ABSTRACT: Turing's celebrated 1950 paper proposes a very general methodological criterion for modelling mental function: total functional equivalence and indistinguishability. His criterion gives rise to a hierarchy of Turing Tests, from subtotal ("toy") fragments of our functions (t1), to total symbolic (pen-pal) function (T2 -- the standard Turing Test), to total external sensorimotor (robotic) function (T3), to total internal microfunction (T4), to total indistinguishability in every empirically discernible respect (T5). This is a "reverse-engineering" hierarchy of (decreasing) empirical underdetermination of the theory by the data. Level t1 is clearly too underdetermined, T2 is vulnerable to a counterexample (Searle's Chinese Room Argument), and T4 and T5 are arbitrarily overdetermined. Hence T3 is the appropriate target level for cognitive science. When it is reached, however, there will still remain more unanswerable questions than when Physics reaches its Grand Unified Theory of Everything (GUTE), because of the mind/body problem and the other-minds problem, both of which are inherent in this empirical domain, even though Turing hardly mentions them.

KEYWORDS: cognitive neuroscience, cognitive science, computation, computationalism, consciousness, dynamical systems, epiphenomenalism, intelligence, machines, mental models, mind/body problem, other minds problem, philosophy of science, qualia, reverse engineering, robotics, Searle, symbol grounding, theory of mind, thinking,Turing, underdetermination, Zombies.

1. Intended Interpretation

William Wimsatt (1954) originally proposed the notion of the "Intentional Fallacy" with poetry in mind, but his intended message applies equally to all forms of intended interpretation. It is a fallacy, according to Wimsatt, to see or seek the meaning of a poem exclusively or even primarily in the intentions of its author: The author figures causally in the text he created, to be sure, but his intended meaning is not the sole or final arbiter of the text's meaning. The author may have been hewing to his Muse, which may have planted meanings in the text that he did not even notice, let alone intend, yet there they are.

A good deal of even our most prosaic discourse has this somnambulistic character if you look at it sufficiently closely, and the greater the creative leap it takes, the less aware we are of where it came from (Hadamard 1949). Even the banal act of pushing or not pushing a button at one's whim comes under a fog of indeterminacy when scrutinised microscopically (Harnad 1982b; Dennett & Kinsbourne 1995).

This paper is accordingly not about what Turing (1950) may or may not have actually thought or intended. It is about the implications of his paper for empirical research on minds and machines. So if there is textual or other evidence that these implications are at odds with what Turing actually had in mind, I can only echo the last line of Max Black's (1952) doubly apposite essay on the "Identity of Indiscernibles," in which the two opposing metaphysical viewpoints (about whether or not two things that there is no way to tell apart are in reality one and the same thing) are presented in the form of a dialogue between two interlocutors. Black effects to be even-handed, but it is obvious that he favours one of the two. The penultimate line, from the unfavoured one, is something like "Well, I am still not convinced," the last line, from the favoured one, "Well, you ought to be."

2. The Standard Interpretation of Turing (1950)

Turing is usually taken to be making a point about "thinking" (Dennett 1985; Davidson 1990). Using the example of a party game in which one must guess which of two out-of-sight candidates is male and which is female on the basis of written messages alone (today we would say on the basis of email-interaction alone), Turing asks what would happen if the two candidates were instead a man and a machine (let us say a digital computer, for the time being) and we could not tell them apart. For then it would follow that we could have no grounds for holding that one of them was thinking and the other was not. Hence either the question "What is thinking?" is meaningless, or it is provisionally answered by whatever proves successful in passing such a test.

This is the gist of the "thought experiment" Turing invented in that paper, but it clearly needs some more elaboration and analysis, for otherwise it is open to misconstrual.

3. Misinterpretations of Turing (1950)

One misconstrual is that the outcome of such a "Turing Test" would accordingly just be a trick: The successful machine candidate would not really be thinking at all, and Turing has simply shown that it might be possible to fool us in this way with a computer.

