3.1 The Turing testTuring (1950) proposed that instead of asking whether a computer could
think, we should see if we could program a machine in such a way that its
responses would fool interrogators into thinking they were dealing with a
man or woman. This test he dubbed ‘the Imitation Game’. Since then,
several programs have been developed which have made a pretty good stab
at passing the Turing test, at least with respect to question-and-answer
sessions which are not too free-wheeling (for example, Weizenbaum’s
ELIZA, Winograd’s SHRDLU, Colby’s PARRY; Schank and Abelson,
1977: references to such attempted ‘simulations of human thought’ are
legion). However, in spite of the ingenuity of the programmers, our in-
tuitive reaction is surely that posing the Turing test really does change the
question, because in order to engage in anything like human thinking it is
not suYcient toimitate the responsesa human being would make – one also
needs to imitate the processes by which such responses are produced.
Thus, for example, using a lot of computing power to sort through a vast
database of statistically sampled responses would be a sort of cheating,
even if the responses did seem fairly natural. And as soon as we learn that a
sophisticated chess-playing computer actually operates on abrute-force
algorithm, searching through thousands upon thousands of feeble vari-
ations which a human chess master would never bother to think about, we
realise that attributions of intentional states such as ‘wanting to avoid
weakening its pawn-structure’, ‘trying to keep its king safe’, ‘intending to
exploit the weakness on the light squares’, and so on, cannot be true in the
way that they might be true of a human player who had chosen the same
moves.
In order to demonstrate this sort of point, Copeland (1993) constructs
the following imaginary example of a computer which passes the Turing
test by pattern-matching. There are aWnite number of possible English
conversations consisting of less than, say, 10,000 sentences, each of which
consists of less than 200 words. Then imagine a computer in which all of
these conversations are listed in a vast look-up table. The computer
operates by matching a given input against its lists, and selecting one
possible continuation at random. Call this computer (which may well be
dependent upon computer technology far in advance of ours!)Superparry.
It seems plain that Superparry would pass the Turing test in connection
with any experimenter who did not actually know or suspect the details of
its program. For we may suppose that no normal human conversation
consists of more than 10,000 sentences, and that no normal human sen-
tence consists of more than 200 words. But it is plain, is it not, that we
would withhold mentality from Superparry as soon as wedidlearn that it
32 Folk-psychological commitments