Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Posting-Version: version B 2.10.1 6/24/83; site mit-athena.ARPA
Path: utzoo!linus!decvax!mit-athena!yba
From: yba@mit-athena.ARPA (Mark H Levine)
Newsgroups: net.ai
Subject: Re: Re: The Turing Test - machines vs. p - (nf)
Message-ID: <206@mit-athena.ARPA>
Date: Wed, 11-Jul-84 19:58:47 EDT
Article-I.D.: mit-athe.206
Posted: Wed Jul 11 19:58:47 1984
Date-Received: Fri, 13-Jul-84 01:02:11 EDT
References: <290@gloria.UUCP>, <500001@ea.UUCP>
Organization: MIT, Project Athena, Cambridge, Ma.
Lines: 19

If a program passes a test in calculus the best we can grant it is that
it can pass tests.  In the famous program ANALOGY (Bobrow's I think)
the computer "passes" geometric analogy tests.  It does not seem to understand
either geometry or analogy outside of this limited domain of discourse.
We make the same mistaken assumption about humans--that is that because
you can pass a "test" you understand a subject.

The Turing test was a "blind" test; in that the Colonel is wrong--someone
reading this over the net or receiving a note from the bank cannot just
"go look".  The idea was to tell via dialog only in a blind situation
(maybe even a double-blind if there are some control situations where
two humans taking the Turing test face each other).

The question of how to evaluate the performance of an AI system has become
an important question.  I am not sure that the question of "understanding"
should even enter into it.  In any case, let's not trivialize it.

-- 
yba%mit-heracles@mit-mc.ARPA		UUCP:	decvax!mit-athena!yba