I don't think this is a correct interpretation of Turing's thought experiment. Turing's point is much more substantive and positive than this. If there is any implication about trickery, it is that we are fooling ourselves if we think we have a more sensitive basis than the Turing Test (TT) for determining whether or not a candidate is thinking: All functional tests of thinking or intelligence are really just TTs (Harnad 1992b).

I will return to this point when I introduce the TT Hierarchy. For now, note simply that the notion that "I have a better way" of discerning whether a candidate is really thinking calls for a revelation of what that better way is, relative to which the TT is merely a trick. It turns out that all the "better ways" are either arbitrary or subsumed by a sufficiently general construal of what the TT itself is. (Again, it matters little whether or not Turing explicitly contemplated or intended this generalisation; it is implicit in his argument.)

A second misconstrual of Turing's argument is that he somehow proved that any candidate passing the TT must be thinking/intelligent: He certainly did not prove that (nor is it amenable to proof, nor is it true of necessity), and indeed there now exists a prominent counterexample -- not a proof either, but a counterargument as strong as Turing's original argument -- for one specific kind of candidate passing one specific version of the TT, namely, a pure symbol-system passing the purely symbolic version of the TT. The counterargument in question is Searle's (1980) notorious "Chinese Room Argument," and I will return to it shortly (Harnad 1989). Suffice it to say that Turing's is an epistemic point, not an ontic one, and heuristic rather than demonstrative.

4. Equivocation About Thinking and Machines

So if passing the TT is neither evidence of trickery nor a guarantee of intelligence, what is it?

To answer this, one must say much more explicitly what Turing's Argument shows, and how. To do this, we have to get rid of the vexed word "thinking," which Turing in any case wanted to conclude was equivocal, if not incoherent. It is equivocal in that the very distinction between "real" and "artificial" here is equivocal. In a trivial sense, any natural thing that we succeed in synthesising is not "real," precisely in that it is not natural but synthetic, i.e., man-made (Harnad 1993b). So the question of what thinking is, and whether a man-made machine can have "real" thoughts would be trivially answered if "real" simply means "natural" (i.e., not man-made, not "synthetic"), for then nothing synthetic is real.

So if the real/artificial distinction is not just the natural/synthetic distinction, what is it then? Even the word "machine" is equivocal in exactly the same sense, for what is a machine? Is it a synthetic, man-made mechanism? Or is it any physical/causal mechanism at all? If cars or planes or robots grew on trees, would that alone make them no longer machines? And if we bioengineered an organic organism from scratch, molecule by molecule, would that make it more a machine than its natural counterpart already is?

No. These distinctions are clearly too arbitrary. The deep issues here are not about the natural/man-made distinction but about something else, something going much deeper into the nature of mechanism itself, irrespective of its provenance.

5. Functional Capacity: Mindful vs. Mindless

It is easy to give a performance-based or functional criterion for "intelligence" (precisely the same point can be made about "thinking"): Intelligence is as intelligence does. It requires intelligence to factor a quadratic equation. Hence, any system that can do that, has, eo ipso, intelligence. But this is clearly not what we mean by intelligence here, and defining it thus does not answer any deep questions about what might be present or lacking in a system along these lines.

What are these lines, then? I think it has become abundantly clear that what we are really wondering about here is whether or not a system has a mind: If it does intelligent, thoughtful, mind-like things, but does so mindlessly, because there is nobody home in there, no one experiencing experiences, feeling feelings, thinking thoughts, then it falls on one side of the dividing line; if it does have a mind, be that mind ever so limited, "thoughtless" and "unintelligent," it is on the other side. And the natural fault marked by this dividing line, coincides, for better or worse, with the mind/body problem.

6. Turing Indistinguishability and the Other-Minds Problem

So Turing is right that we should set aside the equivocal notion of "thinking/intelligence," because nothing nonarbitrary hinges on it. What we really want to know about the candidate is whether or not it has a mind -- and this clearly cuts across the natural/synthetic distinction, for if there is no way to determine whether or not a system has a mind on the basis of its provenance (natural or man-made), then in learning that a TT-passing candidate is a "machine," we have not really learned anything at all (insofar as the question of whether it has a mind is concerned).

