dvm@yale.UUCP (Drew Mcdermott) (03/17/89)
Here is a paper I wrote a few years ago and never did much with,
which seems relevant to recent discussions. Excuse the TeXisms.
-- Drew
The Red Herring Turing Test
Drew McDermott
Yale University
In 1950, Alan Turing (in [<bibref>])
proposed his famous test for whether a machine
is intelligent, namely, to see whether it could, when
interrogated over a teletype, mimic a human so well as to be indistinguishable
from the real thing.\footnote{$^1$}{Turing
actually proposed something different, but this version makes more sense,
so it is what people usually mean by ``Turing test.''}
Since then, this test has become deeply embedded in the thought processes
of cognitive scientists, to the point where many take for granted that
it is an important criterion for whether a system is intelligent.
Turing invented this test as a way of nudging people into entertaining
an implausible and shocking idea, that machines could be intelligent.
He wanted to duck the question, What is intelligence? What is often
overlooked is that, if cognitive science is really ever to become a
science, we do not have the option of ducking this question; we must
answer it. Hence Turing's little scenario has become totally
irrelevant to serious discussion about what thinking is. No one seems
to realize this but me, hence this paper.
What confuses most people is that they mistake Turing's attempt to avoid
the question for an attempt to answer it. But anyone who believes that
Turing's test is an interesting test for intelligence is guilty of
behaviorism, not a crime in itself, but shameful in anyone who believes
in cognitive science, the antithesis of behaviorism. Of course, it
is probably true that a system that could fool a trained panel of
experts into believing it intelligent would in fact be intelligent,
but it is blatant waste of experts' time to have them sit on such
panels, when they should be inquiring about how minds actually {\it work}.
Compare the following hypothetical case: Human explorers land on a
planet whose inhabitants are somewhat technologically backward. The
locals are impressed by human gadgets, especially radio. They decide
to try and understand it, so they rustle up some philosophers in order
first to arrive at a criterion for something's being a radio. Their
first cut is that a radio is a device that emits sounds whenever
similar sounds are made in the control room of the earthlings'
spaceship. But others object that this criterion does not rule out
ordinary telephony, so the criterion is modified. Perhaps they arrive
at something like, ``A radio is a device that emits sounds similar to
those made in the earthlings' spaceship while suspended from the
ceiling by a nonconducting string.''
This is all amusing, but a waste of time if the aliens really want to
understand radio. No one needs an ironclad behavioral criterion for
``radiohood,'' assuming that there are plenty of indisputably genuine
radios around to study. Such a study might eventually lead to a deeper
definition of radio as ``A receiver of signals encoded as modulated
electromagnetic waves,'' but by the time the definition was available
it would be relatively unimportant, when stacked up against the theory
of electromagnetism.
Similarly with intelligence. If we ever have a theory that explains
it, we will no longer care about distinguishing bogus understanding
from the real thing. We will have a rich theory based on concepts we
can now barely imagine, just as radio is based on something as unlikely
as invisible electromagnetic waves.
I point all this out because it is often assumed that to seek
computational theories of thought, the kind AI researchers look for,
requires adopting the Turing Test as a criterion for intelligence, or
even the only criterion. Often in the midst of a discussion about
such issues, I find the Turing Test suddenly glued to me like fly
paper, and I am unable to shake it off. It is not as if I have been
forced to adopt it by the logic of my position, but that it has been
handed to me casually, almost as a favor by whoever I am arguing with.
``You espouse computationalism? Then you won't want to be without
this handy accessory, Turing's Test.'' Flail as I may to get rid of
this unwelcome visitor, it stays glued, and the discussion
degenerates.
A particularly bothersome case is Searle's Chinese Room argument
[<bibref>], in which a non-Chinese-speaking person is hired to execute
an algorithm that supposedly understands Chinese when executed by a
digital computer. The person wouldn't thereby come to understand
Chinese, but there's no essential difference between what the person
does and what a computer would do, so there would be no reason to say
the computer really understood Chinese either. QED
I have always maintained [<bibref>] that the solution to this riddle
is to see {\it two} understanders when an understanding person executes
an understanding algorithm. It's quite simple, really:
Computer executing algorithm: 0+1=1 understander
Person executing algorithm: 1+1=2 understanders
The fact that these two understanders occupy the same body, and the
way the two relate, should make us smile, not choke.
