INTRODUCTION

The artificial intelligence debate come back regularly since the origin of this notion. Anyone can obviously ascend to animals-machines, or to literary myths, since antiquity, about artificial human beings or automaton. But since the fifties, this debate about an artificial intelligence (AI) takes place more specifically, because of the computers existence. While, forty years ago, only some theorists dared this philosophical digression about early computers, from the eighties, the generalization of machines has given the opportunity to spread this technical and philosophical problem to any aware or curious user. Finally, in the nineties, the tangible confrontation with machines affects all areas of activities. This new situation commits us to provide a satisfactory representation in reply to questions that do not fail to appear at consciousness or at users’ imagination. Therefore, it can be legitimate to tell another version of the story than the AI opponents’ one.

Computer science has for long been quasi exclusively technical, and rather isolated in air-conditioned offices. This situation could have seemed to dispense providing an intellectual framework of its justification, which currently belongs to mathematical theory. And the former DP men took the easy way out by maintaining the mystery for those depending on them, especially through some waiting periods, commanding respect or resentment. Something like a distant technocratic image has resulted of these practices, strengthened by the well-known excuse for civil servants: “It’s the computer’s fault!"

From the very beginning, this lack of mastering has given vent to fantasy representations, helped by literary or movies forgeries of the old universal myths, from Frankenstein to disaster movies. The basic argument is the dehumanization or the loss of contact with nature, supposedly direct by the past. The philosophical grounding of these myths can link them to this romanticist lineage, which represents their systematization and coherence. But their relevance, or at least the monopoly preference of these representations, can be discussed, although the literary exploitation of ignorance and fear of science’s repercussions is prevailing today. In fact, the alternative to AI adversaries is quasi exclusively represented by the science-fiction author, Isaac Asimov, with his saga about robots.

Is the resistance at the idea of an artificial intelligence only coming from these literary representations, which essentially aim at reserving intelligence to human beings, until considering artificial intelligence as a contradiction in terms? Indeed, the machine presence can be considered from a simple utilitarian point of view that would hardly have any repercussions. Albeit technical extensions are known to modify consciousness or human being nature, as admitted by some AI’s opponents, often asserting simultaneously the contrary. But those who are practicing data processing, and those whom the prospect of amazing realizations is lured at, both can see that the mechanized part of human activity no longer concerns a motor quality (like transportation), or a technique quality (like automation in general, by opposition at craftsmanship). This mechanized human property, knowledge, even thought therefore, strike at a most fundamental part. And this is all the question of artificial intelligence.

AI’s Opponents

This book will especially study the critiques of artificial intelligence arose by some representative authors, not to say canonical ones. It will often oppose less to their particular arguments than to their conclusion of AI impossibility. While most often, its possibility is concluded from the same reasons [NOTE 1].

More specifically, the debate has taken its current form with the Hubert L. Dreyfus’ critical study, systematic enough in its time, What Computers Can’t Do, A Critique of Artificial Reason (1972, 1979, and 1992). He personifies the champion of the opposition to artificial intelligence. His phenomenological critique has met the public’s acclaim, with the DP men or computer scientists’ blessings, influenced by their school training or by this fashionable philosophical trend.

We can equally retain as significant the standpoint of a research worker in artificial intelligence, Joseph Weizenbaum, that is gone over to the opposition with his Computer Power and Human Reason (1976). He became famous with his AI program ELIZA, which simulated a consultation with a psychotherapist. His objections to artificial intelligence are more classically moralistic, and anti-technoscientific. And the regression of his references on a religious standpoint illustrates very well the difficulty to think modern world with adequate (or merely updated) categories.

A third notorious opposition belongs to the philosopher of language, John Searle, whose book, Minds, Brains and Science, (Conference Reith 1984 of the BBC), has summarized his standpoint. This book argues, philosophically, with most of linguists’ resistances to artificial intelligence. The particularity of this philosophical standpoint consists obviously in turning effective difficulties in formalization of language into ontological impossibilities. The linguistic standpoint is obviously fundamental in artificial intelligence, since it is concerned with categorization of knowledge, the mostly possible in natural language. As quarrels within linguistics chapels add fuel to the flames, this opposition at AI makes way for dogmatism, while, on principle, automatic processing would allow to decide between different formalizations. Wouldn’t it precisely the reason why this resistance is setting up?

It had seemed useful to point out too the Terry Winograd and Fernando Flores’ work, Understanding Computer and Cognition (1986). The first of them, again one of the research workers of origins of AI, had realized a still notorious program, SHRDLU, that allowed to communicate with machine to make it move blocks in an environment. He also has gone over to the enemy for the sake of his phenomenological convictions. The second one, former economy minister of the Salvador Allende’s government before the 1973’s putsch in Chile, especially discusses information systems. And he is questioning the possibility of computerization of human world too, for the sake of the same post-Heideggerian philosophical ideas.

