Accept no imitations

Given that he’s just become a movie star, Alan Turing’s classic paper “Computing Machinery and Intelligence” seems an apt topic for a blog post.  It is in this paper that Turing sets out his famous “Imitation Game,” which has since come to be known as the Turing Test.  The basic idea is as follows: Suppose a human interrogator converses via a keyboard and monitor with two participants, one a human being and one a machine, each of whom is in a different room.  The interrogator’s job is to figure out which is which.  Could the machine be programmed in such a way that the interrogator could not determine from the conversation which is the human being and which the machine?  Turing proposed this as a useful stand-in for the question “Can machines think?”  And in his view, a “Yes” answer to the former question is as good as a “Yes” answer to the latter.

This way of putting things is significant.  Turing doesn’t exactly assert flatly in the paper that machines can think, or that conversational behavior of the sort imagined entails intelligence, though he certainly gives the impression that that is what he believes.  (As Jack Copeland notes in his recent book on Turing (at p. 209), Turing’s various statements on this subject are not entirely consistent.  In some places he explicitly declines to offer any definition of thinking, while at other times he speaks as if studying what machines do can help us to discover what thinking is.)  What Turing says in the paper is that the question “Can machines think?” is “too meaningless to deserve discussion,” that to consider instead whether a machine could pass the Turing Test is to entertain a “more accurate form of the question,” and that if machines develop to the point where they can pass the test, then “the use of words and general educated opinion will have altered so much that one will be able to speak of machines thinking without expecting to be contradicted.”

This is very curious.  Suppose you asked me whether gold and pyrite are the same, and I responded by saying that the question is “too meaningless to deserve discussion,” that it would be “more accurate” to ask whether we could process pyrite in such a way that someone examining it would be unable to tell it apart from gold, and that if we can so process it, then “the use of words and general educated opinion will have altered so much that one will be able to speak of pyrite as gold without expecting to be contradicted.”  Obviously this would be a bizarre response.  Whether pyrite might be taken by someone to be gold and whether pyrite is in fact gold are just two different questions, and what I would be doing is simply changing the subject rather than in any way answering the original question.  How is Turing’s procedure any different?  And how exactly is “Can machines think?” any more “meaningless” than “Is pyrite gold?”

It’s no good, by the way, to object that the cases are not parallel insofar as an expert could distinguish gold and pyrite.  The cases are parallel in this respect, as Turing himself implicitly admitted.  Copeland points out (p. 211) that Turing elsewhere acknowledged that in a Turing Test situation, someone with expertise about machines might well be able to figure out from subtle clues which is the machine.  Turing thus stipulated that the interrogator should be someone who does not have such expertise.  He thought that what mattered was whether the ordinary person could figure out which is the machine.  So, whether an expert (as opposed to an ordinary observer) could figure out whether or not something is pyrite does not keep my example from being relevantly analogous to Turing’s.

So, why might Turing or anyone else think that his proposed test casts any light on the question about whether machines can think?  There are at least three possible answers, and none of them is any good.  I’ll call them the Scholastic answer, the verificationist answer, and the scientistic answer.  Let’s consider each in turn.

What I call the “Scholastic answer” is definitely not what Turing himself had in mind, though in fact it would be the most promising (if ultimately unsuccessful) way to try to defend Turing’s procedure.  The idea is this.  Recall that it is a basic principle of Scholastic metaphysics that agere sequitur esse (“action follows being” or “activity follows existence”).  That is to say, the way a thing acts or behaves reflects what it is.  A defender of the Turing Test could argue that if a machine acts like an intelligent thing, then it must be an intelligent thing.  But competent language use is a paradigmatically intelligent activity (especially for a Scholastic, who would define intellect in terms of the grasp of abstract concepts of the sort expressed by general terms).  Hence (so the argument might go) the Turing Test is a surefire way to test for intelligence.

