Stevan
Harnad
In 1980, the philosopher John Searle published in the journal Behavioral
and Brain Sciences a simple thought experiment that he called the "Chinese Room
Argument" against "Strong Artificial Intelligence (AI)" The
thesis of Strong AI has since come to be called "computationalism,"
according to which cognition is just computation, hence
mental states are just computational states:
Computationalism. According to computationalism, to
explain how the mind works, cognitive science needs to find out what the right
computations are -- the ones that the brain performs in order to generate the
mind and its capacities. Once we know that, then every system that performs
those computations will have those mental states: Every computer that runs the
mind's program will have a mind, because computation is hardware-independent: Any
hardware that is running the right program has the right computational states.
The Turing Test. How do we know which program is
the right program? Although it is not strictly a tenet of computationalism, an
answer that many computationalists will agree to is that the right program will
be the one that can pass the Turing Test (TT), which is to be a system that is
able to interact by email with real people exactly the way real people do -- so
exactly that no person can ever tell that the computer program is not another
real person. Turing (1950) had suggested that once a computer can do everything
a real person can do so well that we cannot even tell them apart, it would be
arbitrary to deny that that computer has a mind, that it is intelligent, that
it can understand just as a real person can.
This, then, is the thesis that Searle set out to show was wrong:
(1) Mental states are just computational states, (2) the right computational
states are the ones that can pass the TT, and (3) any and every hardware on
which you run those computational states will have those mental states too.
Hardware-Independence. SearleÕs thought experiment was
extremely simple. Normally, there is no way I can tell whether anyone or
anything other than myself has mental states. The only mental states we can be
sure about are our own. We canÕt be someone else, to check whether they have
mental states too. But computationalism has an important vulnerability in this
regard: hardware-independence. Since any and every dynamical system (i.e., any
physical hardware) that is executing the right computer program would have to
have the right mental states, Searle himself can execute the computer program,
thereby himself becoming the hardware, and then check whether he has the right
mental states. In particular, Searle asks whether the computer that passes the
TT really understands the emails it is receiving and sending.
The Chinese Room. To test this, Searle obviously
cannot conduct the TT in English, for he already understands English. So in his
thought-experiment the TT is conducted in Chinese: The (hypothetical) computer
program he is testing in his thought-experiment is able to pass the TT in
Chinese. That means it is able to receive and send email in Chinese in such a
way that none of its (real) Chinese pen-pals would ever suspect that they were
not communicating with a real Chinese-speaking and Chinese-understanding
person. (We are to imagine the email exchanges going on as frequently we like,
with as many people as we like, as
long as we like, even for an entire lifetime. The TT is not just a short-term
trick.)
Symbol-Manipulation. In the original version of
SearleÕs Chinese Room Argument he imagined himself in the Chinese Room,
receiving the Chinese emails (a long string of Chinese symbols, completely
unintelligible to Searle). He would then consult the TT-passing computer program,
in the form of rules written (in English) on the wall of the room, explaining
to Searle exactly how he should manipulate the symbols, based on the incoming
email, to generate the outgoing email. It is important to understand that
computation is just rule-based symbol-manipulation, and that
the manipulation and matching is done purely on the basis of the shape of the
symbols, not on the basis of their meaning.
Now the gist of SearleÕs argument is very simple: In doing all
that, he would be doing exactly the same thing any other piece of hardware
executing that TT-passing program was doing: rule-fully manipulating the input
symbols on the basis of their shapes, and generating output symbols that make
sense to a Chinese pen-pal -- the kind of email reply a real pen-pal would
send, a pen-pal that had understood the email received, as well as the email
sent.
Understanding. But Searle goes on to point out
that in executing the program he himself would not be understanding the emails
at all! He would just be manipulating meaningless symbols, on the basis of
their shapes, according to the rules on the wall. Therefore, because of the
hardware-independence of computation, if Searle would not be understanding
Chinese under those conditions, neither would any other piece of hardware
executing that Chinese TT-passing program. So much for computationalism and the
theory that cognition is just computation.
The System Reply. Searle correctly anticipated that
his computationalist critics would not be happy with the handwriting on the
wall: Their ÒSystem ReplyÓ would be that Searle was only part of the
TT-passing system. That whereas Searle would not be understanding Chinese under
those conditions, the system as a whole would be!
Searle rightly replied that he found it hard to believe that he
plus the walls together could constitute a mental state, but, playing the game,
he added: Then forget about the walls and the room. Imagine that I have
memorized all the symbol manipulation rules and can conduct them from memory.
