Can Machines Think
"I propose to consider the question, Can machines think?" is the first sentence of Alan Turing's 1950 paper "Computing Machinery and Intelligence," and the question it poses has outlived every attempt to settle it. The sentence is as much a maneuver as a question: Turing immediately declined to answer it as worded, on the grounds that "machine" and "think" are not defined clearly enough, and replaced it with a question about behavior in a game. The entry is therefore about the question, why the replacement was made and how philosophers have fought over it since, and not about the paper's structure or the game's rules.
Turing's objection to the original wording was about definitions, not about the answer. If the meanings of "machine" and "think" were taken from ordinary usage, he remarked, the answer would have to be sought in something like a statistical survey of public opinion, which he called absurd. His alternative was to trade the vague question for a closely related one in unambiguous words, namely whether a machine could do well at the imitation game. This is an operational move, in the tradition of defining a concept by the procedure used to detect it, as physicists define length by measurement. The trade also had a deliberate rhetorical function: it moved the argument away from the soul, introspection and the ownership of words toward what could be checked from the outside.
The behaviorist flavor of the replacement is both its strength and its target. Behaviorism in psychology and philosophy treated mental terms as shorthand for observable behavior and dispositions to behave, and Gilbert Ryle's 1949 book The Concept of Mind attacked the picture of a ghost inside the machine in a similar spirit. Turing's test inherits that attitude: if a system cannot be told apart from a thinker by its conversational conduct, there is nothing further to demand. Critics reply that two systems can behave alike and differ in what lies behind the behavior. Ned Block's 1981 paper "Psychologism and Behaviorism" imagined a machine that merely looks up pre-stored responses for every possible conversation, which would pass a behavioral test with no intelligence in the ordinary sense at all.
John Searle's 1980 paper "Minds, Brains, and Programs" in Behavioral and Brain Sciences is the best-known attack. His Chinese room imagines a person who does not understand Chinese following a rulebook to manipulate Chinese symbols so well that outsiders believe a fluent speaker is inside. Searle argued that the person, and therefore any program, has syntax without semantics, so running the right program is not sufficient for understanding. Replies were immediate and are still standard: the systems reply says the whole system, not the person, understands; the robot reply says grounding in perception and action would matter; and others accept the argument against symbol shuffling while denying that it touches learned, embodied systems.
Daniel Dennett has been among the most prominent defenders of the behavioral standard. In an essay with the title of this entry from the mid-1980s he argued that the test is demanding if the interrogator is allowed to probe freely, because ordinary conversation requires vast background knowledge, and that anything passing it would deserve to be called a thinker. His wider theory of the intentional stance holds that treating a system as having beliefs and desires is justified when doing so predicts its behavior best, so that thinking is a matter of pattern, not of a special substance. Turing himself, in a 1951 radio talk and a 1952 BBC panel discussion with Max Newman, Richard Braithwaite and Geoffrey Jefferson, kept returning to the point that one should judge by how a machine acts, and that attitudes toward machines would shift with exposure.
The question has not been closed, only recast. Large language models produce fluent text that can fool casual judges while failing in odd ways, and every side claims the evidence: believers cite the fluency, skeptics cite the lack of grounding, and many researchers now prefer concrete questions about specific capabilities, such as reasoning, planning or reliability, over the single question Turing posed. Hubert Dreyfus's 1972 What Computers Can't Do and Roger Penrose's 1989 The Emperor's New Mind argued from phenomenology and physics that computation falls short, which shows the range of positions the opening sentence still sets off.
Two misunderstandings recur. Turing did not claim that machines are conscious, and he said explicitly that there is no need to resolve all mysteries about consciousness before answering the replacement question. And replacing a question is not the same as dissolving it: the Searle-Dennett disagreement is precisely about whether the operational substitute preserves the meaning of the original.
Read "Can machines think?" as a question whose wording Turing refused to accept: he swapped it for one that experiments can answer, and philosophers have argued ever since over whether the swap kept anything that mattered.