Page images

Thus, right at the top, there may be no relationship at all between artificial intelligence and psychology. This is certainly a possible view:

Since the theory rests on analogies between the human and the
mechanical process, Newell et al take some pains to produce
comparisons between human problem solving and the behavior of
the machine. In this effort [LT] they draw upon previously
published descriptions of relevant human behavior. They add
nothing to our further understanding of the living mechanisms,
but they do provide a better understanding of the computer.
(T. Kendler, 1961, pp. 451-452.)

The next stage is where one feels that artificial intelligence provides metaphors, thus making psychologists attend to new phenomena in appropriate ways. This view is the interpretation many scientists put on cybernetics through the forties and fifties. And many people hold it about artificial intelligence now: Psychology and the study of artificial intelligence are both

concerned with intelligent behavior, but otherwise they are
not necessarily related except to the extent that metaphors

borrowed from one discipline may be stimulating to the other.
(A.G. Oettinger, 1969, p. 30.)

No relationship

Metaphor / Attention focussing

Forces operationality

Provides language

Provides base (ideal) models

Sufficiency analysis

Theoretical psychology

Self sufficient

Figure 1: Possible Relationships between AI and Psychology

The next step of engagement is that emphasis on programs and mechanisms forces the psychologist to become operational, that is, to avoid the fuzziness of using mentalistic terms. It is a sort of mental hygiene. Behaviorism is in part a similar sort of mental hygiene, but one that achieves its effect by remaining in the observation language of the experiment (i.e., the behaviors that can be observed). Artificial intelligence offers an operationalism with respect to theory. This view has been very popular, as the following quotations testify:

The advantage of playing this kind of game lies solely in the
fact that, if you talk about machines, you are more certain to
leave out the subjective, anthropomorphic hocus-pocus of

There is still a further step possible along this same road:
the design and construction of actual robots who perform
different human functions as well or better than a man can
The only use that lies in designing an actual robot



is to make sure that, in stating the properties of a function,
we have not left in unwittingly some mystic ambiguous mentalistic
term. (E. Boring, 1946, p. 191.)

On the other hand, the computer program allows us to
specify with complete precision, complex models that certainly
embody what we are vaguely point to with these words. We can
then, as with the concepts "active memory" and "learning" briefly
discussed here, study our models to get a better idea of what we
have been talking about.

The computer is just a powerful tool for clearly specifying
rules that mechanisms must follow in carrying out procedures
that process information. (L. Uhr, 1969, p. 297.)

The next stage sees the language as the major connection: The language of programs and data structures (e.g., list structures) is the appropriate vehicle for describing the behavior of humans, in contradistinction, say, to classical mathematics. An analogous view was strongly held a decade ago in arguing that for the social sciences the appropriate mathematics was that of finite structures (matrix analysis, markov processes, graph theory), as opposed to the mathematics of the continuum (i.e., differential equations). Perhaps, the clearest statements of the language view with respect to artificial intelligence have been made by George Miller:

The computer program can play a double role in psychology: as
a model of an intelligent system and, even more broadly, as a
kind of language in which theories can be expressed. Everyone
recognizes the importance of holding a good theory; the advantages
of speaking a good language, however, are not so often recognized.
(p. 9)

There is much that the psychologist can learn from a study of
computing machines and the structure of their programs. Programm-
ing languages seem to offer an excellent medium for the expression
of psychologicial theories, even though using such languages implies
that men and machines are in some deep sense considered to be equi-
valent functionally, if not structurally. (G. Miller, 1962,
p. 21.)

The stages of metaphor, operationality and language are somehow content free. That is, the gains to psychology are in various behaviors and disciplines of the psychologist. The next stage finally accords the product of the artificial intelligence models significance, even if not their content. Here artificial intelligence is used to provide base lines against which to view actual behavior. These base lines are in the direction of optimum behavior, rather than in the direction of random behavior as in the base lines usually provided for by statistics). Such ideal types are used fruitfully in several places in science. In psychology a good

example is the work of Ward Edwards on behavior in uncertain situations, where
humans are consistently conservative compared to the optimal solution, as computed
from Bayes theorem. Without this comparison with an ideal system, a significant
aspect of the data would be missed. In artificial intelligence this view is
perhaps less common than might be suspected, given that computers are programmed
to do the best job possible. Nevertheless, one finds the attitude expressed

The computer analogues used in some of the model of human
information processing and thought depict ideal intellectual
slaves, experiencing practically no time lag, no loss of memory,
and no reluctance to consider all of the available evidence.

The human to whom our formulations are meant to apply do
unfortunately experience considerable limitations in these
regards. (W.J. McGuire, 1968, p. 159).

The next turn of the screw reflects a unique feature of human cognitive behaviors, namely that they constitute performances for which often we do not know any way that they can be accomplised. Thus, it becomes of interest to discover systems that perform these tasks. If, in addition, no mechanisms are used in these systems that clearly go beyond the capacities of the human, then an initial theory has been provided. This level has been called sufficiency analysis, since it seeks to show that a sufficient set of mechanisms exists for a particular intellectual task. To illustrate, if one develops a chess program that examines 800,000 positions in deciding on a move, then one has not made a contribution; since excellent evidence exists that no human could consider 800,000 separate items of information in ten minutes. But if the chess program only considers around 100 positions, and if there are no other ways in which the program radically violates the general character of human processing capacities, then it may be taken as a first model. An example of this view is the following:

The definitions are both nominal and ostensive in the sense that
when we speak, for example, of "pathogenic conflict" we can

point to a precise procedure in the program which computes
whether two beliefs are in conflict or not. We must postpone
the question, which eventually must be faced, of how closely
this corresponds to the nature of pathogenic conflict in real
persons. But at this point we can say there is a rough match
between the output of the program and typical behavior of
patients in psychotherapeutic sessions. (K.M. Colby and
J.P. Gilbert, 1964, p. 417)

This view has a certain value in itself, since psychology has 1. general ignored the question of explaining how it is that humans can perform the acts of intelligence they routinely accomplish. Thus, it adds a new mode of analysis.

With the next turn, we get artificial intelligence as theoretical psychology. This is analogous to the view of the mathematics of differential equations as theoretical physics. Thus the actual theories of cognitive psychology are to be expressed as artificial intelligence systems. We would expect to find artificial intelligence systems of direct empirical relevance, and also artificial intelligence systems being developed for their own sake, just as in mathematics there is concern with the differential equations of physical interest (e.g., the Mathieu equation) and also the pure theory of differential equations. This view has been often expressed; for instance:

Quite typically, these models express psychological propositions
in terms of individual operations for matching, generating,
transforming, and retrieving information. These operations
are knit together to form systems of complexly organized
structures and processes. Since the structures and processes
are represented explicitly, such models enable us to go behond
measures of the quantifiable and statistical properties of
behavior to investigations of the specific sequences of stimuli
and responses involved. ... By comparing model-generated
behavior with data from humans, we can decide unambiguously

« PreviousContinue »