harnad@mind.UUCP (Stevan Harnad) (10/27/86)
[Until the problem of follow-up articles to mod.ai through Usenet is straightened out, I'm temporarily responding to mod.ai on net.ai.] In mod.ai, in Message-ID: <8610270723.AA05463@ucbvax.Berkeley.EDU>, under the subject heading THE PSEUDOMATH OF THE TURING TEST, PADIN@FNALB.BITNET writes: > DEFINE THE SET Q={question1,question2,...}. LETS NOTE THAT > FOR EACH q IN Q, THERE IS AN INFINITE NUMBER OF RESPONSES (THE > RESPONSES NEED NOT BE RELEVANT TO THE QUESTION, THEY JUST NEED TO BE > RESPONSES). IN FACT, WE CAN DEFINE A SET R={EVERY POSSIBLE RESPONSE TO > ANY QUESTION}, i.e., R={r1,r2,r3,...}. Do pseudomath and you're likely to generate pseudoproblems. Nevertheless, this way of formulating it does inadvertently illustrate quite clearly why the symbolic version of the turing test is inadequate and the robotic version is to be preferred. The symbolic version is equivalent to the proverbial monkey's chances of typing Shakespeare by combinatorics. The robotic version (pending the last word on basic continuity/discontinuity in microphysics) is then no more or less of a combinatorial problem than Newtonian Mechanics. [Concerning continuity/discreteness, join the ongoing discussion on the A/D distinction that's just started up in net/mod.ai.] > THE EXISTENCE OF ...A FUNCTION T THAT MAPS A QUESTION q TO A SET > OF RESPONSES RR... FOR ALL QUESTIONS q IS EVIDENCE FOR THE PRESENCE > OF MIND SINCE T CHOOSES, OUT OF AN INFINITE NUMBER OF RESPONSES, > THOSE RESPONSES THAT ARE APPROPRIATE TO AN ENTITY WITH A MIND. Pare off the pseudomath about "choosing among infinities" and you just get a restatement of the basic intuition behind the turing test: That an entity has a mind if it acts indistinguishably from an entity with a mind. > NOW A PROBLEM [arises]: WHO IS TO DECIDE WHICH SUBSET OF RESPONSES > INDICATES THE EXISTENCE OF MIND? WHO WILL DECIDE WHICH SET IS > APPROPRIATE TO INDICATE AN ENTITY OTHER THAN OURSELVES IS OUT THERE > RESPONDING? The same one who decides in ongoing, everyday "solutions" to the other-minds problem. And on exactly the same basis: indistinguishability of performance. > [If] WE GET A RESPONSE WHICH APPEARS TO BE RANDOM, IT WOULD SEEM THAT > THIS WOULD BE SUFFICIENT TO LABEL [the] RESPONDENT A MINDLESS ENTITY. > HOWEVER, IT IS THE EXACT RESPONSE ONE WOULD EXPECT OF A SCHIZOPHRENIC. When will this tired prima facie objection (about schizophrenia, retardation, aphasia, coma, etc.) at last be laid to rest? Damaged humans inherit the benefit of the doubt from what we know about their biological origins AND about the success of their normal counterparts in passing the turing test. Moreover, there is no problem in principle with subhuman or nonhuman performance -- in practice we turing-test animals too -- and this too is probably parasitic on our intuitions about normal human beings (although the evolutionary order was probably vice versa). Also, schizophrenics don't just behave randomly; if a candidate just behaved randomly it would not only justifiably flunk the turing test, but it would not survive either. (I don't even know what behaving purely randomly might mean; it seems to me the molecules would never make it through embryogeny...) On the other hand, which of us doesn't occasionally behave randomly, and some more often than other?. We can hardly expect the turing test to provide us with the criteria for extreme conditions such as brain death if even biologists have problems with that. All these exotic variants are pseudoproblems and red herrings, especially when we are nowhere in our progress in developing a system that can give the normal version of the turing test a run for its money. > NOW IF WE ARE TO USE OUR JUDGEMENT IN DETERMINING THE PRESENCE OF > ANOTHER MIND, THEN WE MUST ACCEPT THE POSSIBILITY OF ERROR INHERENT > IN THE HUMAN DECISION MAKING PROCESS. AT BEST,THEN, THE TURING TEST > WILL BE ABLE TO GIVE US ONLY A HINT AT THE PRESENCE OF ANOTHER MIND; > A LEVEL OF PROBABILITY. What else is new? Even the theories of theoretical physics are only true with high probability. There is no mathematical proof that our inferences are entailed with necessity by the data. This is called "underdetermination" and "inductive risk," and it is endemic to all empirical inquiry. But besides that, the turing test has even a second layer of underdermination that verges on indeterminacy. I have argued that it has two components: One is the formal theorist's task of developing a device that can generate all of our performance capacities, i.e.,one that can pass the Total Turing Test. So far, with only "performance capacity" having been mentioned, the level of underdetermination is that of ordinary science (it may have missed some future performance capacity, or it may fail tomorrow, or it may just happen to accomplish the same performance in a radically different way, just as the universe may happen to differ from our best physical theory). The second component of the turing test, however, is informal, intuitive and open-ended, and it's the one we usually have in mind when we speak of the turing test: Will a normal human being be able to tell the candidate apart from someone with a mind? The argument is that turing-indistinguishability of (total) performance is the only basis for making that judgment in any case. Fallible? Of course that kind of judgment is fallible. Certainly no less fallible than ordinary scientific inference; and (I argue) no more fallible than our judgments about other minds. What more can one ask? Apart from the necessary truths of mathematics, the only other candidate for a nonprobabilistic certainty is our direct ("incorrigible") awareness of our OWN minds (although even there the details seem a bit murky...). Stevan Harnad {allegra, bellcore, seismo, packard} !princeton!mind!harnad harnad%mind@princeton.csnet (609)-921-7771