Arena Magazine No.3 February-March 1993 FEATURES ARTIFICE AND INTELLIGENCE Alan Roberts The media regularly call our attention to the heavy tread of metallic feet: the robots, we are told, are moving in on us. They will be intelligent robots, their micro-processor brains performing billions of calculations each second, with swift access to thousands of billions of pieces of data in their memories. The coming developments in Artificial Intelligence (AI) may mean that, as a species, we have created our successors . . . Will the computers, mobile or immobile, surpass all human skills in the near future or medium future, then? Are craftspeople, for instance, doomed to the fate of the Indian handloom weavers of the last century - will their bones bleach the plains? The answer is No. But before seeing why, we should try to sum up the present status of work in AI. The current achievements of AI are small but promising: its present and coming performance in strictly delimited fields, with expert systems in particular, is even more promising. But AI holds no promise whatsoever for eventually eclipsing human intelligence. To see more support for this last assertion than can be given here, the best introduction is probably two books (What Computers Can't Do, and Mind Over Machine) by the Dreyfus brothers. This dynamic duo comprises two Berkeley professors, one (Hubert) a philosopher and the other (Stuart) in Industrial Engineering and Operations Research. They have been puncturing various AI balloons for a decade or two now - releasing, appropriately enough, large volumes of hot gas. The leading figures in AI research have certainly seen their work as opening up an apparently unlimited range of human activity to the machine, making craft, apparently, mere child's play. Herbert Simon, a Nobel Prize winner, has declared that '...within ten years most theories in psychology will take the form of computer programs'. Thus equipped with the laws of human behaviour, telling it which stimulus will produce a desired response, a computer- directed robot should certainly be able to toss off a striking tapestry or two. Or a million... Simon again: 'Machines will be capable, within twenty years, of doing any work that a man can do'. Another leader, Marvin Minsky from MIT, has predicted that 'within a generation, the problem of creating "artificial intelligence" will be substantially solved'. We can hardly fail to be impressed by the clarity of these statements, and by their courage in laying down definite time limits: within ten years, within twenty years, within a generation. Before declaring ourselves and our species obsolete, however, we should note when these three predictions were made: in 1957, 1965 and 1967 respectively. Not one has been realized. Not one has even come much closer to fulfilment. When the media cover the issue, they tend even now to repeat such claims while observing a tactful silence about the previous fiascos. Often the stories even adhere to the same decades-old format: no, they can't do these things just yet, but there are exciting results that show they have taken the first step, and the really big developments are just around the corner... But having made the first step does not count for much in attaining your goal if the path you are on is in fact a blind alley, and the picture of human intelligence underlying these 'hard AI' claims seems false enough to guarantee such a misdirection. For the detailed arguments against this 'fallacy of the first step', read the Dreyfus material cited above. A critique that is strongly based on recent physics (and is much gentler) has been given by Roger Penrose (in The Emperor's New Mind), who also considers the less extreme ('soft AI') positions, which are much more defensible. The extravagant claims quoted above have been well and truly exploded by the passage of time. But it is only fair to note the more sober picture that emerges if we turn to more recent AI literature, such as found in The Foundations of Artificial Intelligence: A Sourcebook (1990). It is true that it still exhibits hangovers from the Unbounded Optimism era: "All fields discuss the nature of man. AI tries to do something about it. [A] task of AI as a science is to explain human intelligence... What is common, as intelligent agents, between Einstein, the man-on-the-street, the tribesman on a hunt... is that they all face very similar computational problems..." "[A] recent book by Minsky (one of the founders of AI) offers computational models for phenomena as diverse as conflict, pain and pleasure, the self, the soul, consciousness, confusion, genius, infant emotion, foreign accents, and freedom of will..." But fantasy trips disguised as 'science', like these, now face more realistic competition. Other writings in this collection show that many AI workers are now making serious efforts to define the nature and limitations of their field, in the process grappling with fundamental and very difficult questions. A 'BRAIN' WITHOUT A BODY? But what was wrong with the picture of 'intelligence' that guided these past - and, unfortunately, present - trips into fairyland? It is salutary to look at one line of contrary argument, developed particularly by Hubert Dreyfus. If human capabilities are to be equalled by a computer, they must be based on the following of rules; computers can do nothing else. Of course, sometimes we do resort to rules; a learner driver will mutter: first neutral gear, then the ignition key, then the handbrake off ... But once past the novice stage, there is usually no sign that a conscious rule is being followed. To rescue the 'humanoid computer' project, one has to assume that all our skilled behaviour comes from following a rule, whether we are conscious