But not so fast.  For a Scholastic, the principle agere sequitur esse must, of course, be applied in conjunction with other basic metaphysical principles.  And one of the other relevant ones is the distinction between substantial form and accidental form, a mark of which is the presence or absence of irreducible causal powers.  A plant carries out photosynthesis and a pocket watch displays the time of day, but these causal powers are not in the two objects in the same way.  That a plant carries out photosynthesis is an observer-independent fact about the plant, whereas that a watch displays the time of day is not an observer-independent fact about the watch.  For the metal bits that make up the watch have no inherent tendency to display the time.  That is a function we have imposed on them, from outside as it were.  The plant, by contrast, does have an inherent tendency to carry out photosynthesis.  That reflects the fact that to be a plant is to have a substantial form and thus to be a true substance, whereas to be a pocket watch is to have a mere accidental form and not to be a true substance.  The true substances in that case are the metal bits that make up the watch, and the form of a pocket watch is just an accidental form we have imposed on them.  (I have discussed the difference between substantial and accidental form in many places, such as here, here, and here.  For the full story, see chapter 3 of Scholastic Metaphysics.) 

Now, a computing machine is like a pocket watch rather than like a plant.  It runs the programs it does, engages in conversation, etc. in just the same way that the watch displays the time.  That is to say, it has no inherent tendency to do these things, but does them only insofar as we impose these functions on the parts that make up the machine.  (This is why, as Saul Kripke points out, there is no observer-independent fact of the matter about what program a computer is running, and why, as Karl Popper and John Searle point out, there is no observer-independent fact of the matter about whether something even counts as a computer in the first place.)  To be a computer is to have a mere accidental form rather than a substantial form.

In applying the principle agere sequitur esse, then, we need to determine whether the thing we’re applying it to is a true substance or not, or in other words whether it has a substantial form or merely an accidental form.  If we’re examining bits of metal and find that they display the time, it would silly to conclude “Well, since agere sequitur esse, it follows that metal bits have the power to tell time!”  For the bits are “telling time” only because we have made them do so, and they wouldn’t be doing it otherwise.  Similarly, if I throw a stone in the air, it would be ridiculous to conclude “Since agere sequitur esse, it follows that stones can fly!”  The stone is “flying” only because and insofar as I throw it.  Flying is, you might say, merely an accidental form of the stone.  What matters when applying the principle agere sequitur esse is to see what a thing does naturally, on its own, when left to its own devices -- that is to say, to see what properties flow or follow from its substantial form, as opposed to the accidental forms that are imposed upon it.

Now, seen in this light the Turing Test is just a non-starter.  To determine whether a machine can think, it simply isn’t relevant to find out whether it passes the Turing Test, if it passes the test only because it has been programmed to do so.  Left to themselves, metal bits don’t display time, and stones don’t fly.  And left to themselves, machines don’t converse.  So, that we can make them converse no more shows that they are intelligent than throwing stones or making watches shows that stones have the power of flight or that bits of metal qua metal can tell time.

So, while the Scholastic answer would (in my view, since I’m a Scholastic) be Turing’s best bet, at the end of the day it doesn’t really work.  But of course, Turing was no Scholastic.  Did he have in mind instead what I call the “Verificationist answer”?  The idea here would be this: The meaning of a statement is, according to verificationism, determined by its method of verification.  Now, we can’t peer into anyone else’s mind, in the case of human beings any more than in the case of machines.  So (the argument might continue), the only way to verify whether something is intelligent is to determine whether it behaves in an intelligent way, and intelligent conversation is the gold standard of intelligent behavior.  Hence the only way the question “Can machines think?” can be given a meaningful construal is to interpret it as asking whether machines can behave in an intelligent way.  Since that is precisely what the Turing Test seeks to determine, if a machine passes it, then there is nothing more that could in principle be asked for as evidence that it is genuinely intelligent.  Indeed (so the argument would go), there is nothing more for intelligence to be than the capacity to pass the Turing Test.

Now, verificationism was certainly in the air at the time Turing was writing.  It underlay the “philosophical behaviorist” view that having a mind is “nothing but” manifesting certain patterns of behavior or dispositions for behavior.  But there are serious problems with verificationism, not the least of which is that it is self-defeating.  For the principle of verification is not itself verifiable, which entails that it is, by its own standards, strictly meaningless.  If it were true, then it wouldn’t even rise to the level of being false.  Unsurprisingly, no one defends it any more, at least not in its most straightforward form.

