Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP Path: utzoo!mnetor!seismo!rutgers!princeton!mind!harnad From: harnad@mind.UUCP (Stevan Harnad) Newsgroups: net.ai,net.cog-eng Subject: Re: Searle, Turing, Symbols, Categories Message-ID: <167@mind.UUCP> Date: Fri, 10-Oct-86 11:50:33 EDT Article-I.D.: mind.167 Posted: Fri Oct 10 11:50:33 1986 Date-Received: Sat, 11-Oct-86 19:49:08 EDT References: <158@mind.UUCP> <150@cwrudg.UUCP> <160@mind.UUCP> <2495@utai.UUCP> Organization: Cognitive Science, Princeton University Lines: 120 Summary: Doubts about the turing test can be reduced to doubts about empirical science Xref: mnetor net.ai:1183 net.cog-eng:282 In response to my article <160@mind.UUCP>, Daniel R. Simon asks: > 1) To what extent is our discernment of intelligent behaviour > context-dependent?...Might not the robot version [of the > turing test] lead to the...problem of testers being > insufficiently skeptical of a machine with human appearance? > ...Is it ever possible to trust the results of any > instance of the test...? My reply to these questions is quite explicit in the papers in question: The turing test has two components, (i) a formal, empirical one, and (ii) an informal, intuitive one. The formal empirical component (i) is the requirement that the system being tested be able to generate human performance (be it robotic or linguistic). That's the nontrivial burden that will occupy theorists for at least decades to come, as we converge on (what I've called) the "total" turing test -- a model that exhibits all of our robotic and lingistic capacities. The informal, intuitive component (ii) is that the system in question must perform in a way that is indistinguishable from the performance of a person, as judged by a person. It is not always clear which of the two components a sceptic is worrying about. It's usually (ii), because who can quarrel with the principle that a veridical model should have all of our performance capacities? Now the only reply I have for the sceptic about (ii) is that he should remember that he has nothing MORE than that to go on in the case of any other mind than his own. In other words, there is no rational reason for being more sceptical about robots' minds (if we can't tell their performance apart from that of people) than about (other) peoples' minds. The turing test is ALREADY the informal way we contend with the "other-minds" problem [i.e., how can you be sure anyone else but you has a mind, rather than merely acting AS IF it had a mind?], so why should we demand more in the case of robots? It's surely not because of any intuitive or a priori knowledge we have about the FUNCTIONAL basis of our own minds, otherwise we could have put those intuitive ideas to work in designing successful candidates for the turing test long ago. So, since we have absolutely no intuitive idea about the functional (symbolic, nonsymbolic, physical, causal) basis of the mind, our only nonarbitrary basis for discriminating robots from people remains their performance. As to "context," as I argue in the paper, the only one that is ultimately defensible is the "total" turing test, since there is no evidence at all that either capacities or contexts are modular. The degrees of freedom of a successful total-turing model are then reduced to the usual underdetermination of scientific theory by data. (It's always possible to carp at a physicist that his theoretic model of the universe "is turing-indistinguishable from the real one, but how can you be sure it's `really true' of the world?") > 2) Assuming that some "neutral" context can be found... > what does passing (or failing) the Turing test really mean? It means you've successfully modelled the objective observables under investigation. No empirical science can offer more. And the only "neutral" context is the total turing test (which, like all inductive contexts, always has an open end, namely, the everpresent possibility that things could turn out differently tomorrow -- philosophers call this "inductive risk," and all empirical inquiry is vulnerable to it). > 3) ...are there more appropriate means by which we > could evaluate the human-like or intelligent properties of an AI > system? ...is it possible to formulate the qualities that > constitute intelligence in a manner which is more intuitively > satisfying than the standard AI stuff about reasoning, but still > more rigorous than the Turing test? I don't think there's anything more rigorous than the total turing test since, when formulated in the suitably generalized way I describe, it can be seen to be identical to the empirical criterion for all of the objective sciences. Residual doubts about it come from four sources, as far as I can make out, and only one of these is legitimate. The legitimate one (a) is doubts about autonomous symbolic processes (that's what my papers are about). The three illegitimate ones (in my view) are (b) misplaced doubts about underdetermination and inductive risk, (c) misplaced hold-outs for the nervous system, and (d) misplaced hold-outs for consciousness. For (a), read my papers. I've sketched an answer to (b) above. The quick answer to (c) [brain bias] -- apart from the usual structure/function and multiple-realizability arguments in engineering, computer science and biology -- is that as one approaches the asymptotic Total Turing Test, any objective aspect of brain "performance" that anyone believes is relevant -- reaction time, effects of damage, effects of chemicals -- is legitimate performance data too, including microperformance (like pupillary dilation, heart-rate and perhaps even synactic transmission). I believe that sorting out how much of that is really relevant will only amount to the fine-tuning -- the final leg of our trek to theoretic Utopia, with most of the substantive theoretical work already behind us. Finally, my reply to (d) [mind bias] is that holding out for consciousness is a red herring. Either our functional attempts to model performance will indeed "capture" consciousness at some point, or they won't. If we do capture it, the only ones that will ever know for sure that we've succeeded are our robots. If we don't capture it, then we're stuck with a second level of underdetermination -- call it "subjective" underdetermination -- to add to our familiar objective underdetermination (b): Objective underdetermination is the usual underdetermination of objective theories by objective data; i.e., there may be more than one way to skin a cat; we may not happen to have converged on nature's way in any of our theories, and we'll never be able to know for sure. The subjective twist on this is that, apart from this unresolvable uncertainty about whether or not the objective models that fit all of our objective (i.e., intersubjective) observations capture the unobservable basis of everything that is objectively observable, there may be a further unresolvable uncertainty about whether or not they capture the unobservable basis of everything (or anything) that is subjectively observable. AI, robotics and cognitive modeling would do better to learn to live with this uncertainty and put it in context, rather than holding out for the un-do-able, while there's plenty of the do-able to be done. Stevan Harnad princeton!mind!harnad