of the rule or not. But this notion of 'unconscious rules' has little to support it, and much against it. As Dreyfus notes: "The important thing about skills is that, although science requires that the skilled performance be described according to rules, these rules need in no way be involved in producing the performance." A computer needs rules because the objects it deals with must be clearly defined, and what it is to do next must be unambiguously written in its programme. The data available to the programme is some well-delimited set chosen because of its relevance to the programme's purpose - it is the programmer, of course, who decides on that purpose, and chooses the data which will be relevant to it. Thus the computer operates in a small, self-contained, relatively unpuzzling world. In contrast, we poor humans have to lumber along in a world capable of infinite novelty and do the best we can with it. We have to make do with objects that we can make out only hazily, tentatively change our criteria of relevance in response to new experience, tentatively impose patterns - patterns that generally defy analysis - in attempting to understand a situation. Aware of such valuable features in the human psyche, AI researchers have naturally shown great interest in developing programmes that can learn. Early work here relied strongly on behaviourist theories, in which human learning is seen largely as a matter of simple stimulus- response conditioning, reinforcement and excitation frequency, all easily programmable. But a body of experimental work has now put a large question mark over such theories, and shown how some of the most elementary conditioning results actually depend on the experimental context and may differ according to the subject's choice. Indeed, and even more remarkably, there is now evidence that non-human and even non-mammalian animals may form patterns and use them to shape their perception of the 'input stimulus' variously, in an altogether similar way. If we look at the experimental findings we will be inclined to ask, not whether a computer could 'learn' like a human being, but rather whether it could ever learn as effectively as a pigeon. We might even ask: will any computer ever have as much common sense as a pigeon? For it has turned out - as the discussion above might well lead us to expect - that common sense is immensely harder to program than logical thought. Grappling with this problem, Minsky writes: "...common sense works so well not because it is an approximation of logic; logic is only a small part of our great accumulation of different, useful ways to chain things together." Demurring from Minsky's proposal to program these various 'chaining' methods, Terry Winograd observes in Thinking Machines: "The rules followed by the machine can deal only with the symbols, not their interpretation... There are basic limits to what can be done with symbol manipulation, regardless of how many 'different, useful ways to chain things together' one invents. The reduction of mind to the interactive sum of decontextualized fragments is ultimately impossible and misleading." THE BODY: COSTS AND BENEFITS But if our skills cannot generally be reduced to rules, where do they come from? Dreyfus emphasizes in his account how they originate in, and continually depend upon, bodily experience. It is because we are embodied that we have learnt - that we have had to learn - how to make sense out of experiences that are different in quality and yet simultaneous, what our eyes tell us and what our hands tell us, for example. Originally, it is to negotiate our way through the physical world that we form patterns with which to organize it, and expectations based upon these patterns. Equipped with these patterns and expectations, we can pass most of the time in a world that is now familiar; and yet the patterns and expectations have this extraordinary virtue of remaining open to correction, so that new phenomena do not necessarily leave us floundering. Thus we semi-automate our responses, so that the sensory world does not present itself as an arena we must permanently struggle to understand, and we can cope without a constant drain on our energy and attention. If any human quality deserves the name of 'intelligence', it is hard to think of a better candidate than what we thus start to develop: the ability to cope. As Dreyfus puts it in What Computers Can't Do: "...[A]n embodied agent can dwell in the world in such a way as to avoid the infinite task of formalizing everything... these global forms of recognition are not open to the digital computer, which, lacking a body, cannot respond as a whole but must build up its recognition starting with determinate details.." Thus, in the situations where automated response is all that is required - where objects are well-defined, the body of 'relevant' data is clearly circumscribed, patterns do not need adapting, and expectations are never disappointed - we can just coast along using programmes of response derived from past experience. It is in such fields of activity, where creative and complex thought is not demanded, that formal logic and the computer can thrive quite well. But such cases are rarer than generally believed. For example, most people would probably see mathematics as the archetype of the kingdoms where formal logic holds sway; but this is often because they are thinking of an unrepresentative case, that of arithmetic. In mathematics generally, it is far different. As Penrose (himself a mathematician) writes in The Emperor's New Mind: "People might suppose that a mathematical proof is conceived as a logical progression, where each step follows upon the ones that have preceded it. Yet the conception of a new argument