next up previous
Next: A Summary of Polemical Up: AN UNREASONABLE BOOK Previous: What Does He Say

In Defense of the Unjustly Attacked-Some of whom are Innocent

Here are defenses of Weizenbaum's targets. They are not guaranteed to entirely suit the defendees.

Weizenbaum's conjecture that the Defense Department supports speech recognition research in order to be able to snoop on telephone conversations is biased, baseless, false, and seems motivated by political malice. The committee of scientists that proposed the project advanced quite different considerations, and the high officials who made the final decisions are not ogres. Anyway their other responsibilities leave them no time for complicated and devious considerations. I put this one first, because I think the failure of many scientists to defend the Defense Department against attacks they know are unjustified, is unjust in itself, and furthermore has harmed the country.

Weizenbaum doubts that computer speech recognition will have cost-effective applications beyond snooping on phone conversations. He also says, ``There is no question in my mind that there is no pressing human problem that will be more easily solved because such machines exist''. I worry more about whether the programs can be made to work before the sponsor loses patience. Once they work, costs will come down. Winograd pointed out to me that many possible household applications of computers may not be feasible without some computer speech recognition. One needs to think both about how to solve recognized problems technological possibilities to good use. The telephone was not invented by a committee considering already identified problems of communication.

Referring to Psychology Today as a cafeteria simply excites the snobbery of those who would like to consider their psychological knowledge to be above the popular level. So far as I know, professional and academic psychologists welcome the opportunity offered by Psychology Today to explain their ideas to a wide public. They might even buy a cut-down version of Weizenbaum's book if he asks them nicely. Hmm, they might even buy this review.

Weizenbaum has invented a New York Times Data Bank different from the one operated by The New York Times - and possibly better. The real one stores abstracts written by humans and doesn't use the tapes intended for typesetting machines. As a result the user has access only to abstracts and cannot search on features of the stories themselves, i.e. he is at the mercy of what the abstractors thought was important at the time.

Using computer programs as psychotherapists, as Colby proposed, would be moral if it would cure people. Unfortunately, computer science isn't up to it, and maybe the psychiatrists aren't either.

I agree with Minsky in criticizing the reluctance of art theorists to develop formal theories. George Birkhoff's formal theory was probably wrong, but he shouldn't have been criticized for trying. The problem seems very difficult to me, and I have made no significant progress in responding to a challenge from Arthur Koestler to tell how a computer program might make or even recognize jokes. Perhaps some reader of this review might have more success.

There is a whole chapter attacking ``compulsive computer programmers'' or ``hackers''. This mythical beast lives in the computer laboratory, is an expert on all the ins and outs of the time-sharing system, elaborates the time-sharing system with arcane features that he never documents, and is always changing the system before he even fixes the bugs in the previous version. All these vices exist, but I can't think of any individual who combines them, and people generally outgrow them. As a laboratory director, I have to protect the interests of people who program only part time against tendencies to over-complicate the facilities. People who spend all their time programming and who exchange information by word of mouth sometimes have to be pressed to make proper writeups. The other side of the issue is that we professors of computer science sometimes lose our ability to write actual computer programs through lack of practice and envy younger people who can spend full time in the laboratory. The phenomenon is well known in other sciences and in other human activities.

Weizenbaum attacks the Yale computer linguist, Roger Schank, as follows - the inner quotes are from Schank: ``What is contributed when it is asserted that `there exists a conceptual base that is interlingual, onto which linguistic structures in a given language map during the understanding process and out of which such structures are created during generation [of linguistic utterances]'? Nothing at all. For the term `conceptual base' could perfectly well be replaced by the word `something'. And who could argue with that so-transformed statement?'' Weizenbaum goes on to say that the real scientific problem ``remains as untouched as ever''. On the next page he says that unless the ``Schank-like scheme'' understood the sentence ``Will you come to dinner with me this evening?'' to mean ``a shy young man's desperate longing for love'', then the sense in which the system ``understands'' is "about as weak as the sense in which ELIZA ``understood''. This good example raises interesting issues and seems to call for some distinctions. Full understanding of the sentence indeed results in knowing about the young man's desire for love, but it would seem that there is a useful lesser level of understanding in which the machine would know only that he would like her to come to dinner.gif