Then the whole system is me: WhereÕs the understanding?
Desperate
computationalists were still ready to argue that somewhere in there, inside
Searle, under those conditions, there would lurk a Chinese-understanding of
which Searle himself was unaware, as in multiple personality disorder -- but
this seems even more far-fetched than the idea that a person plus walls has a
joint mental state of which the person is unaware.
Brain Power. So the Chinese Room Argument is right, such as
it is, and computationalism is wrong. But if cognition is not just computation,
what is it then? Here Searle is not much help, for he first overstates what his
argument has shown, concluding that it has shown (i) that cognition is not
computation at all Ð whereas all it has shown is that cognition is not all
computation. Searle also concludes that his argument has shown (ii) that the
Turing Test is invalid, whereas all it has shown is that the TT would be
invalid if it could be passed by a purely computational system. His only positive
recommendation is to turn brain-ward, trying to understand the causal powers of
the brain instead of the computational powers of computers.
But it is not yet apparent what the relevant causal powers of the
brain are, nor how to discover them. The TT itself is a potential guide: Surely
the relevant causal power of the brain is its power to pass the TT! We know now
(thanks to the Chinese Room Argument) that if a system could pass the TT via
computation alone, that would not be enough. What would be missing?
The Robot Reply. One of the attempted refutations
of the Chinese Room Argument Ð the ÒRobot ReplyÓ Ð contained the seeds of an
answer, but they were sown in the wrong soil. A robotÕs sensors and effectors
were invoked in order to strengthen the System Reply: It is not Searle plus the
walls of the Chinese Room that constitutes the Chinese-understanding ÒsystemÓ,
it is Searle plus a robotÕs sensors and effectors. Searle rightly points out
that it would still be him doing all the computations, and it was the
computations that were on trial in the Chinese Room. But perhaps the TT itself
needs to be looked at more closely here:
Behavioral Capacity. TuringÕs original Test was indeed
the email version of the TT. But there is nothing in TuringÕs paper or his arguments
on behalf the TT to suggest that it should be restricted to candidates that are
just computers, or even that it should be restricted to email! The power of the
TT is the argument that if the candidate can do everything a
real person can do Ð and do it indistinguishably from the way real people does
it, as judged by real people Ð then it would be mere prejudice to conclude that
it lacked mental states when we were told it was a machine. We donÕt even
really know what a machine is, or isnÕt!
But we do know that real people can do a lot more than
just email to one another. They can see, touch, name, manipulate and describe
most of the things they talk about in their email. Indeed, it is hard to
imagine how either a real pen-pal or any designer of a TT-passing computer
program could deal intelligibly with all the symbols in an email message
without also being able to do at least some of the things we can all do with
the objects and events in the world that those symbols stand for.
Sensorimotor Grounding of Symbols.
Computation, as noted, is symbol-manipulation, by rules based on the symbolsÕ
shapes, not their meanings. Computation, like language itself, is universal,
and perhaps all-powerful (in that it can encode just about anything). But
surely if we want the ability to understand the
symbolsÕ meanings to be among the mental states of the TT-passing system, this
calls for more than just the symbols and the ability to manipulate them. Some,
at least, of those symbols must be ÒgroundedÓ in something other than just more
meaningless symbols and symbol-manipulations Ð otherwise the system is in the
same situation as someone trying to look up the meaning of a word in a language
Ð letÕs say, Chinese -- that he does not understandÉ in a Chinese-Chinese dictionary!
Emailing the definitions of the words would be intelligible enough to a pen-pal
who already understood Chinese, but they would be of no use to anyone or
anything that did not understand Chinese. Some of the symbols must be grounded
in the capacity to recognize and manipulate the things in the world that the
symbols refer to.
Mind-Reading. So the TT candidate must be a
robot, able to interact with the world that the symbols are about -- including
us -- directly, not just via email. And it must be able to do so
indistinguishably from the way any of the rest of us interact with the world or
with one another. That is the gist of the TT. The reason Turing originally
formulated his test in its pen-pal form was so that we would not be biased by
the candidateÕs appearance. But in todayÕs cinematic sci-fi world we have, if
anything, been primed to be over-credulous about robots, so much more ÒcapableÓ
are our familiar fictional on-screen cyborgs than any TT candidate yet designed
in a cog-sci lab. In real life our subtle and biologically based Òmind-readingÓ
skills (Frith & Frith 1999) will be all we need once cog-sci starts to
catch up with sci-fi and we can begin T-Testing in earnest.
The
Other-Minds Problem.
Could the Chinese Room Argument be resurrected to debunk a TT-passing robot?
