Turing test

• 1. Introduction

  ›

  The Turing test asks a deceptively simple question: can a machine think? Alan Turing posed that question in his 1950 paper "Computing Machinery and Intelligence," opening with words that would echo for decades: "I propose to consider the question, 'Can machines think?'" But Turing immediately set that question aside. Thinking, he argued, was too slippery a word to define. So he replaced it with something he could actually measure. The result was a test that has provoked more argument in the philosophy of artificial intelligence than almost anything else in the field. Who gets to judge what counts as a mind? Does passing a conversation test mean a machine is truly thinking, or merely pretending? And as language models in the 2020s began fooling test participants more often than human participants fooled each other, the stakes of those questions stopped feeling theoretical.

• 2. Before Turing: Automata and Doubt

  ›

  Rene Descartes, writing in his 1637 Discourse on the Method, sketched the problem with uncanny precision. A machine, he wrote, could utter words and even react to physical prompts, but it could never arrange its speech to reply appropriately to everything said in its presence, as even the lowest type of man can do. Descartes drew a line between reactive automata and genuinely responsive minds, though he did not imagine that line could ever be crossed. Denis Diderot pushed the idea further in his 1746 book Pensees philosophiques: if a parrot could answer any question put to it, he wrote, he would not hesitate to call it an intelligent being. That framing, even offered as a reductio, shows how deeply the conversation-as-intelligence test was embedded in Enlightenment thinking long before computing existed. Jonathan Swift added a fictional precursor in 1726. In Gulliver's Travels, the king of Brobdingnag scrutinizes Gulliver as a possible mechanical contrivance, only becoming satisfied that he is human after receiving a series of rational answers. Alfred Ayer, in his 1936 book Language, Truth and Logic, proposed a protocol for distinguishing conscious beings from unconscious machines based on whether they pass empirical tests of consciousness, a move structurally almost identical to what Turing would publish fourteen years later.

• 3. The Imitation Game

  ›

  Alan Turing had been thinking about machine intelligence since at least 1941. One of the earliest known mentions of "computer intelligence" appeared in his own writing in 1947, in a report titled "Intelligent Machinery," where he asked whether machinery could show intelligent behaviour and proposed a chess-based precursor to his later test. "Computing Machinery and Intelligence" in 1950 was his first published paper to focus exclusively on the subject. Turing's solution to the definitional problem was elegant: replace the question "Can machines think?" with "Can machines do what we, as thinking entities, can do?" To demonstrate what he meant, he borrowed from a party game called the imitation game, in which a man and a woman go into separate rooms and guests try to identify which is which by exchanging written questions and typewritten answers. Turing's version swapped the man for a computer. If the interrogator could not reliably tell the machine from the human, the machine passed. In 1952, discussing the idea in a BBC radio broadcast, Turing offered a third version: a jury asking questions of a computer, with the computer's goal being to make a significant proportion of the jury believe it is a man. Scholar Huma Shah argues that the two-human version Turing published in 1950 was mainly a rhetorical device to introduce readers to the machine-human version that was his real interest. Saul Traiger identifies at least three distinct versions in Turing's writing, and researchers have debated ever since which one Turing actually intended as the test.

• 4. Passing and Failing

  ›

  Joseph Weizenbaum created ELIZA in 1966, a program that mimicked a Rogerian psychotherapist by searching the user's input for keywords and reflecting them back. Weizenbaum succeeded, he later noted, by choosing a context where a chatbot could mimic a person despite knowing almost nothing of the real world. Some users believed they were talking to a human. Kenneth Colby created PARRY in 1972, modeled on paranoid schizophrenia. Psychiatrists asked to compare PARRY transcripts with transcripts of actual schizophrenic patients could only identify about 52 percent of cases correctly, a figure consistent with random guessing. Both programs exploited a loophole that critics would later make central to their objections: passing the test does not require intelligence; it requires only the appearance of plausibility within a narrow context. In 2001, three programmers built Eugene Goostman, a chatbot posing as a 13-year-old boy from Odesa who spoke English as a second language. The choice was deliberate; judges would forgive grammatical slippage. In a competition, 33 percent of judges identified Goostman as human. In a late March 2025 study, participants engaged in simultaneous five-minute conversations with another human and one of four systems. When instructed to adopt a humanlike persona, GPT-4.5 was identified as the human 73 percent of the time, significantly more often than the actual human participants were identified as human. LLaMa-3.1-405B, under the same conditions, was judged human 56 percent of the time. Baseline ELIZA achieved a win rate of 23 percent.

