Computing Machinery and Intelligence

Defining them by common usage would make the answer a matter for a statistical survey like a Gallup poll, which he called absurd. Instead he replaced the question with a closely related one in relatively unambiguous words.

Three people: a man (A), a woman (B) and an interrogator (C) in a separate room, who knows them only as X and Y and must decide which is the woman. A tries to make C identify wrongly, B tries to help. Answers are typewritten (ideally by teleprinter) so voices give nothing away.

What happens when a machine takes the part of A: will the interrogator decide wrongly as often as when the game is played between a man and a woman?

It draws a fairly sharp line between physical and intellectual capacities: there’s no point making a thinking machine more human with artificial skin, a machine isn’t penalised for not shining in beauty competitions, nor a man for losing a race against an aeroplane.

That it is weighted against the machine: a machine might do something that ought to be called thinking but is very different from what a man does. His answer: if a machine nevertheless plays the game satisfactorily, this objection needn’t trouble us.

It comes after a pause of about 30 seconds and is wrong: 105621 instead of 105721. That fits Turing’s later point that a machine playing the game would hide its speed and deliberately make arithmetic mistakes, since deadly accuracy would unmask it.

Only digital computers. He wanted to allow any engineering technique, even ones whose workings their builders can’t explain, but to exclude humans born in the usual way; restricting to digital computers does this, and current interest in “thinking machines” was aroused by them anyway.

The question isn’t whether present computers would do well, but whether there are imaginable computers that would.

As a machine meant to carry out any operation a human computer could do: someone following fixed rules from a book without deviating, with unlimited paper for calculations.

The store (the human computer’s paper and memory, including the instruction table), the executive unit (performs individual operations) and the control (makes sure instructions are obeyed correctly and in order).

They let a sequence of operations be repeated until some condition holds, obeying the same instructions over and over rather than fresh ones each time. (Turing’s analogy: a notice in the hall telling Tommy to check on the shoes each morning, instead of asking him afresh.)

Constructing instruction tables: to programme a machine to do A means putting the appropriate instruction table into it. To mimic a human computer one asks him how the operation is done and translates the answer into an instruction table.

Babbage’s Analytical Engine was purely mechanical, and all digital computers are in a sense equivalent, so electricity can’t be of theoretical importance; it just appears wherever fast signalling is needed. Look for mathematical analogies of function instead.

A machine that moves by sudden jumps from one definite state to another, the states being different enough that confusion between them can be ignored. Strictly, nothing physical is like this, but many machines can usefully be thought of so.

	q1	q2	q3
input i0 (lever off)	q2	q3	q1
input i1 (lever on)	q1	q2	q3
output	o0	o0	o1 (lamp)

Turing’s example: a wheel clicking round through three positions, stopped by a lever, with a lamp lit in one position.

In the physical universe tiny errors in initial conditions can have overwhelming later effects (one displaced electron may decide whether a man is killed by an avalanche a year later). In a discrete-state machine this doesn’t happen: reasonably accurate knowledge of the state yields reasonably accurate prediction any number of steps later.

The base-2 logarithm of its number of states, so capacities of combined machines add. The three-state wheel has about 1.6.

Given the table of any discrete-state machine, a digital computer with enough storage and speed can be programmed to mimic it, so the interrogator couldn’t tell them apart. Hence, speed aside, all digital computers are in a sense equivalent, and one machine suitably programmed replaces designing new ones.

Whether one particular digital computer, given adequate storage, greater speed and an appropriate programme, can be made to play the part of A satisfactorily, with a man as B.

That computers with a storage capacity of about \(10^9\) could be programmed to play the imitation game so well that an average interrogator would have no more than a 70% chance of a correct identification after five minutes of questioning, and that by the end of the century educated opinion would allow talk of machines thinking without contradiction.

Scientists don’t proceed inexorably from established fact to established fact; conjectures are of great importance because they suggest useful lines of research, and no harm results as long as it is clear which statements are proved facts and which conjectures.

(1) Theological, (2) “heads in the sand”, (3) mathematical, (4) from consciousness, (5) from various disabilities, (6) Lady Lovelace’s, (7) from continuity in the nervous system, (8) from informality of behaviour, (9) from extrasensory perception.

In its own terms: it restricts God’s omnipotence, since He could confer a soul on an elephant with a suitable brain, or on a machine; building such machines no more usurps His power than having children does. He added that theological arguments have often failed before, as against Copernicus.

“The consequences of machines thinking would be too dreadful; let us hope they cannot.” It reflects a wish to believe Man is superior, strongest among intellectuals. Turing found it too insubstantial to refute: “Consolation would be more appropriate.”