Certainly not. For SearleÕs argument depended crucially on the
hardware-independence of computation. That was what allowed Searle to ÒbecomeÓ
the candidate and then report back to us (truthfully) that we were mistaken if
we thought he understood Chinese. But we cannot ÒbecomeÓ the TT-passing robot,
to check whether it really understands, any more than we can become another
person. It is this parity (between other people and other robots) that is at
the heart of the TT. And anyone who thinks this is not an exacting enough test
of having a mind need only remind himself that the Blind Watchmaker (Darwinian
evolution), our Ònatural designer,Ó is no more capable of mind-reading than any
of the rest of us is. That leaves only the robot to know for sure whether or
not it really understands.
REFERENCES
Anderson
D., & Copeland B.J.
(2002) Artificial Life and the Chinese Room
Argument Artificial Life
8(4): 371-378
Brown S. (2002) Peirce, Searle, and the Chinese Room Argument. Cybernetics
& Human Knowing 9(1) 23-38
Dyer, M. G. (1990) Intentionality and computationalism:
minds, machines, Searle and Harnad. Journal of Experimental and Theoretical
Artificial Intelligence 2(4): 303-319.
French, R. (2000) The Turing Test: The First Fifty Years. Trends
in Cognitive Sciences 4(3): 115-121.
Frith Christopher D. & Frith, Uta (1999) Interacting minds --
a biological basis. Science 286: 1692Ð1695. http://pubpages.unh.edu/~jel/seminar/Frith_mind.pdf
Harnad, Stevan (1989) Minds, Machines and Searle. Journal of
Theoretical and Experimental Artificial Intelligence 1: 5-25. http://cogprints.org/1573/
Harnad, Stevan (1990) The Symbol Grounding Problem. Physica D 42:pp.
335-346. http://cogprints.org/3106/
Harnad, S. (2001) What's Wrong and Right About Searle's Chinese
Room Argument? In: M. Bishop & J. Preston (eds.) Essays on Searle's
Chinese Room Argument. Oxford University Press. http://cogprints.org/4023/
Harnad, Stevan (2001) On Searle On the Failures of
Computationalism, Psycoloquy: 12,#61 Symbolism Connectionism
(28)
http://psycprints.ecs.soton.ac.uk/archive/00000190/
Harnad, Stevan (2003) Can a machine be conscious? How? . Journal
of Consciousness Studies 10(4-5):pp. 69-75.
Harnad, Stevan (2003) Symbol-Grounding Problem. Encylopedia of
Cognitive Science. Nature Publishing Group. Macmillan. http://cogprints.org/3018/
Harnad, Stevan (2004) The Annotation Game: On Turing (1950) on
Computing, Machinery and Intelligence. In: Epstein, Robert & Peters, Grace
(Eds.) The Turing Test Sourcebook: Philosophical and Methodological Issues
in the Quest for the Thinking Computer. Kluwer. http://cogprints.org/3322/
Overill, R.E. (2004) Views into the Chinese Room: New Essays on
Searle and Artificial Intelligence. Journal of Logic and Computation14(2):
325-326
Pylyshyn, Z. W. (1984) Computation and cognition. Cambridge MA:
MIT/Bradford
Searle, John. R. (1980) Minds, brains, and programs. Behavioral
and Brain Sciences 3 (3): 417-457 http://www.bbsonline.org/documents/a/00/00/04/84/bbs00000484-00/bbs.searle2.html
Searle, John R. (1984) Minds, brains, and science. Cambridge,
Mass.: Harvard University Press.
Searle, John.
R. (1987) Minds and Brains without Programs," Mindwaves, C. Blakemore
& S. Greenfield (eds.), Oxford: Basil Blackwell.
Searle, John R. (1990) Explanatory inversion and cognitive
science. Behavioral and Brain Sciences 13: 585-595.
Searle, John. R. (1990) Is the Brain's Mind a Computer
Program?", Scientific American, January 1990.
Searle, John R. (2001) The Failures of Computationalism:
Psycoloquy: 12,#62 Symbolism Connectionism (29)
http://psycprints.ecs.soton.ac.uk/archive/00000189/
Souder, L. (2003) What Are We to Think about Thought Experiments? Argumentation 17(2): 203
- 217
Turing, A. M. (1950) Computing Machinery and Intelligence. Mind
59:pp. 433-460. http://cogprints.org/499/
Wakefield, J.C. (2003) The Chinese Room Argument Reconsidered:
Essentialism, Indeterminacy, and Strong AI. Minds and Machines
13(2): 285 - 319