Path: utzoo!attcan!uunet!seismo!sundc!pitstop!sun!quintus!ok From: ok@quintus.UUCP (Richard A. O'Keefe) Newsgroups: comp.ai Subject: AI and Sociology Keywords: light, research programmes, flame-free Message-ID: <1033@cresswell.quintus.UUCP> Date: 28 May 88 08:13:06 GMT Organization: Quintus Computer Systems, Mountain View, CA Lines: 153 I believe it was Gilbert Cockton who raised the question "why does AI ignore Sociology?" Two kinds of answers have been given so far: (1) AI is appallingly arrogant trash. (2) Sociology is appallingly arrogant trash. I want to suggest that there is a straightforward reason for the mutual indifference (some Sociologists have taken AI research as _subject_ material, but AI ideas are not commonly adopted by Sociologists) which is creditable to both disciplines. Lakatos's view of a science is that it is not so much a set of theories as a research programme: a way of deciding what questions to pursue, what counts as an explanation, and a way of dealing with puzzles. For example, he points out that Newton's theory of gravity is in principle unfalsifiable, and that the content of the theory may be seen in the kinds of explanations people try to come up with to show that apparent exceptions to the theory are not real exceptions. The key step here is deciding what to study. Both in its application to Robotics and its application to Cognitive Science, AI is about the mental processes of individuals. As a methodological basis, Sociology looks for explanations in terms of social conditions and "forces", and rejects "mentalistic" explanations. Let me provide a concrete example. One topic of interest in AI is how a program could make "scientific discoveries". AM and Eurisko are famous. A friend with whom I have lost contact was working on a program to try to predict the kinetics of gas phase reactions. Pat Langley's "BACON" programs are well known. Scientific discovery is also of interest to Sociology. One book on this topic (the only one on my shelves at the moment) is The social basis of scientific discoveries Augustine Brannigan Cambridge University Press, 1981 0 521 28163 6 I take this as an *example*. I do not claim that this is all there is to Sociology, or that all Sociologists would agree with it, or that all Sociological study is like this. All I can really claim is that I am interested in scientific discovery from an AI point of view, and when I went looking for Sociological background this is the kind of thing I found. Brannigan spends chapter 2 attacking some specific "mentalistic" accounts of scientific discovery, and in chapter 3 rubbishes the mentalistic approach completely. If I understand him, his major complaint is that accounts such as Koestler's "bisociation" fail to be accounts of *scientific* *discovery*. Indeed, a section of chapter 3 is headed "Mentalistic models confuse learning with discovery." It turns out that he is not concerned with the question "how do scientific discoveries happen", but with the question "what gets CALLED a scientific discovery, and why?" Which is a very interesting question, but ignores everything about scientific discovery which is of interest to AI people. The very reason that AI people are interested in scientific discovery (apart from immediately practical motives) is that it is a form of learning in semi-formalised domains. If one of Pat Langley's programs discovers something that happens not to be true (such as coming up with Phlogiston instead of Oxygen) he is quite happy as long as human scientists might have made the same mistake. As I read Brannigan's critical comments on the "mentalistic" theories he was rubbishing, I started to get excited, seeing how some of the suggestions might be programmable. Page 35 of Brannigan: "... in the social or behavioural sciences we tend to obfuscate the social significance of familiar phenomena by explaining them in terms of 'underlying' causes. Though this is not always the case, it is true with discovery and learning." This is to reject in principle attempts to explain discovery and learning in terms of underlying causes. "... the equivalence of learning and discovery is a _confusion_. From a social perspective, 'to _learn_' means something quite different from 'to _discover_'." Emphasis his. He would classify a rediscovery as a mere learning, which at the outset rejects as uninteresting precisely the aspects that AI is interested in. Something which is rather shocking from an AI perspective is found on page 64: "... the hallmark of this understanding is the ascription of learning to some innate abilities of the individual. Common sensically, learning is measured by the degree of success that one experiences in performing certain novel tasks and recalling certain past events. Mackay's ethnographic work suggests, on the contrary, that learning consists in the institutional asciprtion of success whereby certain ordered and identified as learning achievements to the exclusion of other meaningful performances." Page 66: "Although as folk members of society we automatically interpret individual discovery or learning as the outcome of a motivated course of inference, sociologically we must consider the cognitive and empirical grounds in terms of which such an achievement is figured. From this position, cleverness in school is understood, not as a function of innate mental powers, but as a function of the context in which the achievements associated with cleverness are made accountable and remarkable." To put it bluntly, if we take statements made by some AI people or some Sociologists at face value, they cast serious doubts on the sanity of the speakers. But taken as announcements of a research programme to be followed within the discipline, they make sense. AI says "minds can be modelled by machines", which is, on the face of it, crazy. But if we read this as "we propose to study the aspects of mind which can be modelled by machines, and as a working assumption will suppose that all of them can", it makes sense, and is not anti-human. Note that any AI practicioner's claim that the mechanisability of mind is a discovery of AI is false, that is an *assumption* of AI. You can't prove something by assuming it! Sociology says "humans are almost indefinitely plastic and are controlled by social context rather than psychological or genetic factors", which is, on the face of it, crazy. But if we read this as "we propose to study the influence of the social context on human behaviour, and as a working assumption will suppose that all human behaviour can be explained this way", it makes sense, and is not as anti-human as it at first appears. Note that any Sociologist's claim that determination by social forces is a discovery of Sociology is false, that is an *assumption* of Sociology. Both research programmes make sense and both are interesting. However, they make incompatible decisions about what counts as interesting and what counts as an explanation. So for AI to ignore the results of Sociology is no more surprising and no more culpable than for carpenters to ignore Musicology (both have some sort of relevance to violins, but they are interested in different aspects). What triggered this message at this particular date rather than next week was an article by Gilbert Cockton in comp.ai.digest, in which he said "But perhaps this is what AI is doing, trying to improve our understanding of ourselves. But it may not do this because of (2) it forbids something that is, any approach, any insight, which does not have a computable expression. This, for me, is anathema to academic liberal traditions ..." But of course AI does no such thing. It merely announces that computational approaches to the understanding are part of _its_ territory, and that non-computational approaches are not. AI doesn't say that a Sociologist can't explain learning (away) as a function of the social context, only that when he does so he isn't doing AI. A while back I sent a message in which I cited "Plans and Situated Actions" as an example of some overlap between AI and Sociology. Another example can be found in chapter 7 of Induction -- Processes of Inference, Learning, and Discovery Holland, Holyoak, Nisbett, and Thagard MIT Press, 1986 0-262-08160-1 Perhaps we could have some other specific examples to show why AI should or should not pay attention to Sociology?