Computing Machinery and Intelligence

• 1. Introduction

  ›

  In 1950, Alan Turing published a paper in the journal Mind that opened with three words that would echo through the following decades: "Can machines think?" Turing himself immediately acknowledged the problem. The words "think" and "machine" were so slippery, so laden with philosophical baggage, that the question was almost unanswerable as posed. So he did something elegant. He threw the question out and replaced it with a better one.

  What followed was "Computing Machinery and Intelligence," a paper that introduced a new kind of test to the general public. Not a laboratory experiment, not a mathematical proof, but a game. And the game's outcome, Turing argued, was the closest a person could get to determining whether a machine was truly intelligent.

  The paper ranged far beyond the game itself. Turing catalogued nine objections to the idea of machine thought, some theological, some mathematical, some philosophical. He considered whether machines could surprise us, whether they could learn, and whether the right way to build a thinking machine was to start with a child's mind rather than an adult's. He even gave extra-sensory perception a hearing.

  By the time readers reached the final pages, the question was no longer whether machines could think. It was how, and when, and what a machine would need to get there.

• 2. The Imitation Game

  ›

  Before Turing modified it, the Imitation Game was a simple party amusement. Three players took part: a man, a woman, and an interrogator who could be of either sex. The interrogator was kept separate, unable to see the other two, and communicated only through written notes. The goal was to determine, through questioning alone, which hidden player was the man and which was the woman. The man tried to mislead; the woman tried to help.

  Turing borrowed this structure and made one substitution. In place of the man, he put a computer. The new question was whether the interrogator would be fooled as often by a machine as by a human playing the deception role. If so, Turing argued, that result would carry the same weight as an affirmative answer to the original question about machine thought.

  The modified version placed three participants in isolated rooms: the computer under test, a human, and a human judge who typed into a terminal to converse with both. Both the computer and the human tried to convince the judge they were human. If the judge could not consistently tell them apart, the computer won.

  As the philosopher Stevan Harnad later framed it, Turing had quietly shifted the ground. The question was no longer whether a machine could think but whether a machine could act indistinguishably from a thinking entity. That reframing sidestepped the need to define "think" at all, and focused instead on observable performance. Turing had been turning this idea over since at least 1941, and one of the earliest known mentions of "computer intelligence" in any document came from him in 1947.

• 3. Why Digital Machines

  ›

  Turing paused before the game to ask which kind of machine was even worth considering. He dismissed the human clone as a man-made object that would prove nothing interesting. The interesting class, he argued, was digital machinery: machines that manipulate the binary digits 1 and 0, rewriting them into memory according to simple rules.

  His justification had two parts. First, digital machines already existed in 1950. There was no need to speculate about whether they were possible. Second, and more importantly, digital machinery was universal in a deep sense. Turing's own earlier research into the foundations of computation had established that a digital computer can, in theory, simulate the behaviour of any other digital machine, provided it has enough memory and time. This is the core insight behind what became known as the Church-Turing thesis and the universal Turing machine.

  The implication was sweeping. If any digital machine could act as though it were thinking, then every sufficiently powerful digital machine could do the same. Turing wrote that "all digital computers are in a sense equivalent." The question narrowed further: take one specific digital computer C, give it adequate storage, increase its speed, provide an appropriate programme, and ask whether C could play the deception role in the game. Turing was careful to clarify that the goal was not to show that all current machines would succeed, or even that any current machine could. The point was whether any imaginable machine could.

• 4. Nine Objections

  ›

  Having sharpened the question, Turing turned to the most persistent arguments against machine intelligence. He listed nine, and noted that they covered nearly all the serious objections that had been raised in the years following the paper's publication.

  The religious objection held that thought was a property of the immortal soul and could not belong to a machine. Turing answered that building such machines was no more a usurpation of divine power than the procreation of children; both were, in his words, "instruments of His will providing mansions for the souls that He creates."

  The "Heads in the Sand" objection was simpler: the consequences of thinking machines would be too frightening to accept. Turing identified this as a fallacy, confusing what should not exist with what can or cannot exist.

  The mathematical objection drew on results like Godel's incompleteness theorem to show that any logical system has limits. Turing replied that humans are frequently wrong themselves and that this could not serve as the decisive difference.

  The argument from consciousness came from Professor Geoffrey Jefferson, who delivered it in his 1949 Lister Oration, an acceptance speech for his 1948 Lister Medal. Jefferson argued that a machine could not truly compose a sonnet or concerto through genuine feeling. Turing replied that we cannot verify the inner experience of any other person either, and added that the mysteries of consciousness did not need to be resolved before answering the question of machine thought. A version of Jefferson's argument was later formalised in 1980 by philosopher John Searle in the Chinese room argument; Turing's reply became known as the "other minds reply."

  Ada Lovelace supplied what Turing called one of the most famous objections. She had written of the Analytical Engine that it "has no pretensions whatever to originate anything" and can only do what it is ordered to do. Turing recast this as the claim that computers "can never take us by surprise," and disputed it directly, arguing that machines can surprise humans whenever the consequences of different facts are not immediately obvious. He added that Lovelace was hampered by the scientific knowledge available in her time.

  The argument from continuity in the nervous system acknowledged that the brain is not purely digital; neurons have analog components in both the timing and probability of their firing. Turing accepted this, but argued that any analog system can be approximated to a reasonable degree of accuracy with enough computing power.

  Extra-sensory perception received an unexpected entry in the list. In 1950, ESP was still an active research topic. Turing admitted to what he described as "overwhelming statistical evidence" for telepathy, likely referring to experiments conducted in the early 1940s by Samuel Soal, a member of the Society for Psychical Research. Turing's solution was practical: design the test conditions so that mind-reading could not affect the outcome.

• 5. The Learning Machine

  ›

  The final section of the paper moved from defending the possibility of machine thought to asking how a machine might actually get there. Turing's answer was that it would have to learn.

  He returned to Lovelace's objection through a new analogy. He compared an unlearning machine to an atomic pile of subcritical size: an idea injected from outside causes a brief disturbance that fades away. A supercritical pile behaves differently. A single neutron entering it triggers a cascade that grows until the pile is destroyed. Turing asked whether the human mind worked like the supercritical pile, and concluded that a small proportion of human minds do: "An idea presented to such a mind may give rise to a whole 'theory' consisting of secondary, tertiary and more remote ideas." The question he left open was whether a machine could be made supercritical in the same way.

  He then proposed that the right approach was not to programme an adult mind but a child's. A child mind would be simpler to programme, like a newly bought notebook. It could then be educated. The process would involve reward and punishment to select desirable patterns, which Turing explicitly compared to natural selection: the structure of the child machine corresponded to hereditary material, changes to it corresponded to mutations, and the experimenter's judgment corresponded to natural selection.

  Turing also identified a feature of learning machines that made them different from conventional computers: the teacher would not know the machine's internal state during learning. Intelligent behaviour, he argued, involved a departure from complete determinism, so long as that departure did not degenerate into pointless loops or mere randomness. Introducing a deliberate element of randomness could even be valuable, particularly when multiple solutions were available or when a systematic search would be inefficient.

  Turing closed by speculating that by the end of the century it would be technically possible to programme a machine to play the imitation game. He suggested two starting points: abstract tasks like chess, and, separately, teaching a machine to understand and speak English. The chess path was later taken by Deep Blue, a computer developed by IBM that defeated world champion Garry Kasparov, though Turing could not have foreseen that specific outcome when he wrote in 1950.