This is merely a special case of an even more general indeterminacy, namely, the fact that no empirical test, TT-passing or otherwise, can tell us whether or not a candidate has a mind! This is one of the lemmas we inherit with the mind/body problem, and it is called the "other-minds problem" (Harnad 1991). (It would have cut short a lot of potential misconstruals of Turing's paper if he had simply called a spade a spade here, but there we are; Turing was not a philosopher but a mathematician.)

Turing did allude to the other-minds problem in passing, but only to dismiss it as leading to "solipsism" and hence the end of rational discourse ("It may be the most logical view to hold but it makes communication of ideas difficult"). Yet the other-minds problem is an inescapable methodological constraint on mind-modelling (Harnad 1984, 1985). For the only way to know for sure that a system has a mind is to be that system. All other ways are risky, and risky in a way that is much more radical than the ordinary Humean risk that goes with all empirical inferences and generalisations. I am not referring here to philosophical scepticism about whether someone as like me as my neighbour has a mind; let us take that for granted. But what happens as we move away in likeness, toward other mammalian species, invertebrates, micro-organisms, plants, man-made artifacts? Where does mere philosophical scepticism leave off and real empirical uncertainty kick in?

7. Out of Sight, Out of Mind: Dissociating Structure and Function

Turing implicitly partitioned this question. "Likeness" can take two forms: likeness in structure and likeness in function. In sending the candidates out of the room in the party game, Turing separated what he took to be irrelevant and potentially biasing aspects of the superficial appearance of intelligence from pertinent manifestations of the functional capacities that constitute it. But banishing physical structure from sight, Turing also inadvertently excluded some potentially relevant aspects of function, both internal and external to the candidate.

For human minds, at least, can do a lot more than just exchange words. And as the TT is clearly predicated on functional likeness (indeed, functional indistinguishability), to banish all our nonverbal functions from it would be to limit the TT's domain of likeness in an almost arbitrary way, and certainly in a way that goes against the spirit of a test that is predicated on functional indistinguishability (Turing-indistinguishability).

8. The Expressive Power of Natural Language and Formal Computation

The special case of only words-in and words-out is not entirely arbitrary, however, because of the remarkable expressive power of language, which some have even suggested might be limitless and universal, including, as it does, not only the capacity of words to express any possible proposition [Steklis & Harnad 1976; Harnad 1996], but also the full computational power of the Turing Machine [Harnad 1982a, 1994b]. Nevertheless, there are things that human beings can do that go beyond mere verbalising; and perhaps more important, these nonvocal capacities may functionally precede and underlie our capacity to have and exchange words: We are able, after all, to interact nonverbally with that world of objects, events, and states that our words are about (e.g., we can detect, discriminate, identify, and manipulate them). Indeed, it is hard to imagine how our words could have the meanings they have if they were not first grounded in these nonverbal interactions with the world (Harnad 1990a, 1992a, 1993a; Crockett 1994).

9. The Turing Hierarchy: Level t1 (Toy Models)

It is accordingly time to introduce the TT hierarchy (Harnad 1994a). The lowest level would properly be T1, but there is no T1, only t1, for here "t" stands not for "Turing" (or "Total"), but for "toy," and t1 launches the TT hierarchy without yet being a proper part of it: The t1 level of Turing modelling and Turing Testing is the level of "toy models." These are models for subtotal fragments of our functional capacity. At worst, these fragments are just arbitrary ones; at best, some might be self-sufficient modules, eventual components of our full-scale capacity, but not stand-alone candidates in the Turing sense. For the TT is predicated on total functional indistinguishability, and toys are most decidedly distinguishable from the real thing. (Toys aren't Us.)

Before moving up to T2 we will make it explicit exactly why toys are ultimately inadequate for the goals of Turing Testing. But note that all of the actual mind-modelling research efforts to date are still only at the t1 level, and will continue to be so for the foreseeable future: Cognitive Science has not even entered the TT hierarchy yet.

