Path: utzoo!mnetor!uunet!mcvax!varol
From: varol@cwi.nl (Varol Akman)
Newsgroups: comp.ai
Subject: Re: Can you sue an expert system?
Message-ID: <148@piring.cwi.nl>
Date: 18 Dec 87 12:42:48 GMT
References: <1788@cup.portal.com> <2366@imag.UUCP>
Organization: CWI, Amsterdam
Lines: 60
Summary: Expert systems cannot be trusted!

Meryem Marzouki writes:
>
> ... material deleted
>
>No comment on this point: I am very trustful in expert systems (I must be so,
>in fact, since I am working in that field), nevertheless, I think that the two
>most important features of expert systems are:
>1) their capacity to verify the consistency of their database(s), and
>2) the domain they are concerned with.
>
> ... material deleted
>
>The expert system was right: it made a deduction from the knowledge it was fed
>on with!
>But the real problem is the domain of expertise, more precisely the suitability
>of an expert system in a particular field.
>It seems to me quite unreasonable to build an expert system for financial
>advice, since this field is continuously in evolution. Moreover, for the
>particular problem of stock market, it is neither a question of months, nor of
>days: it is a question of hours!
>Everybody knows that stock market is particularly precarious, since it can
>easily go up or down, depending on "abstract" parameters, such as feelings, or
>interpretations of official people's declarations (remember the effect of
>Reagan's declarations!), or even the fact that one is tense!
>This kind of knowledge cannot be modeled, at least till now, in an expert system
>database.
>This was to say that the problem is not whether or not to start a discussion on
>qualities/drawbacks of expert systems, but rather on what kind of field is
>suitable for building expert systems.


Expert systems, at this stage of their evolution, are tools for
modeling surface knowledge in an area.  They have no ability
whatsoever to reason about the underlying mechanisms of the domain
that they try to model.  Thus they lack deep knowledge.

Human beings have deep knowledge.  There is also a lot of high quality
work in the area of modeling deep knowledge but this is very much
experimental.  In fact, if we're successful (to an extent) in
modeling deep knowledge, then AI will prove that it is a discipline
which can deal with realistic (read non-toy) problems.

Until then, expert systems will serve as advisors whose advice need
close scrutiny (sp?) by human experts.  I would never try to sue an
expert system because I KNOW that it can't be trusted, given their
level of sophistication at this time.

I can't trust something if it is the subject matter of my field of
research because my field of research is very much in its infancy.
To me that kind of trust is probably the worst thing that I may have.
Programs should be trusted not because we feel a parental warm affinity
towards them.  They should be trusted if they are worth our trust.
The road to that trust is not a path of roses; it is a path full of
hard work, correctness proofs, wide and general field tests, etc.

Until then let's just work and hope that everythings turns out
to be allright at the end.

-Varol Akman
CWI, Amsterdam