Results like Gödel’s theorem and Turing’s own (1937) show any given machine of this kind must answer some questions wrongly or not at all, e.g. “Will this machine ever answer ‘Yes’ to any question?” about a machine related to itself, a disability the human intellect is claimed not to share.

It is proved that any particular machine has limitations, but only asserted that humans have none. We often give wrong answers ourselves, and a triumph over one machine says nothing about all machines: there might be men cleverer than any given machine, but other machines cleverer again.

(Jefferson) No machine equals a brain until it writes a sonnet because of thoughts and emotions felt, and knows it has. Taken to the extreme, the only way to know a machine (or a man) thinks is to be it: solipsism. Rather than accept that, we adopt the polite convention that everyone thinks, so critics would probably accept the test.

An oral examination used to find out whether someone really understands something or has learnt it “parrot fashion”, essentially the imitation game without player B. His sample viva about a sonnet (“Shall I compare thee to a summer’s day”) shows the kind of sustained answers that would be hard to dismiss as “artificially signalling”.

Claims that machines will never be able to do some X: be kind, have a sense of humour, fall in love, learn from experience, use words properly, do something really new… He saw them as unconscious scientific induction from the small, special-purpose machines people had seen, whose limitations mostly come from very small storage capacity.

Errors of functioning come from mechanical or electrical faults; abstract machines can’t have them. Errors of conclusion arise when meaning is attached to the output and the machine outputs a falsehood, which nothing prevents. A machine playing the game would make deliberate arithmetic mistakes to avoid being unmasked by its accuracy.

In the sense that its operations can concern itself: helping to write its own programmes, predicting the effects of changes to its structure, or modifying its programmes by observing the results of its behaviour.

Of Babbage’s Analytical Engine: “The Analytical Engine has no pretensions to originate anything. It can do whatever we know how to order it to perform.”

The evidence available to her didn’t show the property, not that machines lack it; a universal machine could mimic any machine that has it. To the variant “a machine can never take us by surprise”: machines surprise him often, and assuming all consequences of known facts are seen at once is a fallacy.

The nervous system isn’t a discrete-state machine: small differences in an impulse can matter a lot. But under the game’s conditions the interrogator can’t exploit this: a digital computer imitating a continuous differential analyser asked for π could answer 3.12–3.16 with suitable probabilities and be indistinguishable.

No set of rules can cover every circumstance, so men can’t be machines. Turing separated rules of conduct (precepts one can follow, like stopping at red lights) from laws of behaviour (laws of nature applied to the body); we can never be sure there are no complete laws of behaviour, and even a small program can be impossible to predict from its replies.

He believed the statistical evidence for telepathy was overwhelming, which would threaten the idea that bodies follow known physical laws. A telepathic human could beat a machine at guessing cards, so the test would need tightening: put the competitors in a “telepathy-proof room”.

Atomic pile: most minds (and machines) are subcritical, an injected idea dies away, but some are supercritical, one idea sets off a whole theory; can a machine be made supercritical? Onion skin: strip off the mechanically explicable functions of the mind and see whether a “real” mind remains. Turing called both “recitations tending to produce belief”, not proofs.

He estimated the brain at \(10^{10}\)–\(10^{15}\) binary digits and expected about \(10^9\) to suffice, practicable with foreseeable engineering. But at a thousand digits of programme a day, about sixty workers would need fifty years, so a more expeditious method was needed.

Simulate a child’s mind and educate it. The child brain is like a notebook from the stationer’s: rather little mechanism and lots of blank sheets. The problem splits into the child programme and the education process.

structure of the child machine	hereditary material
changes of the child machine	mutation
judgment of the experimenter	natural selection

It should be faster than evolution, since the experimenter can use intelligence and isn’t limited to random mutations.

Events shortly before a punishment become less likely to repeat, and those before a reward more likely: an early statement of reinforcement learning. But rewards alone carry little information, so “unemotional” channels are needed too, e.g. teaching the machine by rewards to obey orders in a symbolic language.

The rules that change during learning are less pretentious ones, with only ephemeral validity, like amendments under a fixed constitution.

Unlike ordinary computation, where the programmer tracks the machine’s state at every step, a machine educated from a well-tried child programme will have internals its teacher can’t follow, though he may still predict its behaviour. This contradicts the claim that machines only do what we know how to order them to do.

Learning is a search for behaviour that satisfies the teacher, and with many satisfactory solutions random search avoids being stuck in a large region without solutions and needs no record of what has been tried, as in evolution.

He didn’t know: either a very abstract activity like chess, or giving the machine the best sense organs money can buy and teaching it to understand and speak English like a child. Both should be tried.