These four works allow us to cover a field of arguments large enough of the opposition at artificial intelligence. Each of them develops more specifically an aspect of the critique only outlined by others. But they are also useful to check, one against another, some partisan excess, some mistakes, or simply to reveal limitations of their philosophical standpoint. Many further opponents usually repeat these very theses [NOTE 2]. They represent the foundations of philosophical vulgate to apprehend the AI new phenomenon and even the data processing in general [NOTE 3].

The influence of the critique of AI, especially because of the authority of the dominant philosophy (even more in France), demanded therefore an answer the less possible respectful of academic conformism. And, as AI practical results are not accepted, this debate has to be held in a general ideas level. This book is elsewhere born in reaction to a conference, organized by ARC (Association of Cognitive Research, France), in which a psychoanalyst argued against AI for the sake of incommensurability of subjectivity. Besides (or because of) his stage fright in front of a meeting of computer scientists and knowledge engineers whom he could consider like the incarnation of the evil, his labored statement (or lesson) were placed exclusively on the ground of canonical authority of psychoanalysis or phenomenology. But at any time, no one could have been able to see an expression of the so-called individuality he had reminded, except through the automatic releasing of an academic defense mechanism!

It is also necessary to notice that these arguments provided by this critiques of AI depreciate the research programs of this field with decision‑makers or public. All of those who are not involved in it can be too happy to have a ready-made catalogue to avoid to be interested in an area they ignore, and make them anxious. From an alternate point of view, the work of partisans of AI, Edward Feigenbaum and Pamela McCorduck, The Fifth Generation, argued easily on anti-Japanese topic of international scientific competition. More reasonably, it can be admitted that if a country does not begin any researches, it will be late to whine if others obtain results first. But this is the same with all other economic sectors.

With this opportunity, the Dreyfus’ confession about the origin of his interest for AI can be reminded. His students of M.I.T. had made him notice that, in laboratories where they were working, old philosophical problems were mechanically solved, while philosophy itself just go on about the same old things. But Dreyfus should have been considering his own bad mood as a warning: if research workers in human sciences are not interested in these questions, those they consider as vulgar technicians are soon going to exceed them. This would not be the first time that something like this happen in history.

Assumed Debate

This book is also explicitly directed at compensate for slackness that reigns in the academic world, including the concerned field of AI. This systematic avoidance of the comparison has already been noted by one of the first advocate of artificial intelligence:

“Newell and Simon, who are Dreyfus’ prime targets, concluded that any formal rebuttal would only propagate Dreyfus further. […] During the period immediately after ‘Alchemy and Artificial Intelligence’ [NOTE 4] appeared, people often raised Dreyfus’ charges during the question-and-answer periods following lectures Newell and Simon gave, and so they actually prepared answers. But aside from that, and from Herb Simon’s open letter to the Sigart Bulletin, the two have refused to answer Dreyfus directly.” (McCorduck, Machines who think, p. 202).

On the contrary, Dreyfus has not abandoned his undermining activity against AI, which he has made a kind of specialty. He has even just republished his work, in 1992, with the title – ironic rather than descriptive, due to fact it is almost identical – What Computers STILL Can’t Do. A Critique of Artificial Reason.

We will therefore face up the debate, often praised as indispensable, but hindered by a lounging concealed under the industrious excuse of the sacrosanct pre-eminence of the research work:

“Papert’s response is an unfinished document, largely because Papert lost interest in it, feeling that with a finite life it was more interesting things to do in science than to defend against what he considered as an intellectually irresponsible attack.” (McCorduck, Machines who think, p. 203).

On the contrary, it is possible to consider that answering questions is part of science, and anyone can take it on. It remains true whether these questions are asked by colleagues from the same field, or by colleagues from other fields. One of the constraints of science should be not to neglect a minimum overall coherence, at least for not undermining its authority with the public. Ignorant people, elsewhere, have the same prerogatives. Scientists have to submit to this demand of answer: either for pedagogical or democratic motives, due to public financing of the research programs, or to security constraints of self-financing technology, or for their own security, like in the case of geneticists’ elimination by Lysenko’s partisans under Staline. Finally, it is necessary to dissipate false ideas on AI, to provide some right ones, because the classic production of myths or fantasy cannot be prevented anyway.

The Artificial Intelligence Problem

The AI field, and its critique, concerns currently a multitude of scientific areas (algorithmic, logic, linguistics, psychology, philosophy, sociology, robotics, etc.) and it is not circumscribed with precision. That, obviously, is inciting to think that it does not concern an autonomous area, even that it does not concerns anything scientific at all… Besides, an up-to-date technology can be made fit every occasion (which is not specific). More, the question of human intelligence itself is far from being solved, which distort all possibility of comparison. But anyone can specifically distinguish several issues of this question, or its controversy.

