From newshub.ccs.yorku.ca!torn!cs.utexas.edu!usc!sol.ctr.columbia.edu!destroyer!ais.org!umeecs!dip.eecs.umich.edu!marky Mon Aug 24 15:40:42 EDT 1992
Article 6608 of comp.ai.philosophy:
Newsgroups: comp.ai.philosophy
Path: newshub.ccs.yorku.ca!torn!cs.utexas.edu!usc!sol.ctr.columbia.edu!destroyer!ais.org!umeecs!dip.eecs.umich.edu!marky
>From: marky@dip.eecs.umich.edu (Mark Anthony Young)
Subject: Re: Turing Test Myths
Message-ID: <1992Aug13.035633.15424@zip.eecs.umich.edu>
Sender: news@zip.eecs.umich.edu (Mr. News)
Organization: University of Michigan EECS Dept., Ann Arbor
References: <BILL.92Aug11105853@ca3.nsma.arizona.edu> <1992Aug12.063425.13479@zip.eecs.umich.edu> <BILL.92Aug12122254@ca3.nsma.arizona.edu>
Date: Thu, 13 Aug 1992 03:56:33 GMT
Lines: 45

%r bill@nsma.arizona.edu (Bill Skaggs)
>
>Okay, I was wrong.  It was me who was confused.  But it's hard for me
>to believe that Turing actually realized what he was saying when he
>wrote this.  Suppose it turned out that interrogators could
>distinguish between men and women 100% of the time, and they could
>also distinguish between men and computers 100% of the time.  Would
>this imply that computers think as much like men as women do?
>Obviously not.
>
You're right.  If people could distinguish with 100% accuracy, the
test would be useless.  You have run into the "ceiling effect".
Your test is too easy, so you can't separate the smart from the
dumb (it's like giving a calculus exam with one question: "What is
1+1 equal to?").  In order for the test to be useful you must have
a success rate that is away from the end points of the scale.
To get a useful number you can manipulate the amount of time
allowed for the decision (you might even be able to collect "best
guesses" at various points and make a curve).

>I believe that, given an hour of interrogation, I would have a better
>than 90% probability of distinguishing between a man and a woman.  I
>might be right, or I might be wrong, but it doesn't seem reasonable
>that the question whether I'm right or wrong has anything to do with
>the question whether machines can think.
>
The particular datum of whether _you_ were right or wrong is not
directly related to whether machines can think.  It is only as part
of the aggregate data that it is relevent.

Suppose you had taken a test, and the professor handed it back to
you and you saw that you got an 85 (of 100).  Did you do well?  How
can you tell if you did well?  If the class average was 61, then
you did well.  If the class average was 98, you did very poorly.

The man-woman part of the Turing Test provides the class average
(the base line).  The question of whether _you_ were right when
_you_ played the immitation game corresponds to the question "Did
you get a point for question 37(a)?"

(If we had Klingons and Vulcans and Romulans and the like running
around, we could come up with an even better base-line.  Still, we
have to work with what we've got.)

...mark young