But Turing does not in any case appeal to verificationism in the paper, and I don’t think that’s really what’s going on.  What I think he was at least tacitly committed to is what I call the “Scientistic answer” to the question of why anyone should think the Turing Test casts light on the question whether machines can think.  Turing’s view, I suspect, was essentially that there is no way to study intelligence scientifically other than by asking what a system would have to be like in order to pass the Turing Test.  Hence that is, in his view, the question we should focus on.  Notice that this is not (or need not) be the same position as that of the verificationist.  His talk about “meaninglessness” notwithstanding, Turing need not say that it is strictly meaningless to ask whether something could pass the Turing Test and yet not truly be thinking.  He could say merely that since there is no scientific way to investigate that particular question, there is no point in bothering with it, and we should just focus instead on what the methods of the empirical scientist might shed light on.

If this is what Turing is up to, then he is essentially doing the same thing Lawrence Krauss does when he pretends to answer the famous question why there is anything at all rather than nothing.  And what Krauss does, as I have discussed several times (here, here, here, and here), is to pull a bait-and-switch.  He pretends at first that he is going to explain why there is something rather than nothing, but then changes the subject and discusses instead the question of how the universe in its current state arose from empty space together with the laws of physics -- which, of course, are very far from being nothing.  His justification for this farcical procedure is essentially that physics has something to tell us about the latter question, whereas it has nothing to tell us about why there is anything at all (including the fundamental laws of physics themselves) rather than nothing.  What we should focus on, in Krauss’s view, is the question he thinks he can answer rather than the question we originally asked.

Now this is exactly the same fallacy as that of the drunk who insists on looking for his lost car keys under the lamp post, on the grounds that that is the only place where there is enough light by which to see them.  The fact that that is where the light is simply doesn’t entail that the keys are there, and neither does it entail that there is any point in continuing to look for the keys under the lamp post after repeated investigation fails to turn them up, or that there is no point in trying to find ways to look for the keys elsewhere, or that we should look for something else under the lamp post rather than the keys.  Similarly, the fact that the methods of physics are powerful methods doesn’t entail that those methods can answer the question why there is anything at all rather than nothing, or that we should replace that question with some other question that the methods of physics can handle, or that there is no point in looking for other methods by which to investigate the question.  To assume, as Krauss does, that the question simply must be one susceptible of investigation by physics if it is to be rationally investigated at all is to commit what E. A. Burtt identified as the fallacy of “mak[ing] a metaphysics out of [one’s] method” -- that is, of trying to force reality to conform to one’s favored method of studying it rather than conforming one’s method to reality. 

Turing seems to be guilty of the same thing.  Rather than first determining what thought is and then asking what methods might be suitable for studying something of that nature, he instead starts by asking what sorts of thought-related phenomena might be susceptible of study via the methods of empirical science, and then decides that those are the only phenomena worth studying.  The fallaciousness of this procedure should be obvious.  Characterizing “thought” as the kind of thing that a machine would exhibit by virtue of passing the Turing Test is like characterizing “keys” as the sort of thing apt to be found under such-and-such a particular lamp post.

In general, there is (as I have argued many times) simply no good reason to accept scientism and decisive reason to reject it.  There are at least five problems with it: First, formulations of scientism are typically either self-defeating or only trivially true; second, science cannot in principle offer a complete description even of the physical world; third, science cannot even in principle offer a complete explanation of the phenomena it describes; fourth, the chief argument for scientism -- the argument from the predictive and technological successes of science -- is fallacious; and fifth, the widespread assumption that the only alternative to natural science is a dubious method of doing “conceptual analysis” is false.  (See chapter 0 of Scholastic Metaphysics for detailed exposition of each of these points.)  So, the “Scientistic answer” also fails.

Needless to say, Turing was a brilliant scientist, and all of us who use and love computers are in his debt.  But his foray into philosophy resulted, I think, not in any positive contribution but only in an interesting and instructive mistake.