Formally speaking, AI is one consequence of the development of logic. Questions to which AI is linked to are those of reasoning, automatic demonstration, calculability. An immediate interest of current computer sciences is to find a real application to logic, because it is the fact these old logic lessons drown into bluntness or vacuity – judiciously noted by the French 17th century Port-Royal logicians – of the scholastic way of practicing Aristotelian logic. We can integrate to this formal area the question of human intelligence itself, through cognitive psychology. From this point of view, it is undeniable that formal automatable operations are already existing, which represent some aspects of human reasoning. How can we understand the meaning of a negation of this reality? But then the principle of these questions is independent from the existence of the machine. Calculations can be made in writing, or orally in the discussion.

From a technical point of view, AI results from evolution of calculus machine, or automata, since the first model of calculator developed by Pascal, through the machine of Babbage, until Hollerith’s punch card machines (origin of IBM) and the first valve computers. Here, we can speak about an automation of logic and the AI adversaries’ resistances plainly concern opposition to technique or science.

Programs usually considered as actual AI are often about human language processing or understanding, and of what derives from them, when interacting with a computer. A superior stage is the reproduction of reasoning or expert systems [NOTE 5], even pattern recognition or other specialized performances. Here, resistances rather concern the limitations or the fragmentary aspect of these realizations. We can however admit intermediate stages, contrary to purists, linguists or philosophers. Who can ask a science, especially at its beginning, to solve all problems by a single overall theory, and by an instantaneous general application? Nevertheless, we will see that philosophical holistic practice does not hesitate in denying technical realizations, or at challenging each step of progress of AI concerning knowledge.

Finally, an essential issue of AI (or its debate) concerns data, or meaning, processed by the machine. The essential interest of data processing, management data processing generally (but AI too), consists precisely in processing symbolic data. Information thus concern human activities first, quantitative or qualitative data hasn’t any importance from this point of view: the Hollerith’s machine had been built to process data from the American census; and any databases really include information about characteristics of the real world of human activities, either professional or private. While systems or organization analysis give shape to contents of a human context, through computerization process; and new programming techniques (called object programming), or expert systems, allow us to handle knowledge in term of categories close to those of human reasoning – aiming to provide a natural language human-machine interaction.

Generality of Artificial Intelligence

One of the foundations of this book, contrary to what one could expect, is a demystification of AI too, for it is considered as a generalization of the traditional data processing. This way of thinking is integrating one of the aspects of the skepticism about AI, as expressed by Philippe Kahn, former chairman of Borland, who satirized once by telling that AI is used when programs do not work, elsewhere it is software. Precisely, from the beginning, the classic data processing is equivalent to what is called universal machine. From this initial point of view, it simply opposes to specialized machines (like a loom – first machine of this kind –, or like a washing machine, an elevator, a pocket calculator…). The computer is working by loading in its RAM (Random Access Memory, i.e. erasable) any programs it needs, that it finds in its ROM (Read-only Memory, i.e. permanent) or in hard disks, diskettes, magnetic bands, or by telecommunication (formerly on punched cards). AI does not bring anything to this technical point of view, except with neural systems, which add to the analogy with human beings, but which can be simulated on traditional computer, on the same universal machine principle.

Technical characteristics are causing sometimes a few confusions, often self-interested. As Winograd and Flores emphasized, the name fifth generation, used sometimes to characterize AI, is rather concerning microprocessors. The first generation was made of vacuum tubes (in 1940), the second one of transistors (in 1950), the third one of integrated circuits (in 1960), and the VLSI (Very Large Scale Integration) one integrates several tens of thousands of components on a chip, and the fifth current generation includes millions. But we can also consider another thing that the simple quantitative accumulation, by attracting the attention on wired logical operation itself. Actually, microprocessors (CISC or RISC [NOTE 6]) only represent the hardware version of programs (soft). In the string of programming languages building, we can already observe the progressiveness to closer human communication:

First generation languages, the binary coding at the beginning of the data processing, were equivalent at the maximal operation breaking down, close to the level of components of microprocessors. Programmers were speaking then the language of machine. The principle was the direct application of the machine of Turing that undertakes simple displacement of signs in locations of the memory, and defines admissible transformations.

Second generation languages, assemblers, were merely equivalent to the translation of the former string of binary digits (0000101010101) in codes more easily memorable (ADD for the addition, etc.), first avoiding keystroke mistakes. We can already consider the part of the program that attends to translate assembler in machine language as an improvement of human/machine communication. But each assembler still concerns only one type of microprocessor.