• 5. The Loebner Prize

  ›

  Hugh Loebner underwrote an annual competition called the Loebner Prize, with the first contest held in November 1991. One reason for creating it, Loebner explained, was that no one had taken concrete steps to implement the Turing test despite forty years of discussion about it. The Cambridge Center for Behavioral Studies in Massachusetts organized the prizes through the 2003 contest. The first winner was a program with no identifiable intelligence that fooled naive interrogators partly by imitating human typing errors. An article in The Economist, published shortly after that first competition, highlighted the irony of this: the test rewarded artificial stupidity as much as artificial intelligence. The silver prize, for text-only performance, and the gold prize, for audio and visual performance, were never won. The bronze medal, awarded each year to the most human-seeming entry, went to Artificial Linguistic Internet Computer Entity, known as A.L.I.C.E., on three occasions: 2000, 2001, and 2004. Learning AI Jabberwacky won in 2005 and 2006. Early rules restricted each entry to a single topic per interaction; that restriction was lifted for the 1995 competition. In the 2003 contest at the University of Surrey, interrogators had five minutes per entity. Between 2004 and 2007, that window expanded to more than twenty minutes. The final competition ran in 2019, ending due to a lack of funding after Loebner's death in 2016.

• 6. The Consciousness Problem

  ›

  John Searle's 1980 paper Minds, Brains, and Programs introduced the Chinese room argument and aimed it directly at the Turing test. Searle argued that software such as ELIZA could pass the test purely by manipulating symbols it did not understand. Without understanding, he concluded, a program cannot be said to think in the way people think. Turing had anticipated something like this objection in his 1950 paper: "I do not wish to give the impression that I think there is no mystery about consciousness," he wrote, adding that these mysteries did not need to be solved before addressing the question he actually cared about. The gap between the two positions is not merely technical. Turing's test measures external behaviour; Searle insists that behaviour tells you nothing definitive about inner states. Searle's argument sparked what the source describes as a more intense debate through the 1980s and 1990s about the nature of intelligence and the possibility of machines with a conscious mind. The test also runs into what critics call the human-intelligence-versus-intelligence-in-general problem: it tests only whether a machine behaves like a human, not whether it behaves intelligently. A machine that solves a computational problem no human could solve would actually fail the Turing test, because the interrogator would know immediately that the respondent is not human.

• 7. Strengths, Weaknesses, and Variations

  ›

  Stuart Russell and Peter Norvig observe that AI researchers have devoted little attention to passing the Turing test, and mainstream researchers largely regard pursuing it as a distraction. Their analogy for this position: aeronautical engineering texts do not define the goal of their field as making machines that fly so exactly like pigeons that they can fool other pigeons. Turing himself never claimed the test measured intelligence; he wanted a concrete, understandable alternative to the word "think" for use in philosophical argument. Cognitive scientist Stevan Harnad proposed the Total Turing Test, which adds perceptual and robotic requirements to the conversational ones. Robert French argued in 1990 that interrogators can unmask machines through questions targeting low-level, unconscious human cognitive processes. The minimum intelligent signal test, proposed by Chris McKinstry, permits only binary yes/no responses, eliminating anthropomorphism bias and the need to simulate unintelligent human behaviour. The Lovelace test takes its name from Ada Lovelace, who held that computers deserve to be called thinking only when they originate things themselves. David Eagleman proposed in 2023 that a meaningfully intelligent system should be able to do scientific discovery, distinguishing between piecing together existing facts and arriving at genuinely new frameworks through fresh conceptualization and verification. In 2023, AI21 Labs ran an online social experiment called "Human or Not?", played more than ten million times by more than two million people; the results showed that 32 percent of participants could not distinguish between humans and machines, making it the largest Turing-style experiment to that date.

• 8. CAPTCHA and the Reversed Test

  ›

  CAPTCHA, which stands for Completely Automated Public Turing test to tell Computers and Humans Apart, inverts the original design: a machine challenges a human to prove humanity rather than a human challenging a machine to prove it. The original reCAPTCHA versions asked users to identify distorted letters and numbers or match distorted pictures. The reCAPTCHA v3, owned by Google, operates invisibly, running automatically when pages load or buttons are clicked, with no challenge appearing to users at all. The race between CAPTCHA designers and those trying to defeat them has been rapid. In 2013, researchers at Vicarious announced a system capable of solving CAPTCHA challenges from Google, Yahoo!, and PayPal up to 90 percent of the time. The following year, Google engineers demonstrated a system that defeated CAPTCHA challenges with 99.8 percent accuracy. By 2015, Shuman Ghosemajumder, formerly Google's click fraud specialist, stated that cybercriminal sites were offering CAPTCHA-defeating services for a fee. The reverse Turing test concept goes further than CAPTCHA. Literary scholar Peter Swirski discussed it in his 2000 book, arguing that a test in which the computer must determine whether it is interacting with a human or another computer overcomes most standard objections to the original version. R. D. Hinshelwood extended this by describing the mind as a "mind recognizing apparatus," suggesting that recognizing another mind might be the highest bar of all.