Newsgroups: comp.ai.philosophy
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!howland.reston.ans.net!gatech!rutgers!argos.montclair.edu!hubey
From: hubey@pegasus.montclair.edu (H. M. Hubey)
Subject: Re: Turing test (was Penrose and Searle)
Message-ID: <hubey.786763482@pegasus.montclair.edu>
Sender: root@argos.montclair.edu (Operator)
Organization: SCInet @ Montclair State
References: <38tqh6$5qk@percy.cs.bham.ac.uk> <hubey.786306282@pegasus.montclair.edu> <D05IFt.CwK@spss.com> <hubey.786396076@pegasus.montclair.edu> <D0Cqoy.BLr@spss.com>
Date: Wed, 7 Dec 1994 01:24:42 GMT
Lines: 68

markrose@spss.com (Mark Rosenfelder) writes:

>In article <hubey.786396076@pegasus.montclair.edu>,
>H. M. Hubey <hubey@pegasus.montclair.edu> wrote:
>>markrose@spss.com (Mark Rosenfelder) writes:
>>>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?

>Nothing in your reply addresses this objection.  Your original posting 
>claimed that errors in the test results don't matter because "It's a 
>statistical test."  In fact statistical methods don't help a bit if a
>measurement process cannot produce the discriminations that are asked of it.

I don't see the problem.  You already have the best instruments for
measuring or discriminating intelligence from non-intelligence; the
instrument is called "a human".

YOu can pick 500 humans and run the test and then take a simple average
of the type I and II errors (and you know this because you know the
results already) or you can make it more strict by using only experts.

In fact, the fact that some people get easily fooled by Eliza
doesn't mean much. As soon as programs like this become common
commodities, even high school dropouts will be able to tell
canned programs. I test Eliza in my classes with average students (not
computer science majors). They can see after a few minutes that
it starts creating nonsense.


>(A nice thing about the NDDC statistic, by the way, is that it doesn't
>require a master value; it just relies on a collection of measurements
>from the process being investigated.  In the case of the TT, this means
>that you don't need to know whether a given measurement is "correct" or
>not (whether the thing being tested really is intelligent).)

There's much neither statistics, nor logic can handle. I can't put my
finger on it but certain kinds of quite natural behavior seem abnormal
if looked at from the point of view of logic and correlation-regression
requires that you already have to supply the form of the relationship.
If the relationship is not supplied then you can try all types and then
try to select some "best" one.  

In the case of the TT, we leave everything alone and ask humans (who
seem to be the most intelligent on earth) to recognize if some other
entity has it.  The fact that humans make mistakes is not a big deal.
It's the mistakes that count.  In fact, it's often a puzzle, who really
is "smarter"; the guy who got A's in HS and college and became a HS
math teacher or the guy with the C grades who started his own business
and became "successful", or the 6'5" guy who didn't even study but
became a basketball player.


>We don't know what the NDDC for the Turing Test is, of course.  Till we
>do, it's a thought experiment, not a scientific test.

It can easily be turned into a scientific test based on the ratio of the
type I and II errors for both humans and machines. 

--
						-- 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
