The imitation game is Alan Turing's proposal to replace the question "Can machines think?", which he considered "too meaningless to deserve discussion," with a test of verbal behavior (Turing, 1950, sec. 6). Today we know it as the Turing test.
The game
The original version has three players: a man (A), a woman (B), and an interrogator (C) in another room who communicates with them in writing, ideally by teleprinter. C must work out which one is the man; A tries to mislead C and B tries to help. Turing then asks what would happen if a machine took the part of A (Turing, 1950, sec. 1). Written communication draws a line between physical and intellectual capacities: nobody requires the machine to have human skin.
Turing also made a precise prediction: in about fifty years it would be possible to program computers with a storage capacity of about 10⁹ so that an average interrogator would have no more than a 70% chance of making the right identification after five minutes of questioning. He added that by the end of the century the use of words would have changed so much that one could speak of machines thinking without expecting to be contradicted (Turing, 1950, sec. 6).
The objections
Much of the paper answers nine objections: the theological one, the "heads in the sand" one, the mathematical one (based on Gödel), the argument from consciousness, the arguments from various disabilities, Lady Lovelace's objection (the machine only does what we order it to do), the argument from continuity in the nervous system, the argument from informality of behavior, and the argument from extrasensory perception, which Turing took surprisingly seriously (Turing, 1950, sec. 6; M. Mitchell, 2019, ch. 3). The final section proposes programming not an adult mind but a "child machine" and educating it: the child's brain would be like a notebook fresh from the stationer's, with "rather little mechanism, and lots of blank sheets" (Turing, 1950, sec. 7). It is a remarkable anticipation of machine learning.
Critiques
The most influential objections came later. Searle (1980) called the test "unashamedly behavioristic and operationalistic" (1980, p. 423) (see The Chinese Room), and Block (1981) imagined a machine that would pass it with a table of prerecorded answers (see Blockhead). Chollet (2019) considers it a red herring for the field: it hands evaluation to unreliable human judges and, besides, Turing's game was mainly an argumentative device rather than a literal test (2019, sec. I.1). Practice seemed to vindicate the skeptics: in 2014 the chatbot Eugene Goostman, posing as a 13-year-old Ukrainian boy and changing the subject whenever a question got hard, fooled 33% of the judges at a competition held at the Royal Society; as M. Mitchell (2019) notes, the result revealed more about the judges and the test than about the machines (2019, ch. 3).
What LLMs did to the test
With large language models the test stopped being a horizon. In preregistered three-party, five-minute experiments, Jones & Bergen (2026) found that GPT-4.5, prompted to adopt a plausible human persona, was judged to be the human 73% of the time, more often than the real humans; LLaMa-3.1-405B, 56%; ELIZA and GPT-4o, 23% and 21%. By Turing's own criterion, the test was passed. The general reaction was a shrug, which half-confirms his other prediction: the use of words did change. Mahowald et al. (2024) diagnose the underlying confusion, the fallacy that whoever is good at language must be good at thinking. The public wager between Mitchell Kapor and Ray Kurzweil, with much stricter rules (three judges, two-hour interviews), will be settled before the end of 2029 (M. Mitchell, 2019, ch. 3).