize, begun in 1991, staged annual restricted conversations, and winning bots relied on typos, evasions, jokes and topic changes rather than comprehension. In 2014 a chatbot presenting itself as a 13-year-old Ukrainian boy, Eugene Goostman, was announced as having passed at an event at the Royal Society; many researchers rejected the claim because the persona excused ignorance and the test was short. Judges are fallible, and a test of fooling people measures the judges as much as the machine.
Philosophers have attacked the inference from behavior to understanding. John Searle's Chinese room argument (1980) imagines a person who follows a rulebook to manipulate Chinese symbols and produces fluent replies without understanding a word, suggesting that symbol manipulation alone is not comprehension. Others reply that the whole system, rulebook included, might understand, or that behavior is all anyone ever has as evidence of other minds. Turing anticipated many objections in the paper itself, including the theological objection, the argument from consciousness and Lady Lovelace's claim that machines can only do what they are told. The dispute is not settled, but it clarifies what the test can and cannot claim: it addresses whether a machine is behaviorally indistinguishable from a person, not whether anything is experienced inside.
Modern language models have changed the debate from fantasy to measurement. Fluent open-ended conversation is now routine, and controlled studies have been run in which judges chat with a human and a model and choose which is human; recent three-party experiments by Cameron Jones and Benjamin Bergen reported that a persona-prompted large language model was often judged to be the human more frequently than the real human participants. These results show that conversational mimicry no longer discriminates well, and they sharpened a long-standing view that the test is a weak yardstick of general ability. Researchers have proposed alternatives that probe specific skills, such as the Winograd schema challenge (Hector Levesque and colleagues, 2012), which tests commonsense pronoun resolution, and the broad task suites used to evaluate models today. Practical concerns have also moved toward disclosure and impersonation, since a system that passes as human raises questions about fraud and consent.
Common misconceptions concern what passing would mean. The test does not require a machine to be correct, only convincing, and a perfectly informed assistant would fail it by answering arithmetic instantly and never making human slips; Turing's own example dialogue includes a machine pausing and giving a wrong sum. It is also not a single historical event, since no agreed protocol, duration or judge pool exists, so claims of "passing" must always be read with the setup in mind. Nor was it ever a formal definition of intelligence: Turing offered it as a sufficient criterion for attributing thought, not a necessary one, leaving open that a machine could think differently from people and still fail. Treating it as a milestone to be crossed obscures its value as a thought experiment.
Read the Turing test as a provocation about evidence rather than a benchmark: it asks what observable behavior would justify calling a machine intelligent, and its history, from ELIZA to chatbots that fool judges while understanding nothing in the human sense, shows that being convincing in conversation is both achievable and insufficient for settling what thinking really is.