Posted Sun Jul 27 13:35:33 BST 1997
Newsgroups: comp.ai.philosophy,comp.ai,sci.cognitive,sci.psychology.theory
References: <5q8mkn$gtc@ux.cs.niu.edu> <01bc908f$51ef6da0$e37a61ce@asdf> <5qgb5g$c5p@usenet.srv.cis.pitt.edu> <868992479snz@longley.demon.co.uk> <5qjejt$np6@usenet.srv.cis.pitt.edu>
From: AaronSloman@cs.bham.ac.nospam (Aaron Sloman See text for reply address)
Subject: Re: Natural Judgement & Artificial Intelligence: A Reconceptualisation

[I apologise in advance if my comments below missed the point because I
have not read the preceding articles. I don't have time to do more than
browse occasionally, and in any case not all the articles posted reach
our site, nor do they reach us in a sensible order.]

andersw+@pitt.edu (Anders N Weinstein) makes some very strange
comments, apparently provoked by one of David Longley's
non-sequiturs.

> Date: 16 Jul 1997 21:31:41 GMT
> Organization: University of Pittsburgh
[DL]
> >An even better way is to appreciate that there are problems  with
> >the use of psychological "predicates" full stop. The reason being
> >that  (loathe as we are to recognize it), *we* are  mot  unitary,
> >consistent  agents at all. We have multiple views  (dispositions)
> .....
> >in relation to all of our experience and only fail to  appreciate
> >this because of a rationality "assumption".

The obvious reply to this, which Anders does not like, is that being
inconsistent and irrational (as humans indeed are much of the time) does
not mean that we cannot be the subject of psychological predicates.
In fact among the characteristics of being a human are having the
ability to make mistakes, change one's mind, become confused, become
irrationally angry or jealous, develop new standards, become more
mature, become more forgetful, etc.

In fact if we could not apply psychological predicates to irrational
individuals we could not say that they were irrational!

The claim that psychological predicates are applicable only to
consistent, or rational entities is popular among some philosophers and
cognitive scientists (e.g. it's part of the definition of Dennett's
"intentional stance" and Newell's "knowledge level") but is just a
mistake.

We can talk about a human infant or a cat as having mental states (e.g.
perceiving things, wanting things) without imputing rationality to them.
(At least not the kind of rationality that Anders talks about -- see
below.)

That's because (according at least to my analysis, which remains to
be tested fully) our concepts of mentality do not presuppose
rationality in the things we apply them to, rather they presuppose
the existence of a suitably functionally differentiated information
processing architecture.

I.e. attributions of intentionality are normally (among ordinary
people who have not studied philosophy) based on what Dennett calls the
"design stance", not the intentional stance.

E.g. in treating an animal or an infant as being able to see we do not
consciously or unconsciously presuppose it to be rational.

It can see if it has certain kinds of optical transducers which provide
information about the environment which is used in a variety of ways
(which depend on what sort of agent it is), including triggering new
actions, controlling actions via feedback loops, triggering conceptual
development, creating or updating long term information stores,
generating new goals (if the animal has goal processing capabilities),
and possibly many more.

It's one of the goals of AI to investigate the variety of forms of
such processes and the kinds of architectures and mechanisms which
support them.

Then how can we ever know that other people can see, feel, hear, etc.?
Do we have to first check out their information processing
architectures?

Obviously not.

As I have explained in this news group more than once, I suspect we are
born with innate predispositions to interpret the behaviour of
surrounding agents in terms of a presumed model of the information
processing architecture producing their behaviour.

This model is partly influenced by the culture in which each child
develops, but the initial framework is innate, designed by a long period
of evolution. (It's part of a much richer repertoire of innate
dispositions, including, for instance, the disposition to interpret
perceived events as causally connected: which is not something which
could possibly be learnt from experience.)

All this makes it unnecessary for newborn infants to solve the "other
minds" problem by going through a process seriously proposed by some
philosophers, of arguing "they behave as I do so they probably have
minds like mine", which is totally implausible, not least because
most of the people in the environment of an infant do not behave at
all like infants. In fact, the infant's survival depends on the fact
that they don't.

(There's also the fact, usually ignored, that babies cannot observe
their own behaviour unless left constantly in front of a mirror, which
is not a requirement for normal infant development. If they use
proprioceptive feedback to "observe" themselves that's no help in
arguing about other minds, because they are not hooked into the
proprioceptive feedback of the people they observe. So they need to
compare their own proprioceptive data with the visual data concerning
behaviour of others: a task that's impossible without a host of prior
innate assumptions. So either way infants have to make assumptions that
go beyond what they can sense. This is not surprising, there are many
largely innate behaviours in all sorts of organisms: several thousand
million years of evolution can be expected to produce a store of
useful in-built information.)

Rationality, as Anders points out later, requires a type of self
awareness and self evaluation (e.g. the ability to judge one's decisions
and reasoning as correct or incorrect) which it is not reasonable to
expect of cats and infants.