10. Level T5 (Grand Unified Theories of Everything [GUTEs])

Toy models, that is, subtotal models, of anything at all, have the liability that they have more degrees of freedom than the real, total thing that they are modelling. There are always more different ways to generate a fragment of a system's function (say, chess-playing -- or simple harmonic motion) than there are to generate the system's total function (say, everything else a human mind can do -- or all of Newtonian or Quantum Mechanics), if for no other reason than that the Total model must subsume all the functions of the subtotal toy models too (Schweizer 1998; cf. Lucas 1961).

We spoke of Humean indeterminacy earlier; philosophers are not yet of one mind on the number of candidate Total Theories there are likely to be, at the end of the empirical/theoretical road for, say, Physics. At that Utopian endpoint, when all the data -- past, present and future -- are Totally accounted for, will there be one and only one Grand Unified Theory of Everything (GUTE)? Or will there be several rival GUTEs, each able to account for everything? If the latter, will the rivals posit an equal number of parameters? and can we assume that the rival GUTEs will just be notational variants of one another? If not, will we be right to prefer the miniparametric GUTE, using Occam's Razor?

These are metaphysical questions, and physicists would be more than happy to have reached that Utopian stage where these were the only questions left to worry about. It is clear, however, that current physics is still at the stage of subtotal theories (it would be churlish for any other science to venture to call them "toys," as physics's are surely the most advanced such toys any science has yet produced). Subtotal models always run the risk of being very wrong, although that risk diminishes as they scale up toward a closer and closer approximation to totality (GUTE); the degrees of freedom shrink, but they may not vanish altogether, even after we have reached Utopia.

Utopia is also the top Level T5 of the Turing Hierarchy. We will return to this after describing Levels T2-T4.

11. Utopian Differences That Make No Difference

The Humean risk that our theory is wrong, instead of shrinking to zero even at totality, may remain in that nonzero state which philosophers call the "underdetermination" of theory by data. The ever-present possibility that a theory could be wrong, even after it accounts for all the data, and that another theory might instead be right, is a liability that science has learned to live with. By way of consolation, one can note that, at the point of miniparametric totality -- when all the data have been accurately predicted and causally explained by the GUTE(s) with the fewest parameters, the difference between being right and wrong would no longer be one that could make any palpable difference to any of us.

Consider the status of "quarks," for example. (I am not speaking now of actual quarks, which have lately become experimentally observable, but the original notion of a quark, which was a largish entity that could not occur in "free state," but only tightly "bound" up together with other quarks into a much tinier observable entity such as an electron.) The existence of quarks was posited not because they were experimentally observed, but because hypothesising that they existed, even though they could never be observed empirically, made it possible to predict and explain otherwise unpredictable and inexplicable empirical data. But quarks have since turned out to be observable after all, so I am now speaking about what GUTEs would have been like if they had contained unobservable-in-principle quarks. [A more current example might be superstrings.]

At the end of the day, once all data were in and fully accounted for by GUTEs that posited (unobservable-in-principle) quarks, one might still wonder about whether quarks really exist. Maybe they do, maybe they don't. There is no empirical way left to find out. Maybe the GUTEs that posit quarks are wrong; but then other, quarkless, GUTEs must have components that do for them the predictive/explanatory work that the quarks do for the quarkful GUTEs. In any case, it is a fact about the quarkful GUTEs that without positing quarks they cannot successfully predict/explain Everything.

12. Empirical Underdetermination and Turing Testing

I rehearse this indeterminacy explicitly in order to make it clear how little is at stake: All data are successfully predicted and explained, and we are just worrying about which (if any) of the successful Total, Turing-Indistinguishable explanations is the real one; and no one can ever be the wiser except an omniscient deity who simply knows the truth (e.g., whether or not there really are quarks).

This is normal scientific underdetermination. Notice that the name of Turing has been introduced into it, because with GUTEs we are actually at the top (T5) level of the Turing hierarchy, which turns out to be a hierarchy of (decreasing) degrees of empirical underdetermination. The t1 (sub-Turing) level, for both physical science (world-modelling) and cognitive science (mind-modelling), is that of subtotal, toy models, accounting for some but not all of the data; hence underdetermination is maximal in toyland, but shrinks as the toys grow.

