COMMENT ⊗   VALID 00002 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	Challenges to weak AI, and perhaps especially to Schankian AI
C00008 ENDMK
C⊗;
Challenges to weak AI, and perhaps especially to Schankian AI

	We propose a small meeting intended to devise the
easiest possible problems that present challenges to performance
of AI systems.  Conceptual issues of whether a given performance
would involve genuine belief, wanting, etc. are out of order at
this meeting.  Skeptics about AI may present the easiest performances
they think no AI systems can exhibit, and skeptics about present
systems can present performances that they think present ideas
for implementation can't do.  The motivation for the discussion
is that philosophical puzzles can often be put in the form of
a contention that an AI system based on certain ideas will not
perform adequately on problems that exhibit certain philosophical
phenomena.

	Here is an example of a possible challenge to Schank's
restaurant script program.

	A man comes into a restaurant, sits down and
when the waiter comes, snarls, "Gimme a hamburger and make goddamn
sure its rare".  Just before the waiter returns with the hamburger,
his scowling face smooths out and when the waiter brings the
hamburger, he says politely, "Pardon me, I like well done hamburgers.
Would you be so kind as to ask them to cook it a little more?"
The waiter takes it back, but before he brings it, he notices
that the benign countenance has again turned to a scowl, so he
quickly takes a rare hamburger that had been intended for another
customer and sets it before the scowling man.  The latter looks
at it, takes a bite and says, "Lucky for you that namby-pamby
Percy didn't get you to cook it more, or I'd have busted your
snoot".  He then takes a large bite of the hamburger and sits
back in the chair.  After a while, he again assumes a benevolent
expression and calls the waiter, who brings the other hamburger.
How many hamburgers are now on the table?  Which hamburger does
the customer now eat?

	Understanding this story requires the reader to get the
idea of two personalities in one body.  Human ability to get this
idea does not depend on a previous formulation of common sense that
distinguishes between personality and body.  Most people go through
their entire lives without ever entertaining the possibility.
An AI system that could answer the above questions might have to
be flexible enough to split a concept - in this case that of
a person - into two concepts on an ad hoc basis.
The ability to do circumscription or other non-monotonic reasoning
is required for a system that can identify body and person in
all ordinary case but make the distinction when an occasion arises.

The above challenge needs to be polished a bit.  It would be worthwhile
if the story could be told in vocabulary already present in Schank's
system.  Moreover, it would be worthwhile to try to anticipate and
prevent cheating solutions such as a simple correlation
of benevolent expressions with well done hamburgers.

	This example is perhaps too complicated.  It was suggested
by notion that Searle missed the Jekyll and Hyde solution to his
Chinese conversation procedure problem.  Surely there are easier
problems that challenge present AI ideas of how programs should
be constructed.

	Actually a simpler problem for humans may be an equal
challenge to a Schankian script based program.  Suppose that
the two personalities are explicitly identified and explained
as Mr. Hyde and Dr. Jekyll.  Can the program even accept this
and then answer questions about who is in control on the basis
of behavior.