Path: utzoo!utgpu!news-server.csri.toronto.edu!rutgers!mcnc!rti!ntpdvp1!kenp
From: kenp@ntpdvp1.UUCP (Ken Presting)
Newsgroups: comp.ai
Subject: Re: Hayes vs. Searle
Summary: A more rigorous version of Searle's original argument
Message-ID: <619@ntpdvp1.UUCP>
Date: 7 Aug 90 03:50:42 GMT
References: <129.26a5feab@csc.fi> <14385@venera.isi.edu> <25618@cs.yale.edu> <1620@oravax.UUCP>
Organization: SNA Solutions Inc., Contract Programming Group
Lines: 115

In article <1620@oravax.UUCP>, daryl@oravax.UUCP (Steven Daryl McCullough) writes:
> . . . . It seems to me that judging
> intelligence from behavior must come first.

Daryl, you are simply overlooking an immense variety of possibilites, but
rather than list more of them, I want to discuss the CR from the perspective
of the *program* that generates the behavior.

> (Ken Presting) wrote:
> > Evidence of all these types is objective and public, and has
> > nothing to do with anything as confusing as introspection.
> 
> I think you would see the role of introspection in all of this if you
> would only use a little 8^) The question that started all this Chinese
> Room stuff off was: "Is behavior sufficient to determine whether a
> system is intelligent"? The Strong AI position assumes that the answer
> to this question is "yes". Your disussion of radical translation seems
> to show your general agreement with this position. Yet, in my reading
> of Searle, he is in disagreement with this position.

Searle certainly thinks that radical interpretation is inadequate to
establish the meaning of words in a language (I appreciate David Chalmers
correcting me on this).  But he does *NOT* think that behavior in the
wide sense you advocate is inadequate to determine the intelligence of
a system.  He says so in the original BBS article, under "The Combination
Reply".  Searle is opposed to behaviorism, which is usually defined more
narrowly than the position you have taken.

But the CR is not based on behavior - it is based on programs.  Searle
conspicuously defines Strong AI in terms of its position on what programs
say about the systems that implement them.  If Searle is right (and I think
he is not) then the inadequacy of the Turing Test would follow.  IMO, the
TT needs to be more specific, but you, me, and Searle all agree that
empirical observations of behavior over time are sufficient to convince
any reasonable person of another's intelligence.

> 
> . . .  If introspection didn't tell us that there was a
> difference, why would it occur to us to make such a distinction?
> 
> If you ignore introspection, I would see no plausibility to Searle's
> Chinese Room argument at all. As it is, I still don't find it
> compelling for precisely the reason that the introspection of the man
> in the room is irrelevant.
> 

If I thought the argument had any dependence on introspection at all, I
would probably take no interest in it at all.  I agree that introspection
by the man in the room is irrelevant!

Let me try to make another example.  Suppose you and I each go the library
and read a book about Chinese.  I read a book called "Introduction to
Chinese Syntax" and you read one called "Introduction to Chinese Semantics".
We do *not* need to perform any introspection to infer that I will learn
something about syntax but little about semantics, while you will learn
something about semantics but little about syntax.  Searle draws the 
inference that the man in the room cannot understand the symbols, because
the books which hold the program do not say anything about Chinese semantics. 

This is supposed to be so obvious that nobody could possibly think 
otherwise, and I think that's why Searle is so recalcitrant about the 
systems reply.  From his perspective, it very simply misses the point.
Searle thinks that the program contains information only about the syntax
of Chinese conversations, and therefore neither he nor any other implement
can acquire semantic information from the program.  If you grant that
understanding requires semantic information, then his conclusion follows.
(Recall that my position is that *some* programs *can* contain semantic
information.  I am NOT defending Searle.)

Let me put it another way.  Suppose we write a program for factoring large 
numbers.  It should be obvious that memorizing this program will not give
anyone the ability to calculate fast Fourier transforms.  That is the basic
argument structure of the Chinese Room Example - obvious and trivial.  Just
by assuming that the syntax and the semantics of a language are two different
bodies of knowledge, Searle sets the stage for a straightforward conclusion.
(This easy conclusion is independent of the biology/silicon "causal powers"
issue, which is a little more complicated.)

So the issue becomes, what kind of program can contain semantic information?
The problem is that whatever program you write for any automaton, it is
equivalent to a language-recognition automaton, which by definition has
only semantic information in it.  This is just mathematics, and you can
find it in Hopcroft and Ullman, "Formal Languages and their Relation to
Automata".


> > There is no need to settle too quickly for immediate I/O activity
> > alone as a criterion for deciding the success of AI, even if we are
> > deliberately excluding criteria based on the internals of a system.  
> 
> You seem to have a much narrower notion of "behavior" than I have. To
> me, a system's behavior is the relationship between the system's past
> history and its future actions, and not simply "immediate I/O
> activity". If you deliberately exclude criteria based on internals of
> a system, then behavior is all that one has left.

To mathematically describe the operation of a program which is dependent
on more than immediate I/O activity, you cannot restrict yourself to
Turing machines operated in the usual way.  You need some sort of
permanent memory (which everday computers always have).  I don't think
Searle can be blamed too much for assuming the properties which are
universally attributed to Turing machines in his argument.  TM programs
are purely formal, and TM's (and their implementations) are purely
syntactic devices.  That folows directly from Church's thesis, nothing
more.  No TM can do anything a context-sensitive grammar couldn't do
just as well.

I have tried to show that Searle's conclusion follows from his premises,
by adding a few steps that involve no introspection and no assumptions
of non-behavioral observations.  Nobody (except Husserl fans like
Stephen Smoliar) wants to base any conclusions on introspection.  Searle
certainly does not need introspection in his argument.


Ken Presting   ("Thou shalt not covet thy neighbor's program")