13. Basic Science Vs. Engineering: Forward and Reverse

Physical science and cognitive science are not really on a par, because physical systems with minds are of course a part of the world too. Cognitive science is probably better thought of as a branch of reverse bioengineering (Dennett 1994): In normal engineering, the principles of physics (and engineering) are applied to the design of physical systems with certain functions that are useful to us. In reverse bioengineering, those systems have already been designed by Nature (by the Blind Watchmaker, Darwinian Evolution), and our task is to figure out how they work. The actual methodology turns out to resemble forward engineering, in that we still have to design systems that are more and more like the natural ones, starting with toy models, and converging on Total Turing-Indistinguishability by a series of tighter and tighter approximations.

Which brings us back to the Turing Hierarchy. Before we leave the t1 level, let it be stressed that, in principle, there can be a distinct Turing Hierarchy for each species' functional capacities; hence there can be toy models of mollusk, mouse and monkey, as well as man. Other species no doubt have minds (I personally think all animals with nervous tissue, vertebrate and invertebrate, do, and I am accordingly a vegetarian), but, as mentioned earlier, our confidence in this can only diminish as we move further and further from our own species. So it is not arbitrary that Turing Testing should end at home (in the total performance capacities of our own species), even if the toy modelling begins with the reverse engineering of "lesser" species (Harnad 1994a).

So at the bottom level of the hierarchy are subtotal toy models of subhuman and human functional capacities. Such toy models are also sub-Turing (which is why this level is called t1 instead of T1), because the essential criterion for Turing Testing is Total Indistinguishability from the real thing, and toys are by definition subtotal, hence distinguishable. Nevertheless, the idea of converging on Totality by closer and closer approximations is implicit in Turing's paper, and it is the basis for his sanguine predictions about what computers would successfully scale up to doing by our day. (Unfortunately, his chronology was too optimistic, and this might have been because he was betting on a particular kind of candidate, a purely computational one; but if so, then he shouldn't have been, for the Turing spectrum admits many other kinds of candidate mechanisms too.)

14. Level T2: The Pen-Pal Turing Test (Symbols-in/Symbols-out)

T2 is the conventional TT, the one I've taken to calling the "pen-pal" version: words in and words out (Harnad 1989). It is the one that most people have in mind when they think of Turing Testing. But because of the unfortunate "imitation game" example with which it was introduced, T2 has been subject to a number of misunderstandings, which I will quickly try to correct here:

14.1. Life-Long: T2 is not a game (or if it is, it's no less than the game of life!). It is a mistake to think of T2 as something that can be "passed" in a single evening or even during an annual Loebner Prize Competition (Loebner 1994; Shieber 1994). Although it was introduced in the form of a party game in order to engage our intuitions, it is obvious that Turing intends T2 as a serious, long-term operational criterion for what would now be called "cognitive science. The successful candidate is not one that has fooled the judges in one session into thinking it could perform indistinguishably from a real pen-pal with a mind. (Fooling 70% of the people one time is as meaningless, scientifically, as fooling 100% of the people 70 times.) The winning candidate will really have the capacity to perform indistinguishably from a real pen-pal with a mind -- for a lifetime, if need be, just as unhaltingly as any of the rest of us can. (This is a variant of the misconstrual I mentioned earlier: T2 is not a trick; Harnad 1992b.)

14.2. Symbolic/Nonsymbolic: T2 is restricted to verbal (i.e., symbolic) interactions: symbols in, symbols out. But it is not restricted in principle to a computer doing only computation (symbol manipulation) in between (although Turing so restricts it in practise). The candidate "machine" can in principle be anything man-made, including machines other than digital computers (e.g., sensorimotor transducers, analog computers, parallel distributed networks, indeed any dynamical system at all; Harnad 1993a).