is hardly likely actually to proceed in this way. There is a globality and seemingly vague conceptual content that is necessary in the construction of a mathematical argument..." (Of course, the argument can always be reconstructed as a logical progression.) What may seem a paradoxical conclusion emerges from this: a major reason why the computer will never duplicate human intelligence is that it has no body with which to meet the world. This is where we come to some interesting consequences for the computerizing of skills, craft skills in particular. CRAFT, SKILL AND 'VIRTUAL REALITY' Craft relies - perhaps more heavily than 'high art' - on those crucial features of human understanding that derive from and depend upon bodily experience: how materials resist or yield, what they feel like - the staggeringly complex integration of sensory data and instinctual needs that is implied by a phrase like 'security blanket'. A computer might well be programmed to control machines that duplicate or clumsily plagiarize an existing craft work, just as colour photocopiers can be expected eventually to turn out remarkably similar copies of the Mona Lisa. (According to press reports, they are already supplying passable imitations [literally!] of a fifty-dollar note.) But no original pieces of creative craft need ever be expected to roll off the numerically controlled production line. (It is worth observing that the output from such a production line need not be monotonously uniform. Numerical control can easily avoid the perfect sphere or the rigidly straight line, adding random deviations that are perceptible to the eye or hand. If there is enough demand, those who just want bumps for the sake of bumpiness will have factories catering to their needs.) It might be pertinent to suggest a special significance for craft in today's world. That world is one where 'virtual reality' threatens to substitute for reality. Face-to-face human encounters are increasingly undermined and constricted as they are replaced by mediated and abstract relationships; 'technocratic rationality', economic efficiency, the inevitability of 'progress' are assumed to justify changes that are often dubious, unwelcome to most people and even horrendous. If 'high art' celebrates such developments, or even presents them neutrally, it risks losing its expected critical edge. And if we reject the ideology which insists that such a daunting world is 'inevitable', we might detect, and be glad to see, a critique at least tacit in craft work which revives older experiences now in danger of being lost: the feel of a blanket, the weight of stone - not in 'virtual reality', but in one's hand. Of course, such a critique can be either reactionary or forward-looking. It may convey a naive desire for the past to be re-created just as it was - for the lost virtues to emerge in the same social surroundings as they once did, for people to be close inside the (patriarchal) extended family or the (narrow-minded and exclusionary) village community. Or, on the other hand, it might express a desire to recover what is lost or endangered but to situate it in a framework of more expansively convivial relationships. That these can be envisaged only dimly might be regarded as no excuse to abandon the hope for them. We might draw a salutary lesson from all this. Even if 'technological progress' has been made into an ideology, it is certainly not a pure myth and we need to be alert to those of its promises and threats which happen to be real; but we also need to keep a sharp lookout for 'scientific' con-jobs. Recall how, in the field of nuclear power, the vision of 'electricity too cheap to meter' was dangled before us to silence the doubters. Remember how freeways were going to solve the problems of expanding road traffic. And once upon a time, the benevolent scientific genie was about to wave his wand and reduce the working week to twenty hours, to ten hours... The claims of AI researchers above are probably extreme examples of such a ploy, and they are by no means innocent. After considering the social role of AI in general, Terry Winograd - one of the more enlightened leaders in the field - offers a disturbing conclusion (his own italics): "[T]he techniques of artificial intelligence are to the mind what bureaucracy is to human social interaction." The alleged march of the robots gives this bureaucratic nightmare its intimidating clincher. If we are all nearing our use-by date anyway, due for replacement by smarter, more durable beings made from steel, carbon fibre and plastic, then obviously there is little point in struggling to preserve the merely human... Actually, if we felt unkind, we could say that those who peddle such daydreams and nightmares are refuting their own claims in the very act of making them. Surely they are demonstrating that no machine will ever out-do homo sapiens in the specifically human skill of wanking. Alan Roberts, originally a physicist, now researches in theoretical ecology at Monash University.This article is based on an paper to the craft conference 'Interventions', July 1992. An earlier version appeared in the Winter 1992 issue of Artlink. ----------------------------------------------------------- Arena Magazine is published six times a year by: Arena Printing and Publishing Pty Ltd 35 Argyle Street, Fitzroy, 3065, Australia. Email: . The material in Arena Magazine is copyright. Permission is given to republish articles, in either electronic or paper form, so long as: (1) it is for non-profit purposes; (2) that the text of all work remains intact (including this copyright notice); and (3) notification is sent to Arena Publishing, either by snail or email. Applications from commercial publishers will be considered.