The Searly comeback is this: ``Come now, it's preposterous to imagine
two souls inside one body like this. One of those understanders is
there (and knows he's there), but the other is sanctioned only by the
Turing Test, which you computationalists insist so much on. Without
this behavioral test, what gives you the right to call it an
understander at all?'' His actual words [<pageref>] are these (he has
cast himself in the role of understander, hence the first-person
pronouns):
``What {\it independent} grounds are there supposed to be
for saying that the agent must have a subsystem within him which literally
understands stories in Chinese? As far as I can tell the only grounds
are that in the example I have the same input and output as
native Chinese speakers and a program that goes from one to the other.
... The only motivation for saying there {\it must} be a subsystem in
me which understands Chinese is that I have a program and I can pass
the Turing test, I can fool native Chinese speakers. But precisely
one of the points at issue is the adequacy of the Turing test.''
The fallacy here (one is tempted to call it sophistry) is an
unjustified revision of the initial assumption halfway through a {\it
reductio ad absurdum}. The argument starts by assuming there exists
an algorithm $A$ which understands Chinese. No such algorithm has
been written, and if none ever is the whole issue will be moot. But
that's okay --- the argument is designed to reduce someone's position
to an absurdity, so we should start by assuming that someone's
position is true. Whose position? --- the computationalists'. But if
AI ever succeeds in producing algorithm $A$, presumably it will be due
to the discovery of a nontrivial theory of understanding Chinese. It
is {\it this theory}, not Turing's Test, that will say whether an
entity understands or does not understand Chinese. One consequence of
this theory is that any entity that executes $A$ will understand
Chinese. Searle may feel that this idea is absurd, but he is not
allowed to cite that feeling at this point. One cannot get away with
a {\it reductio ad absurdum} of the form:
``Suppose my opponent's position were true. But that
would be absurd.''
The argument has to cover some ground, and has to hit a conclusion
that everyone agrees is absurd. Since Searle's argument doesn't do
this, he is forced to drag in Turing's Test, and pronounce {\it it}
absurd. But it is crucial to realize that {\it he} brought Turing's
Test in; it didn't come with the position he is attacking. In other
words, it is completely false that ``one of the points at issue is the
adequacy of the Turing test.'' When he asks for ``independent
grounds'' for believing there are two understanders, he overlooks
where they come from: the theory behind $A$.\footnote{$^2$}{By the
way, his use of the term ``subsystem'' is an effort at disinformation.
No such subsystem is proposed by his opponents; rather, the entire
system embodies two understanders.}
Let me dramatize this rebuttal by returning to the radio-free planet.
Suppose that some budding electrical engineer proposes that diodes have
something to do with radio reception, and proposes a simple circuit to
demodulate radio waves. An alien Searle might respond thus: ``If this
proposal were correct, we could make a radio by having someone hold
two halves of a diode in each hand, while electricity was conducted
through his gut. Then, since we could suspend him from the ceiling
by a nonconducting rope, your theory would have the consequence that
this person would be a radio! This is ridiculous on its face, not
to mention the consequence that his gut, which is manifestly a
digesting organ, would also have to be a conducting organ as well,
which doesn't seem possible.''
This argument is not in detail analogous to Searle's Chinese Room
argument, but it does show two similar features:
\item{1} Part of its force comes from its painting a ludicrous
picture. In this case it's a picture of a person suspended from the
ceiling and receiving radio waves. In the original argument, it's a
picture of a person apparently understanding Chinese without knowing
it. In the case of the radio, we are not fooled, since we know that
you really could build a radio this way. In the case of the
understanding algorithm, where we don't yet know if it would work,
mere silliness should still carry little weight.
\item{2} The rest of the argument's force comes from its invocation of
a pointless behavioral criterion. The electrical engineer doesn't
need the definition of a radio as a device that emits sounds when
suspended from the ceiling. If he didn't lack philosophical street
smarts, he would insist that this crazy criterion be thrown out. In
the original argument, there is no need at all for AI researchers to
defend Turing's Test, but philosophers often succeed via
presupposition in convincing us that we have some stake in it.
\noindent When these two features are eliminated, neither argument has any
force at all.
Searle ends up attributing to his opponents the premise that a
computational theory of mind exists and has been confirmed by Turing's
Test. But I don't want that premise; I want what Searle calls
``strong AI,'' the theory that any computational device that simulates
thoughts and emotions in the right way would {\it have} those thoughts
and emotions. I don't want this premise ``as confirmed in a certain
way''; I just want the premise, plain and simple. As Searle says at
the outset [<pageref>]:
One way to test any theory of the mind is to ask oneself
that would it be like if my mind actually worked on the principles
that the theory says all minds work on.