14.3. Man-Made: The reason the candidate needs to be man-made is that all the tests in the TT Hierarchy are tests of mind-models. Interposing a natural candidate that we did not build or understand would not be a test of anything at all. Turing was indifferent about whether the thinking thereby generated by the system was real or not -- he thought that that was an uninteresting question -- but he did want to successfully synthesize the capacity, in toto. So we could say that he did not care whether the result would be the reverse bioengineering of the real mind, or the forward engineering of something mind-like but mindless, but the engineering was certainly intended to be real, and to really deliver the behavioral goods!

14.4. Mindful: Turing was not interested in the question of whether the candidate would really have a mind, but that does not mean that the question is of no interest. More important, he ought to have been interested, because, logically speaking, without the mental criterion, T2 is not a test of anything: it is merely an open-ended catalogue of what a particular candidate can do. The indistinguishability is between the candidate and a real person with a mind. So, like it or not, if T2 is not merely to be testing the difference between two arbitrary forms of cleverness, it must be testing mindful cleverness against mindless. And once they are indistinguishable, then we no longer have a basis for affirming of the one a predicate we wish to deny of the other (other than trivial predicates such as that one is man-made and the other is not). And "mindful/mindless" is just such a (nontrivial) predicate.

14.5. Conscious: Note that T2 is rightly indifferent to whether the candidates are actually doing what they do consciously (Harnad 1982b; Searle 1990). They might be doing it somnambulistically, as discussed earlier in connection with the intentional fallacy. This does not matter: the real person has a mind, even if it's only idling in the background, and that would be a world of difference from a candidate that had no mind at all, foregrounded or backgrounded, at that moment or any other. And if they're indistinguishable (for a lifetime), then they're indistinguishable; so if you are prepared to impute a mind to the one, you have no nonarbitrary basis for denying it to the other.

14.6. Empirical/Intuitive: Turing Indistinguishability is actually two criteria, one empirical and one intuitive: The empirical criterion is that the candidate really has to have total T2 capacity, for a lifetime. The intuitive criterion is that that capacity must be totally indistinguishable from that of real people with minds to real people with minds. (Note the plural: not only is T2 not a one-night affair; it is not merely a dyadic one either. Real pen-pals have the capacity to correspond with many people, not just one.) The intuitive criterion may be almost as important as the empirical one, for although (because of the other-minds problem) we are not really mind-readers, our intuitive "theory-of-mind" capacities may be the next best thing (Heyes 1998, Mitchell & Anderson 1998), picking up subtleties that are dead give-aways to our fine-grained intuitions while still out of the reach of the coarse-grained reverse-engineering of our overall cognitive capacities.

15. Searle's Chinese Room Argument

Only corrective (14.2) above is unique to T2. The rest are applicable to all the levels of the TT Hierarchy. T2 is a special level not just because it is the one at which Turing originally formulated the TT, and not just because of the special powers of both natural language and formal computation on which it draws, but because it is the only level that is vulnerable to a refutation, in the form of a counterexample.

So let us now briefly consider this celebrated counterexample to T2, for it will serve to motivate T3 and T4: But first, note that the intentional fallacy applies to Searle's (1980) paper as surely as it applies to Turing's (1950) paper: Searle thought he was refuting not only cognitive "computationalism" (which he called "Strong AI" and will be defined in a moment), but also the Turing Test. In fact, he was refuting only T2, and only in the special case of one specific kind of candidate, a purely computational one: an implementation-independent implementation of a formal symbol system. T2, however, is not the only possible form a TT can take, and a pure symbol system is certainly not the only possible candidate.

Searle (1980, 1984, 1993) defined his target, "Strong AI," as the hypothesis expressed by the following three propositions:

(1) The mind is a computer program.

(2) The brain is irrelevant.

(3) The Turing Test is decisive.

Similarly, what Searle should have said instead of (2) was that the physical details of the implementation are irrelevant -- but without the implication that implementing the code at all was irrelevant. Hence the rest of the details about the brain are irrelevant; all that matters is that it is implementing the right code.