Third generation languages, FORTRAN, COBOL, C, ALGOL, BASIC, etc., are also an improvement of communication. These new instructions are equivalent at summaries of string of instructions in assembler. The elementary instructions string of an addition in assembler (“LD x, ADD y, STO z,” loading a register, using the binary addition rules, stocking the result) really shows, in a third language generation, a correct legibility (“A = B + C”). The program code is portable as well on any microprocessor.

Fourth generation languages (like SQL – Structured Query Language), more specifically concern the processing in database. Their instructions can express strings of sub-statements more explicitly (SELECT “customers” WITH “orders” > 10000) to find records in a file, to select them according to some criteria, to modify them, or to output forms on printer, etc. Automatic software engineering, that we can classify at this level too, allows to built databases, which directly represent an information system, or the reality to formalize.

Fifth generation languages, actual artificial intelligence ones like LISP or PROLOG, or “objects extensions” (functioning in term of classes, etc.) from the two previous generations languages, thus allow to represent knowledge (in theory) in declarative form (roughly: “grandparent = parent of parent” in PROLOG). The inference engine included in the language itself allows then calculating if consequences can be deduced of questions asked to the system. And some systems (called non-monotonic) can modify themselves their own knowledge, in interaction with the environment. The user is then less a programmer than a part of this environment exchanging information with the machine.

According to this principle, artificial intelligence represents a new step in front of the traditional, or imperative, programming, in which the programmer has to describe precisely the algorithm; i.e. each steps describing processing (the operations order of a recipe is the best example of it). AI has developed the approach of a declarative programming, consisting in providing structured information using definitions. Linguistic analysis is the best example of the declarative form:

sentence = nominal group + verbal group

nominal group = determiner + noun

nominal group = determiner + adjective + noun, etc.

We can then sum up AI as the capacity for a program, i.e. for inference rules and data, to find all alone the best way to solve a problem. Designers of the first machines, or those having formalized their theoretical possibility, have very rapidly realized they could, recursively, give to programs the possibility to consider themselves as a datum. That represents all the artificial intelligence essence.

Thus, anyone can easily extrapolate by imagining it will be possible someday to communicate directly with the machine in human language (more or less simplified). In that purpose, it is enough to add, like in previous cases, a new layer used as intermediary between human language (called natural language in the computer world) and the operative code of the microprocessor. It is necessary not to forget the actual working way of computers: each level is permanently translated in the inferior or superior level. AI should thus first allow avoiding the learning of an artificial language, whose main function was to suppress ambiguities. Nowadays, even in a simple dialogue with a database provide often-insufficient answer, by default, or by excess. While a cooperative answer (more or less in natural language) allows to answer to bad questions (by proposing alternatives), or to consider the context (by considering previous questions).

Finally, the critique of AI is grounded on a kind of inversion of the optimistic, self‑congratulating, or frankly advertising discourses, prevailing in data processing or computer science. In this technological field, a kind of enthusiasm or exaltation is found, linked to quasi-daily announcements of progress. It is true that the multiplier coefficient can really give dizziness: only for microcomputers since 1981, the speed of processing has grown from 7 MHz (megahertz) to 166 MHz (1995) for the professional standard machine; the RAM memory has grown from 64 Ko (thousands of characters) to 8 Mo (millions). At the beginning of microcomputers, there were only diskettes containing 360 Ko as storage, without hard disk; in 1995, they contain respectively, 1,44 Mo (100 Mo for some back up systems) and 1 Go (a billion characters) for the hard disk [NOTE 7].

In this context, it is conceivable that some partisans of AI, rather limited at the beginning of the field, could consider, right or wrong, that their problems would be solved later. As much that in this field, the changing rhythm happens every 12 or 18 month approximately, when the depletion of a cycle of production is five years in traditional industries (and a career is thirty-five years for academic or professional qualifications). A retrospective look could make even think today that dissatisfaction should have been larger yesterday, with so limited means.

Resistances to AI seem to be equivalent to the Zenon’s paradoxes: never the arrow will reach the target, Achilles never will rejoin the tortoise, and the machine will never be as intelligent as human. Even for some DP men, computer scientists, or AI research workers, resistances appear sometimes insurmountable. Even those working in one of its divisions can easily use AI as a whole as Aunt Sally. A bias, maybe sectarianism for their sub-specialty, can make them disparage others.

The classic step of the automation has concerned manual work. The program of the loom already was a list of instructions that told a machine what to do. The main interest of the computing has been to apply these instructions to symbolic or digital data, while being able to trigger instructions to peripheral devices (machines, printers, screens, etc.). The step in question with artificial intelligence is merely concerning automation of intellectual capacities, whatever level of complexity they are. The resistance to this automation is equivalent to others. It is only more easily heard, for it concerns jobs that deal with arguments and ideas.

Notes