Contrast Weizenbaum's demanding, more-human-than-thou attitude to Schank and Winograd with his respectful and even obsequious attitude to Chomsky. We have ``The linguist's first task is therefore to write grammars, that is, sets of rules, of particular languages, grammars capable of characterizing all and only the grammatically admissible sentences of those languages, and then to postulate principles from which crucial features of all such grammars can be deduced. That set of principles would then constitute a universal grammar. Chomsky's hypothesis is, to put it another way, that the rules of such a universal grammar would constitute a kind of projective description of important aspects of the human mind.'' There is nothing here demanding that the universal grammar take into account the young man's desire for love. As far as I can see, Chomsky is just as much a rationalist as we artificial intelligentsia.

Chomsky's goal of a universal grammar and Schank's goal of a conceptual base are similar, except that Schank's ideas are further developed, and the performance of his students' programs can be compared with reality. I think they will require drastic revision and may not be on the right track at all, but then I am pursuing a rather different line of research concerning how to represent the basic facts that an intelligent being must know about the world. My idea is to start from epistemology rather than from language, regarding their linguistic representation as secondary. This approach has proved difficult, has attracted few practitioners, and has led to few computer programs, but I still think it's right.

Weizenbaum approves of the Chomsky school's haughty attitude towards Schank, Winograd and other AI based language researchers. On page 184, he states, ``many linguists, for example, Noam Chomsky, believe that enough thinking about language remains to be done to occupy them usefully for yet a little while, and that any effort to convert their present theories into computer models would, if attempted by the people best qualified, be a diversion from the main task. And they rightly see no point to spending any of their energies studying the work of the hackers.''

This brings the chapter on ``compulsive computer programmers'' alias ``hackers'' into a sharper focus. Chomsky's latest book Reflections on Language makes no reference to the work of Winograd, Schank, Charniak, Wilks, Bobrow or William Woods to name only a few of those who have developed large computer systems that work with natural language and who write papers on the semantics of natural language. The actual young computer programmers who call themselves hackers and who come closest to meeting Weizenbaum's description don't write papers on natural language. So it seems that the hackers whose work need not be studied are Winograd, Schank, et. al. who are professors and senior scientists. The Chomsky school may be embarassed by the fact that it has only recently arrived at the conclusion that the semantics of natural language is more fundamental than its syntax, while AI based researchers have been pursuing this line for fifteen years.

The outside observer should be aware that to some extent this is a pillow fight within M.I.T. Chomsky and Halle are not to be dislodged from M.I.T. and neither is Minsky - whose students have pioneered the AI approach to natural language. Schank is quite secure at Yale. Weizenbaum also has tenure. However, some assistant professorships in linguistics may be at stake, especially at M.I.T.

Allen Newell and Herbert Simon are criticized for being overoptimistic and are considered morally defective for attempting to describe humans as difference-reducing machines. Simon's view that the human is a simple system in a complex environment is singled out for attack. In my opinion, they were overoptimistic, because their GPS model on which they put their bets wasn't good enough. Maybe Newell's current production system models will work out better. As to whether human mental structure will eventually turn out to be simple, I vacillate but incline to the view that it will turn out to be one of the most complex biological phenomena.

