- In the ballet Coppelia, a dancer mimics a clockwork dancing doll simulating a dancer. The imitating movements, dancing twice removed, are predictably "mechanical," given the discrepancies of outward resemblance between clockwork dancers and real ones. These discrepancies may diminish to zero with the technological progress of clockwork, until a dancer mimicking a clockwork dancer simulating a dancer may present a spectacle of three indiscernible dancers engaged in a pas de trois. By behavioral criteria, nothing would enable us to identify which is the doll, and the lingering question of whether the clockwork doll is really dancing or only seeming to seems merely verbal - unless we adopt a criterion of meaning much favored by behaviorism that makes the question itself nonsensical.
- The question of whether machines instantiate mental predicates has been cast in much the same terms since Turing, and by tacit appeal to outward indiscernibility the question of whether machines understand is either dissolved or trivialized. It is in part a protest against assimilating the meaning of mental predicates to mere behavioral criteria - an assimilation of which Abelson and Schank are clearly guilty, making them behaviorists despite themselves - that animates Searle's effort to mimic a clockwork thinker simulating understanding; to the degree that he instantiates the same program it does and fails to understand what is understood by those whom the machine is designed to simulate - even if the output of the three of them cannot be discriminated - then the machine itself fails to understand. The argumentation is picturesque, and may not be compelling for those resolved to define (such terms as) "understanding" by outward criteria. So I shall recast Searle's thesis in logical terms which must force his opponents either to concede machines do not understand or else, in order to maintain they might understand, to abandon the essentially behaviorist theory of meaning for mental predicates.
- Consider, as does Searle, a language one does not understand but that one can in a limited sense be said to read. Thus I cannot read Greek with understanding, but I know the Greek letters and their associated phonetic values, and am able to pronounce Greek words. Milton's daughters were able to read aloud to their blind father from Greek, Latin, and Hebrew texts though they had no idea what they were saying. And they could, as can I, answer certain questions about Greek words, if only how many letters there are, what their names are, and how they sound when voiced. Briefly, in terms of the distinction logicians draw between the use and mention of a term, they knew, as I know, such properties of Greek words as may be identified by someone who is unable to use Greek words in Greek sentences. Let us designate these as M-properties, in contrast to U-properties, the latter being those properties one must know in order to use Greek (or any) words. The question then is whether a machine programmed to simulate understanding is restricted to M-properties, that is, whether the program is such that the machine cannot use the words it otherwise may be said to manipulate under M-rules and M-laws. If so, the machine exercises its powers over what we can recognize in the words of a language we do not understand, without, as it were, thinking in that language. There is some evidence that in fact the machine operates pretty much by pattern recognition, much in the manner of Milton's unhappy daughters.
- Now I shall suppose it granted that we cannot define the U-properties of words exhaustively through their M-properties. If this is true, Schank's machines, restricted to M-properties, cannot think in the languages they simulate thinking in. One can ask whether it is possible for the machines to exhibit the output they do exhibit if all they have is M-competence. If not, then they must have some sort of U-competence. But the difficulty with putting the question thus is that there are two ways in which the output can be appreciated: as showing understanding or as only seeming to, and as such the structure of the problem is of a piece with the structure of the mind-body problem in the following respect. Whatever outward behavior, even of a human being, we would want to describe with a psychological (or mental) predicate - say that the action of raising an arm was performed - has a physical description that is true whether or not the psychological description is true - for example, that the arm went up. The physical description then underdetermines the distinction between bodily movements and actions, or between actions and bodily movements that exactly resemble them. So whatever outward behavior takes a (psychological) Ψ-predicate takes a (physical) Φ-predicate that underdetermines whether the former is true or false of what the latter is true of. So we cannot infer from a Φ-description whether or not a Ψ-description applies. To be sure, we can ruthlessly define Ψ-terms as Φ-terms, in which case the inference is easy but trivial, but then we cannot any longer, as Schank and Abelson wish to do, explain outward behavior with such concepts as understanding. In any case, the distinction between M-properties and U-properties is exactly parallel: anything by way of output we would be prepared to describe in U-terms has an M-description true of it, which underdetermines whether the U-description is true or not.
- So no pattern of outputs entails that language is being used, nor hence that the source of the output understands, inasmuch as it may have been cleverly designed to emit a pattern exhaustively describable in M-terms. The problem is perfectly Cartesian. We may worry about whether any of our fellows is an automaton. The question is whether the Schank machine (SAM) is so programmed that only M-properties apply to its output. Then, however closely (exactly) it simulates what someone with understanding would show in his behavior, not one step has been taken toward constructing a machine that understands. And Searle is really right. For while U-competence cannot be defined in M-terms, an M-specified simulation can be given of any U-performance, however protracted and intricate. The simulator will only show, not have, the properties of the U-performance. The performances may be indiscriminable, but one constitutes a use of language only if that which emits it in fact uses language. But it cannot be said, to use language if its program, as it were, is written solely in M-terms.
- The principles on the basis of which a user of language structures a story or text are so different from the principles on the basis of which one could predict, from certain M-properties, what further M-properties to expect, that even if the outputs are indiscernible, the principles must be discernible. And to just the degree that they deviate does a program employing the latter sorts of principles fail to simulate the principles employed in understanding stories or texts. The degree of deviation determines the degree to which the strong claims of AI are false. This is all the more the case if the M-principles are not to be augmented with U-principles.
- Any of us can predict what sounds a person may make when he answers certain questions that he understands, but that is because we understand where he is going. If we had to develop the ability to predict sounds only on the basis of other sounds, we might attain an astounding congruence with what our performance would have been if we knew what was going on. Even if no one could tell we didn't, understanding would be nil. On the other hand, the question remains as to whether the Schank machine uses words. If it does, Searle has failed as a simulator of something that does not simulate but genuinely possesses understanding. If he is right, there is a pretty consequence. M-properties yield, as it were, pictures of words: and machines, if they encode propositions, do so pictorially.
Text Colour Conventions (see disclaimer)
- Blue: Text by me; © Theo Todman, 2020
- Mauve: Text by correspondent(s) or other author(s); © the author(s)