Artificial Knowing Gender and the Thinking Machine by Alison Adam
2
AI IN CONTEXT
Marvin Minsky (1968: v) defines AI in these terms: ‘Artificial Intelligence
is the science of making machines do things that would require intelligence
if done by men.’ However, in describing what AI is about, I doubt whether
it is ultimately useful to offer an immutable definition, and certainly not
one which defines it in terms of men’s intelligence! The ‘artificial mind’
myth has contributed to a view of AI which, as well as being too mystical,
may well be some distance from the intentions of those working in the
field. This is especially problematic if, as I have suggested, AI researchers
tend to view their discipline in terms of engineering, of designing and
uilding computing artefacts; this does not tie in neatly with an ‘artificial
mind’ view. It is hard to know why such a position has proved so per-
sistent. It may be that for many, at least outside the confines of computing,
their introduction to the subject comes from one of the widely known
philosophical critiques, such as that of Dreyfus (1979; 1992) or Searle
XXXXXXXXXXIn such a case it would be easy to fix on the idea that the aim of
AI is primarily to create an artificial mind, and that the success or failure
of the whole AI project should be judged against this one goal.
GENERAL PROBLEM-SOLVING - THE EARLY DAYS OF AI
A researcher entering the field of AI at the end of the twentieth century
enters a mature discipline with clear boundaries and a set of problems
which are deemed to be appropriate for the subject, what Thomas Kuhn
(1970) would have termed a ‘paradigm’, or Imre Lakatos (1970), a ‘research
programme’. Forty or more years earlier, in an entirely new subject area,
the choice of appropriate problem was not so clear. Essentially the defi-
nition of what constituted an appropriate problem for AI was still open
in the mid-195Os, when there were just a few key players, including He
ert
Simon at Carnegie Technical Institute (later Carnegie Mellon University)
collaborating with Allen Newell of the RAND Corporation, and Marvin
Minsky collaborating with John McCarthy at the Massachusetts Institute
of Technology (MIT), in efforts to produce a working computer system
34
AI IN CONTEXT
which would set the standard for the nascent discipline. I argue that the
kind of problems which were chosen in that period was remarkably sig-
nificant, as it set the style for the symbolic AI programme which was to
dominate AI research for the next three decades or so. In deciding what
constituted appropriate intelligent, behaviour to be modelled in their com-
puter systems, the new AI researchers naturally looked to themselves. As
Tom Athanasiou (1985: 13, quoting the researcher Bob Wilensky, explains,
They were interested in intelligence, and they needed somewhere to
start. So they looked around at who the smartest people were, and
they were themselves, of course. They were all essentially mathema-
ticians by training, and mathematicians do two things - they prove
theorems and play chess. And they said, hey, if it proves a theorem
or plays chess, it must be smart.
Although it was not the very first AI program to be developed, in the
official history of the subject, the first significant AI program: is widely
taken to be Allen Newell, J. C. Shaw and He
ert Simon’s XXXXXXXXXXLogic
Theorist. Simon initially considered three tasks for the program: chess,
geometry and logic theorem proving - the latter for no deeper reason,
apparently, than he happened to have the two volumes of Bertrand Russell
and Alfred North Whitehead’s Principia, the ‘bible’ of predicate logic, at
home. In the history of AI, Logic Theorist is highly significant as it
mapped out the field for AI search strategies and the use of heuristics
which were developed from Simon’s own work on decision theory. In this,
it can be seen how the ideal of rational decision-making was ca
ied over
into the concept of search in AI.
Simon (1976: 20) characterizes rational decision-making as a process of
listing alternative strategies, determining the consequences of each and then
comparatively evaluating each consequence in turn. Decision theory applies
sophisticated mathematics to decisions in a number of areas based on these
precepts. Similarly, the idea of searching for a solution to an AI problem
involves characterizing the problem as a number of discrete and formally
described states, one or more of which will be a starting state of the
problem and one or more of which will be a goal or solution state.
Operations or rules, which move the problem from one state to another,
and a test or evaluative function, which determines whether the problem
has reached its goal or solution state, must also be defined. The problem
then is seen in terms of a search for a solution, going from one state to
another and another and so on until the goal is reached. Hence, the
problem is moved from one formally defined state to another in some way
which is regarded as rational, perhaps guided by a heuristic or rule of
thumb which may help to find a solution more quickly.’
