Artificial intelligence is the attempt to make machines that think like humans. In this article the abbreviation AI will be used to refer to this enterprise, and the phrase artificial intelligence technologies (AIT) will be reserved for new technologies that initially grew out of AI but that mimic only some aspects of human abilities, such as a subset of speech recognition or pattern recognition, while avoiding the deep problems. AIT and AI are often confused with one another.
AI proper has an important bearing on sociology in general, and social studies of science in particular, because of the light it can shed on the notion of the social. Most sociologists believe that most of a person's capacities are gained through the person's embedding in social groups. If machines could succeed in mimicking human reasoning, then either humans would have learned to "socialize" machines or there would be something wrong with the idea of "the social." At the moment, humans have no idea how to socialize machines; there are no machines that can be raised from birth and learn language within a family, nor any that are imprinted with a ready-made set of social abilities and a capacity to continue to build them through social interaction. From time to time such abilities have been claimed for machines as they have evolved. For example, neural nets appear to be capable of learning by themselves, but only in a crude, behavioristic way, as one might train a pigeon or the like, so the deep problem of socialization has not been approached. This means that any real successes in AI would threaten the sociologist's idea of the social.
This is not the only kind of relationship between artifical intelligence and the social sciences. Sociologists are interested in the way new technologies change society, and the changes brought about by AIT are one such area of inquiry. One might think of machines of all sorts as already an integral part of society, but this is to use the notion of "the social" in a way that bears less directly on sociology as an enterprise. Also, social studies of science have a legitimate concern with the development of AITs of various kinds, and especially their relationship to military projects.
Returning to AI proper, the attempt to automate scientific discovery is a hard case for those who would wish to maintain the idea of the social. At the heart of the sociology of scientific knowledge (SSK) is the idea that even scientific knowledge is deeply invested with the social, whereas dominant models of science take it to be a paradigm of universality divorced from social influence. Thus, imagine a human community that has developed in isolation from other communities: There would be no grounds to expect such a community to develop, say, the English language, still less the nuances of any particular dialect spoken at a given moment in history. One accepts that such capacities would not develop in the absence of social contact between the community and the social group that embodied the dialect. On the other hand, the dominant view of science would lead us to be less surprised should such an isolated community rediscover many of our scientific and mathematical laws. It is this view of science that is challenged by SSK, which treats any body of scientific knowledge as very much like a dialect in a natural language. If the current generation of asocial machines could rediscover scientific laws on their own, this would support the dominant view and challenge the sociological view of science. The sociology of scientific knowledge is, then, a hard case for the larger argument about AI and the social. If the sociologists can hold their ground in the case of science, then the ground can be held much more easily in the case of social activity as a whole.
Attempts to develop AI can be seen, then, as an expensive experiment to test the deep ideas of sociology in general and of SSK in particular. Like all experiments, however, this one suffers from indeterminacies in its outcome associated with the experimenter's regress and the like. One important source of confusion is the confusion between AIT and AI, which is amplified by humans' ability to "repair" the deficiencies in others' communication and attribute far more competence to partners in discourse that they deserve. The need for repair is crucial to ordinary communication, because speech is normally indistinct, broken, overlaid with other sounds, and invested with allusion to shared but unspoken contexts. It is only by "reading" within context and repairing the "mistakes" that we are able to make sense of others' speech and action. On these tendencies depends the success of confidence tricksters and fraudsters of various kinds, who can rely on the "mark: to do most of the work necessary to see what they do as a competent performance.
Thus, supposed tests of artifical intelligence as they are normally designed, such as versions of the Turing test, which are used in much-publicized computer challenges, trade on humans' ability and willingness to repair the broken discourse of computers, whereas the computers deal with only complete or near-complete language input. A fairer test would involve more symmetry between what the computer and the human have to accomplish in the way of context-dependent understanding. The asymmetry that continues to exist is revealed in attempts to build speech transcribers and such devices, which continue to founder on the inability of machines to understand contexts that they could grasp only through immersion in society. To provide a simple example, the following is a true statement: "My spell checker will correct weerd processor but not weird processor." This is a difficulty for my spell checker, but not as big a difficulty as the fact that in the context of this essay, both spellings should remain uncorrected (which the human editor, understanding the context, immediately sees). Careful attention to the deficiencies of automated speech-processing devices, from spell checkers to speech transcribers, shows that the world is still a safe place for the idea of the social as conceived by sociology as a whole.
Turning back to the hard case of science, the issue is more complex and turns on a careful separation of what is being done by the humans and what by the machines (Collins and Kusch talk of polymorphic and mimeomorphic actions). For example, a pocket calculator appears to be a competent arithmetic machine, though the sociologist of science would like to say that even arithmetic is a social activity. We can resolve the problem by pointing out that arithmetic consists of a mechanical part—something that can be likened to stacking pebbles—and a part that fits this kind of activity into human life including, say, approximation. The number of decimal points retained in a calculation depends on whether we are converting, say, a person's height from one kind of unit to another or calculating the course of a spacecraft, and these continually changing subtleties are, once more, like the subtleties of dialect. The problem is visible even inside arithmetic in the buildup of errors within calculations. Thus, simple calculators will usually make a mistake at about 7.0000/11.0000 × 11.0000, and are even more likely to get it wrong if the computation is done in two stages. (Though some calculators will get this simple calculation right, the deeper problem of automation of arithmetic recurs in more complex tasks.)
Computers can be programmed to discover systematic relationships between variables given a database, and it has been said that this means that scientific discovery is an individual process. Once more this is easy to understand so long as one thinks of discovery as the mechanical part of science (as the spellings of words in a dictionary are the mechanical part of language). But before any computer can do this, it has to separate data from noise, and in science this is a socially approved judgment. In the same way, the output of such discovering computers must be filtered. Not just any relationship is scientifically interesting, or science would be no more than "data dredging." Once more, a close examination of supposed independent discovery machines shows that they depend on humans to filter their output in the same way as a modern spell checker generates a number of potential corrections but leaves it to the human to make the appropriate choice. In sum, even the hard case of science turns out to support the sociologists' view; the supposed discovering machines can help, but not replace, the humans in the scientific process.
The continuing development of various generations of intelligent machines continues to refine our understanding of the social, but it comes no closer to resolving its deep problems.
H. M. Collins
Collins, Harry M., and Martin Kusch. The Shape of Actions: What Humans and Machines Can Do. Cambridge, MA: MIT Press, 1998.
Holton, Gerald. The Scientific Imagination: Case Studies. Cambridge, U.K.: Cambridge University Press, 1978.
Kuhn, Thomas S. "The Function of Measurement in Modern Physical Science." ISIS 52 (1961): pp.162-176.
Langley, Pat, Herbert A. Simon, Gary L. Bradshaw, and Jan M. Zytkow. Scientific Discovery: Computational Explorations of the Creative Process. Cambridge, MA: MIT Press, 1987.
Maurer, David W. The Big Con: The Story of the Confidence Man and the Confidence Game. New York: Bobbs Merrill, 1940.
From Science, Technology, and Society