To hold that the capacity to pass the TT (3) is what demonstrates that the "right" code is being implemented, the one that implements a mind, is unobjectionable, for that is indeed Turing's argument -- except of course that Turing was not necessarily a cognitive computationalist! If just a computer, running the right code, could successfully pass the pen-pal version of the TT (T2), Turing would indeed have held that we had no better or worse grounds for concluding that it had a mind than we would have in the case of a real pen-pal, who really had a mind, for the simple reason that we had no way to tell them apart -- their performance would be Turing indistinguishable.

However, Turing would have had to concede that real pen-pals can do many other things besides exchanging letters; and it is hard to see how a computer alone could do all those other things. So Turing would have had no problem with an augmented candidate that consisted of more than just a computer implementing the right code. Turing would have been quite content, for example, with a robot, interacting with the world Turing indistinguishably from the rest of us. But that version of the TT would no longer be T2, the pen-pal version, but T3, the robotic version.

We will return to T3 shortly. It must also be pointed out that Turing is not irrevocably committed to the candidate's having to be merely a computer even in the case of T2. The T2-passing system could be something other than just a computer running the right code; perhaps no computer/code combination alone could successfully generate our lifelong pen-pal capacity, drawing as the latter does on so much else that we can all do. We could, for example, include a photo with one of our letters to a pen-pal, and ask the pen-pal to comment on the quality of the blue object in the photo (or we could simply enclose a blue object). A mere computer, without sensorimotor transducers, would be completely nonplussed by that sort of input. (And the dynamical system consisting of a computer plus sensorimotor transducers would no longer be just a computer.)

Never mind. A (hypothetical) T2-passing computer is nevertheless fair game, in this thought-experimental arena that Turing himself created. So Searle takes Turing at his word, and supposes that there is a computer program that is indeed capable of passing T2 (for a lifetime); he merely adds the stipulation that it does so in Chinese only. In other words, the candidate could correspond for a lifetime with Chinese pen-pals without any of them having any reason to suspect that they were not communicating with a real (Chinese-speaking) person. Turing's argument is that at the end of the day, when it is revealed that the candidate is a computer, one has learned nothing at all to over-rule the quite natural inference one had been making all along -- that one was corresponding with a real pen-pal, who really understood what all those letters (and his own replies) had meant.

It is here that Searle makes a very shrewd use of the other two premises -- that what is at issue is the code alone (1), and that although that code must be implemented, the physical details of the implementation are irrelevant (2): each and every implementation of the right code must have a mind, if any of them does, and if they indeed have a mind only because they are implementations of the right (i.e., T2-passing) code.

So Searle too makes a simple appeal to our intuitions, just as Turing did in appealing to the power of indistinguishability (i.e., in noting that if you can't tell them apart, you can't affirm of one what you deny of the other): Searle notes that, as the implementation is irrelevant, there is no reason why he himself should not implement the code. So we are to suppose that Searle himself is executing all the computations on the input symbols in the Chinese messages -- all the algorithms and symbol manipulations -- that successfully generate the output symbols of the pen-pal's replies (the ones that are indistinguishable from those of real life-long pen-pals to real life-long pen-pals).

Searle then simply points out to us that he would not be understanding the Chinese messages he was receiving and generating under these conditions (any more than would a student mindlessly but successfully going through the motions of executing some formal algorithm whose meaning he did not understand). Hence there is no reason to believe that the computer, implementing exactly the same code -- or indeed that any implementation of that code, merely in virtue of implementing that code -- would be understanding either.

What has happened here? Turing's "no way to be any the wiser about the mind" argument rested on indistinguishability, which in turn rested on the "other-minds" barrier: The only way to know whether any candidate other than myself has or hasn't a mind is to be that candidate; but the only one I can be is me, so Turing indistinguishability among candidates with minds is all I have left.

16. Penetrating the Other-Minds Barrier Through Implementation-Independence

What Searle has done is to exploit one infinitesimally narrow window of vulnerability in cognitive computationalism and T2: No one can show that a computer (or a rock) has or hasn't a mind, because no one can be that computer or that rock. But the computationalism thesis was not about the computer, it was about the running code, which was stipulated to be implementation-independent. Hence the details of how the code is executed are irrelevant; all that is needed is that the code be executed. And it is indeed being executed by Searle, who is thereby being the running code, being the candidate (there is no one else in sight!). And he (truthfully) informs us that he does not understand a word: he is merely mindlessly manipulating code.

