Newsgroups: comp.ai.philosophy
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!news.mathworks.com!udel!gatech!swrinde!cs.utexas.edu!rutgers!argos.montclair.edu!hubey
From: hubey@pegasus.montclair.edu (H. M. Hubey)
Subject: Re: Turing test (was Penrose and Searle)
Message-ID: <hubey.786396076@pegasus.montclair.edu>
Sender: root@argos.montclair.edu (Operator)
Organization: SCInet @ Montclair State
References: <38tqh6$5qk@percy.cs.bham.ac.uk> <3bfphr$6sj@news-rocq.inria.fr> <D03Lp9.L4H@gpu.utcc.utoronto.ca> <hubey.786306282@pegasus.montclair.edu> <D05IFt.CwK@spss.com>
Date: Fri, 2 Dec 1994 19:21:16 GMT
Lines: 161

markrose@spss.com (Mark Rosenfelder) writes:

>I hope you know more statistics than this.  The fact that a given test cannot
>distinguish between men and humans does not prove that there is no 
>difference to find.  It may be that the test is lousy.


What's wrong?  There are four possibilities whenever we are faced with
the truth/falsity of a statement. Let me restate it


>>11	We can call it True(1) when it is in fact True(1)
>>10	We can call it True when it is False  (Error)
>>01	We can call it False when it is TRue  (Error)
>>00	We can call it False when it is False

1. We don't have any instruments for measuring inteligence [except
for those things called IQ tests.]  
2. Even those tests are themselves subject to the same constraints 
as above.
3. The TT does not even pretend to have produced a standard set of
questions which could be used [even in principle] to be used by
humans as a test of whether the entity being tested is a computer
or a human. So it's even more flexible and the same comments above
still apply, except that in this case, the questions/answers can
be made up of humans as they go along, and we can test with
as many humans as we like, so that there's no restriction on 
whether we are allowed to pick up yokels from the streets or that
we can use experts in various fields to make the test even better.


In real life we do this all the time. Let me paraphrase it as it
has already been done centuries ago.

00  Those who know not that they know not are fools, shun them.
10  Those who know they know not are students, teach them.
01  Those who know not that they know are asleep, wake them.
11  Those who know they know are wise, follow them.


We are up against the same problem. I didn't want to confuse the
issue even further. How does anyone know if X is type 00 or 11 ?
It's relative.  Let's restrict this to a specific domain. Then type
11 is the expert [one who knows what he/she knows and also knows
what he doesn't know]. Type 10, or 01 might not be able to tell type
00 from a type 11. Just look at politics, why do people vote for
people like Hitler or Mussolini. Look at Zhirinovsky's success.
Look at how the public voted anti-Clinton [i.e. anti-democratic]
that's the same public that voted Clinton into office.

A blustering bluffer (00) might be able to fool 01 or 10 types
but will not fool a 11. Anyway this is all detail. The TT does
not put any restrictions on what kinds of humans you can choose
for the test.

And finally, we can't do any better in real life. How do you know
if someone you know really is intelligent [knowledgeable, etc etc]?
How does anyone know? How do you know that chimps are not as
intelligent as humans? How do you know chimps are more intelligent
than dogs? So on and so forth...



>There are statistical tests to show whether or not a measuring system is
>capable of making the discriminations asked of it.  (I helped write a
>product here that makes such tests.)  Any measurement process will have
>some fuzz-- some randomness in its results.  If the spread of this fuzz
>is large enough in comparison with the thing measured, the measurement
>process is useless.  

Yes, so?


>We don't know if the measurement errors in the Turing Test are enough to
>swamp the results-- and that's a problem right there.  But the known false
>positives-- the people who judged ELIZA and the Loeb entries as human--
>suggest that the fuzz is indeed very large.

The TT says Yes/No. Despite the fact that formalists make strict
distinctions between probability, fuzzy logic, and bivalent logic, in
effect, fuzzy logic is a simplification of prob. theory when we don't
have enough knowledge (or don't care since it's not going to be
very important because of the "fuzz" as you call it). In a similar vein
one can think of logic is a further simplification of many-valued
or infinite-valued logic into only two values.


PS. The fact that some people judged Eliza to be human is not
a serious problem [or put another way, is no more serious than similar
problems in real life]. All it says from one point of view is that
programs can be intelligent [after all, some people have been fooled].
From the other perspective, it says that the test is bad because
we know it's a computer program but people get fooled anyway.

This is essentially the gist of the argument right from the moment
these started to get discussed on this newsgroup and others. One is
the Searlian position which says "it's not intelligent because it's
not human". The other says "yes, but it already acts intelligent,
therefore it is. If you can't tell the difference, then on what basis
do you claim it's not intelligent?"

This is like the problem of whether we invent truth or discover it.
What you believe depends on your attitude.


>Has anyone tried?

1)What's stopping them?  
2) What are IQ tests [all brands]?

This is like the misbehavior of a bad-tempered sheep dog who doesn't
eat the grass and doesn't allow the sheep to eat it.

On the one hand they can't produce diddly-squat. On the other hand,
they won't stop criticizing efforts because they're not perfect. Someone
should tell them that simplifying things and then solving the simple
versions is how progress is made.


>>What would be really 
>>interesting is if some one wrote a program that takes those standardizes
>>IQ tests that pschologists seem to be fond of giving to humans :-)..

>Programs designed to read and solve algebraic word problems were written
>back in the '50s.

>As a generalized test of intelligence, IQ tests are even worse than the 
>Turing Test; the notion of "intelligence" they cover is ludicrously narrow.

Yes. Exactly the point. Psychologists are the only ones who seem to
even try and their tests are even worse. Even if a simple version of
Schank's script/frame ES were implemented on a powerful computer, it
could probably pass the reading comprehension tests. The mathematical
problems will be even easier to pass. Then what?

What's intelligence anyway?  The counterarguments lack conviction because
they always seem to be versions of these:

1) Submarines don't really swim because they don't wiggle their tails
2) Airplanes don't fly because they don't flap their wings.
3) Dish washers don't really wash dishes because they only squirt soap
water, blah blah.
4) Cars don't really run fast, because running refers to locomotion
with the legs, and cars don't have legs.
5) add more...


All the arguments are about "what is intelligence?" Turing has answered it
in the best way. We have it. And we can tell if others have it?

If aliens from another galaxy who are 15 million years beyond us visit
the earth and they decide that to them we are just steaks [just like
cows to us], then we can discuss the weaknesses of the Turing test.
Other than that, let the naysayers put up something in its place or
shut up.


--
						-- Mark---
....we must realize that the infinite in the sense of an infinite totality, 
where we still find it used in deductive methods, is an illusion. Hilbert,1925