But it is equally unreasonable to infer that they therefore cannot have
any mental states or processes. They merely don't have typical RATIONAL
adult human mental states and processes.

Instead of pointing this out, Anders goes off in another implausible
direction in commenting on the assumption that humans are rational:

[AW]
> The point is, it is not any kind of assumption or hypothesis, any more
> than the statement that John has promised to come is an *assumption*
> that he will come. It is rather a normative *demand* we place on
> people. To the extent their various psychological states are
> inconsistent or otherwise dissociated from one another, the people are
> irrational, and can justifiably be called to account for it.

It is totally unreasonable to claim that a cat or an infant can
justifiably be called to account for its inconsistencies, dissociations,
or what may appear to be irrational decisions, preferences, or actions
(I'd prefer to call them "non-rational" when the capability for
rationality is not (yet) present).

Similar comments can be made about adults who have suffered various
kinds of brain damage or senile decay. Have you ever had a close
encounter with an Altzheimer's victim?

But just because infants, cats, brain damaged individuals are not
capable of rationality it does not follow that no psychological
predicates are applicable to them, e.g. that they cannot see, feel pain,
etc.

[AW]
> Thinkers as such are *constituted* by relations to normative
> standards.

Such philosophers' talk about being *constituted* is just a disguised
and confusing way of making semantic claims, i.e. claims about
definitions (which may be implicit or explicit). I.e. this is a claim
that the notion "thinker" is *defined* (implicitly or explicitly) by
relations to normative standards.

That may be your definition of "thinker". Is there any reason why other
people should follow it?

Of course, if you mean, not "thinker", but "rational thinker" then we
can agree with you. But why confuse the more general concept with the
narrower one, which is applicable only to a subset of people?

[AW]
> Although imperfect, they count as thinkers only by their
> manifesting a relation to an ideal they grasp themselves.

Including infants, cats and people with brain damage?

[AW]
> This is why
> persons fall in a completely different ontological category than
> computers or machines generally.

Perhaps you exclude very young humans and people with senile decay from
the category "person"?

We can all then remember that what Anders means by "person" is "rational
person", while many other people use the word more broadly.

He then goes on to make a sweeping generalisation about computers.

[AW]
> The computers don't *conduct*
> themselves, they don't manifest sensitivity to standards or striving
> after ideals in their behavior.

Which computers are you talking about?

What would you say if you encountered computer-based robots that had
been designed to start off like human infants with a minimal collection
of cognitive and other capabilities, and over a long period of time
bootstrap themselves, as children do, so that they absorb a culture,
develop a collection of concepts adequate to cope with their
environment, learn the local language, acquire a whole lot of new
tastes, standards, preferences, attitudes, ideals, and learn to assess
their own reasoning and decision making behaviour according to those
criteria?

Or are you saying that that's impossible? If so, perhaps you could give
some argument. I don't see anything impossible about it, though
producing such robots is a very difficult long term project requiring
the design of much richer information processing architectures than
anything so far found in AI.

It may turn out to be impossible for some reason that we don't yet
understand, e.g. concerned with processing limitations of computers, or
some hitherto unappreciated requirements for human-like capabilities.
But making it impossible by definition is just an unjustified attempt to
stifle research.

(Some people say X is impossible, if THEY cannot think of any way to
achieve it, but I would not expect Anders to make such a simple mistake.
Some people say X is impossible if they wish it were impossible, e.g
because they wish to preserve a unique status for human beings. I
sometimes get hints of such wishful thinking in what Anders writes. I
wonder if he would be truly upset to discover he was wrong, or merely
say "OK I made a mistake, as we all do from time to time...", applying
his own standards of rationality to himself.)

[AW]
> Rather we who design them impose
> normative standards of proper functioning from without.

That may be true of most existing computing systems. It does not have to
be true of self-modifying systems.

E.g. for the infant robot growing up in our culture there need be no
more imposition of standards from without than there is in the case of
human children. And as you are probably well aware not all humans adopt
the standards which the surrounding culture hopes they will adopt (hence
the need for criminal courts, etc.).

The same sort of rejection of external standards could happen in a
developing robot child.

(Some science fiction writers have supposed that this could be ruled out
by building self-limiting mechanisms into robot brains. I do not believe
that in general such constraining mechanisms would be consistent with
fully functional human-like learning and development. I.e. it may be
possible only for robots that are so limited in the kinds of things they
can do that they will never become as intelligent and creative as a
normal person, i.e. not the kind of robot you would want as a friend, or
perhaps as a philosophy tutor for your children. My reasons are too
complex to present here, but are essentially a consequence of the
computational intractability of checking that any new piece of software
meets all its requirements and constraints, except in relatively trivial
cases.)

> But those norms
> are nothing to the machine. Deep Blue is not ashamed of itself if it
> makes a blunder.

You have fallen into the same kind of trap that Longley often falls
into:

    a is P and a is Q, therefore All P's are (necessarily) Qs.

Cheers
Aaron