As Simon himself realized, the idealized model of decision-making was
arely, if ever, achievable in a real situation, since an individual could never
35
AI IN CONTEXT
know all the alternatives and their consequences. But Simon’s critique was
not an objection to the rationalistic approach as such, instead it was an
objection to the assumption of full knowledge. Applying this idea to
computer evaluations of decisions, the computer must operate within a
ounded rationality. These ideas developed into more general theories of
problem solving which I argue can be seen more particularly in the widely
accepted idea of the search procedure in AI.
It should not necessarily be taken for granted that solutions to problems
are things to be searched for. Generally speaking, the idea of search, which
is such a fundamental part of symbolic AI, is based on the ideal Cartesian
method of deduction. This disguises the need to look at how other forms
of problem solving based on intuition (a less prestigious form of reasoning)
could be represented, where a search is not ostensibly part of the process.
In addition, the emphasis of early AI on search was based on a model of
ational, one step at a time, or serial, decision-making, in the same Cartesian
mould. Such a process is excessively deterministic, and even by the admis-
sion of its originator, impossible to achieve. Chapter three describes the
limitations of this model as an ideal as it was based on limited empirical
esearch on individuals making decisions in constrained circumstances.
In the same vein, planning may be thought of as an adjunct to searching
as it involves making the plan of which states to search, while searching
involves actually ca
ying out the plan. Later empirical research on
decision-making shows the extent to which individuals neither search for
solutions nor plan their actions in a path towards a solution. Suchman
(1987) shows the way that individuals react contingently to the situations
in which they find themselves, in interactions with intelligent machines.
They do not plan a serial, rational, step-by-step path to their goal, rather,
they marshal a range of resources in order to deal with a variety of often
unexpected settings. In other words computing systems built on a planning
model tend to confuse people’s plans with their situated actions. Plans
neither act as an adequate reconstruction of situated action nor do they
determine its course.
In the mid-19% programming a computer was no mean feat. Indeed
such was the difficulty that Newell and Simon were led to ‘hand-simulate’
their program before implementing it on a computer. This they did by
giving each of Simon’s wife and children a sub-routine to ‘execute’ when
called upon to do so, an experience which his children apparently never
forgot (Crevier XXXXXXXXXXWe may be struck by the irony of having the bodily
immanence of one’s children simulate the ideal of Cartesian reason in this
way; certainly Simon was lucky to have a big enough and willing enough
family to execute all his sub-routines. Human computers apart, the point
I wish to make is not that there was a deliberate choice to start up the
field of AI with an example which clearly venerated male reason over
female reason; that, for example, pure mathematics was consciously chosen
36
AI IN CONTEXT
instead of knowledge of child rearing or whatever. Rather, I argue that this
kind of problem was the natural choice of workers in the field; an example
of what is taken to be the highest form of reasoning, something that people
find highly abstract and difficult, a masculine standard drawn from their
own lives, which was then to form the subject matter of the first significant
AI program.
Newell’s and Simon’s XXXXXXXXXXlater development of GPS (General Problem
Solver) was produced as an attempt to mimic contextless general problem
solving abilities, in the form of ‘means-ends analysis’, which was derived
from subjects’ think-aloud protocols in solving logic problems, in a series
of psychological experiments. The idea is that human subjects will select
the most appropriate means to satisfy a given end, gradually reducing the
difference between the start and the solution to the given problem, until
the co
ect path is found from the starting position to the answer. Both
GPS and the later Soar system are based upon a highly constrained problem
solving situation, with an artificial and formally defined problem domain,
and with only a rather limited amount of empirical data. It is significant
that these authors extrapolated from such a bounded problem solving
situation to make an important claim about the nature of general problem
solving. Nevertheless GPS is regarded as an important milestone in the
history of AI.
The goal of general problem solving, where the system itself is context-
less, is now seen as overambitious by many AI researchers, yet GPS
does not appear to have attracted substantial criticism for involving this
extrapolation. Its failure has been seen as more of an implementation
problem, and hence necessarily productive