I regard Forrester's models as incapable of taking into account qualitative changes, and the world models they have built as defective even in their own terms, because they leave out saturation-of-demand effects that cannot be discovered by curve-fitting as long as a system is only rate-of-expansion limited. Moreover, I don't accept his claim that his models are better suited than the unaided mind in "interpreting how social systems behave", but Weizenbaum's sarcasm on page 246 is unconvincing. He quotes Forrester, ``[desirable modes of behavior of the social system] seem to be possible only if we have a good understanding of the system dynamics and are willing to endure the self-discipline and pressures that must accompany the desirable mode' ''. Weizenbaum comments, ``There is undoubtedly some interpretation of the words `system' and `dynamics' which would lend a benign meaning to this observation''. Sorry, but it looks ok to me provided one is suitably critical of Forrester's proposed social goals and the possibility of making the necessary assumptions and putting them into his models.

Skinner's behaviorism that refuses to assign reality to people's internal state seems wrong to me, but we can't call him immoral for trying to convince us of what he thinks is true.

Weizenbaum quotes Edward Fredkin, former director of Project MAC, and the late Warren McCulloch of M.I.T. without giving their names. pp. 241 and 240. Perhaps he thinks a few puzzles will make the book more interesting, and this is so. Fredkin's plea for research in automatic programming seems to overestimate the extent to which our society currently relies on computers for decisions. It also overestimates the ability of the faculty of a particular university to control the uses to which technology will be put, and it underestimates the difficulty of making knowledge based systems of practical use. Weizenbaum is correct in pointing out that Fredkin doesn't mention the existence of genuine conflicts in society, but only the new left sloganeering elsewhere in the book gives a hint as to what he thinks they are and how he proposes to resolve them.

As for the quotation from (McCulloch 1956), Minsky tells me ``this is a brave attempt to find a dignified sense of freedom within the psychological determinism morass''. Probably this can be done better now, but Weizenbaum wrongly implies that McCulloch's 1956 effort is to his moral discredit.

Finally, Weizenbaum attributes to me two statements - both from oral presentations - which I cannot verify. One of them is ``The only reason we have not yet succeeded in simulating every aspect of the real world is that we have been lacking a sufficiently powerful logical calculus. I am working on that problem''. This statement doesn't express my present opinion or my opinion in 1973 when I am alleged to have expressed it in a debate, and no-one has been able to find it in the video-tape of the debate.

We can't simulate ``every aspect of the real world'', because the initial state information is never available, the laws of motion are imperfectly known, and the calculations for a simulation are too extensive. Moreover, simulation wouldn't necessarily answer our questions. Instead, we must find out how to represent in the memory of a computer the information about the real world that is actually available to a machine or organism with given sensory capability, and also how to represent a means of drawing those useful conclusions about the effects of courses of action that can be correctly inferred from the attainable information. Having a sufficiently powerful logical calculus is an important part of this problem--but one of the easier parts.

[Note added September 1976 - This statement has been quoted in a large fraction of the reviews of Weizenbaum's book (e.g. in Datamation and Nature) as an example of the arrogance of the "artificial intelligentsia". Weizenbaum firmly insisted that he heard it in the Lighthill debate and cited his notes as corroboration, but later admitted (in Datamation) after reviewing the tape that he didn't, but claimed I must have said it in some other debate. I am confident I didn't say it, because it contradicts views I have held and repeatedly stated since 1959. My present conjecture is that Weizenbaum heard me say something on the importance of formalization, couldn't quite remember what, and quoted "what McCarthy must have said" based on his own misunderstanding of the relation between computer modeling and formalization. (His two chapters on computers show no awareness of the difference between declarative and procedural knowledge or of the discussions in the AI literature of their respective roles). Needless to say, the repeated citation by reviewers of a pompous statement that I never made and which is in opposition to the view that I think represents my major contribution to AI - is very offensive].

The second quotation from me is the rhetorical question, ``What do judges know that we cannot tell a computer''. I'll stand on that if we make it ``eventually tell'' and especially if we require that it be something that one human can reliably teach another.


next up previous
Next: A Summary of Polemical Up: AN UNREASONABLE BOOK Previous: What Does He Say

John McCarthy
Tue Oct 17 20:28:09 PDT 2000