\noindent But if strong AI is actually true, then perforce any
creature, including our deluded human CPU, will indeed exhibit
understanding when he runs algorithm $A$. If strong AI is true, it
will not matter whether it has been confirmed or even noticed by the
human race. If it's true, then any system executing $A$ will
understand, and that's that.
I realize that some may find my counterargument perplexingly vacuous.
Nothing I have said here provides the slightest evidence that anything
like an AI account of mind will prove correct. But my goal was only
to knock Searle's argument down, not set up one of my own. In order
to refute a {\it reductio ad absurdum} argument, it is not necessary
to find reasons for believing the starting assumption after all; it is
only necessary to find a flaw in the route to the alleged
contradiction.
In fact, we know almost nothing about what a computational theory of
mind would look like. Why should this bother us, seeing as how Searle
is willing to reveal almost nothing about his alternative ``causal''
theory? My only point is that, given the anemic state of our
theorizing, it is ridiculous to demand that we provide in advance any
empirical test for the presence of understanding, intelligence,
consciousness, or any other phenomenon. By the time a theory of these
concepts is available, they will no doubt have all been revised beyond
recognition anyway. (Suppose that early physicists had been required
to provide in advance a test for whether their concept of energy
matched the phenomenon exhibited by a person ``bursting with
energy.'') Hence, we do not have to embrace Turing's Test, or provide
any substitute, at this point in our efforts. In fact, there is no
reason for anyone to be drawn into an argument about the a priori
plausibility of AI, except for the fascination and strongly
conflicting intuitions that the subject affords.
If a computational theory of mind is ever found, Turing's Test will
play no role in it. For one thing, it can never hope to provide a
necessary condition for intelligence, but only a sufficient one.
Presumably a full-blooded theory will allow us to locate many
different degrees and types of intelligence in all sorts of entities
that could never pass Turing's Test. Contrariwise, although it is
implausible that something could pass Turing's Test and not be
intelligent according to such a theory, it is possible that this is
so. (For instance, people might have some blind spot which caused
them to be inevitably gullible about certain patterns of behavior
which a machine could duplicate without any real intelligence.) But
even if passing the Turing Test does turn out to be a sufficient
condition for intelligence, it will still be the theory of
intelligence that will matter, not the test.reiter@babbage.harvard.edu (Ehud Reiter) (03/20/89)
As a followup to Drew McDermott's excellent article on the Turing test
and the Chinese Room problem, let me add one small note. There is
a certain mindset that equates intelligence with whatever humans do.
Now, we know that there are plenty of types of reasoning which humans
are pretty bad at. Multiplying large numbers is an obvious case.
More interesting perhaps are
- probabilistic reasoning. Kahneman and Tversky have shown that
people make fundamental mistakes, such as ignoring priors and assuming
that p(A&B) can be greater than p(A).
- predictive tasks. A long literature, dating back to Paul Meehl,
shows that statistical techniques usually out-perform expert human judgement,
provided that the data is quantifiable.
(I sometimes wonder what the expert system people have to say about Meehl's
findings. If a simple linear regression can do a better job than a human
expert, why bother building a computer system that attempts to emulate
human judgements?).
(see JUDGEMENT UNDER UNCERTAINTY: HEURISTICS AND BIASES, edited by D.
Kahneman, P. Slovic, and A. Tversky. Especially chapters 1 and 28).
A machine that could pass the Turing test would have to be programmed to do
as badly as humans at multiplying, probabilistic reasoning, and predictive
tasks. But is it really critical to the definition of intelligence that, say,
an entity ignore prior probabilities when making probabilistic judgements?
I doubt it, and suggest that finding out how to do a good job on the above
reasoning tasks is more important than finding out how to replicate the
mistakes humans make.
Ehud Reiter
reiter@harvard (ARPA,BITNET,UUCP)
reiter@harvard.harvard.EDU (new ARPA)mayoung@bnr-di.UUCP (Mark Young) (03/21/89)
In article <1441@husc6.harvard.edu>, reiter@babbage.harvard.edu (Ehud Reiter) writes: > A machine that could pass the Turing test would have to be programmed to do > as badly as humans at multiplying, probabilistic reasoning, and predictive > tasks. This is not strictly true. The machine would only have to be able to emulate this behaviour. Aftir all, I can imatate a guy who cant spel to good, even if my own spelling is (reasonably) good. Remember that the point of the TT was for the machine to _fool_ the observer into thinking it was a person. If the machine feels that giving the wrong answer to a multiplication problem will help it in this task, it will give the wrong answer. > > Ehud Reiter Mark Young