This counterargument is not a proof (any more than Turing's own indistinguishability argument is a proof). It is logically possible that simply going through the motions of manipulating the right symbols does give birth to another understanding mind -- not Searle's, for he does not understand, but whoever in his head is answering the Chinese messages in Chinese. (It is, by the same token, logically possible that if we taught someone illiterate and enumerate how to apply the symbol-manipulation algorithm for solving quadratic equations, that he, or someone else inside his head, would thereby come to understand quadratic equations and their solution.)

17. Multiple Minds? The Searlean Test [ST]

This is spooky stuff. We know about multiple personality disorder -- extra minds cohabiting the same head, not aware of one another -- but that is normally induced by early sexual abuse rather than by learning to manipulate mindlessly a bunch of meaningless symbols.

So I think it makes more sense to give up on cognitive computationalism and concede that either no implemented code alone could ever successfully pass the lifetime T2, or that any code that did so would nevertheless fail to have a mind; that it would indeed be just a trick. And the way to test this would be via the "Searlean Test" (ST), by being the candidate, executing the code oneself, and failing to experience the mental state (understanding) that is being attributed to the pen-pal. As one has no intuitive inclination whatsoever to impute extra mental states to oneself in the face of clear 1st-person evidence that one lacks them, the ST should, by the transitivity of implementation-independence, rule out all purely computational T2-passers.

A purely computational T2-passer, in other words, is Turing-distinguishable from a candidate with a mind, and the way to make the distinction is to be the candidate, via the ST [Footnote 1].

18. Grounding Symbols in Their Objects Rather than in the Mind of an External Interpreter

Has Turing lost a lot of territory with this? I don't think so. The other-minds barrier is still there to ensure that no other candidate can be penetrated and hence unmasked in this way. And besides, there are stronger and more direct reasons than the ST for having doubts about cognitive computationalism (Harnad 1990a): The symbols in a symbol system need to be grounded in something other than simply more symbols and the interpretations that our own minds project onto them (as was the case with the symbolic pen-pal), no matter how coherent, self-confirming, and intuitively persuasive those external interpretations might be (Harnad 1990b, 1990c). The symbols must be grounded directly and autonomously in causal interactions with the objects, events and states that they are about, and a pure symbol-cruncher does not have the wherewithal for that (as the photo/object accompanying the letter to the pen-pal demonstrates).

What is needed to ground symbols directly and autonomously in the world of objects? At the very least, sensorimotor transducers and effectors (and probably many other dynamical internal mechanisms); and none of these are implementation-independent; rather, they are, like most things, other-minds-impenetrable, and hence immune to the ST and the Chinese Room Argument. The TT candidate, in other words, cannot be just a disembodied symbol-cruncher; it must be an embodied robot, capable of nonsymbolic, sensorimotor interactions with the world in addition to symbolic ones, all Turing-indistinguishable from our own. This is T3, and it subsumes T2, for pen-pal capacities are merely subsets of robotic capacities (Harnad 1995a).

19. Level T3: The Robotic Turing Test

Note that just as scaling up from toys to totality is dictated by Turing's functional equivalence/indistinguishability constraint, so is scaling up from T2 to T3. For pen-pal capacities alone are clearly subtotal ones, hence distinguishable as such. More important, pen-pal capacities themselves surely draw implicitly on robotic capacities, even in the T2, as demonstrated not only by enclosing any nonsymbolic object with a letter for comment, but surely (upon a little reflection) also by the very demands of coherent, continuous discourse about real objects, events, and states via symbols alone (French 1990; Harnad 1992a; Crockett 1994, Watt 1996).

Why did Turing not consider T3 more explicitly? I think it was because he was (rightly) concerned about factoring out irrelevant matters of appearance. Differences in appearance certainly do not in and of themselves attest to absence of mind: Other terrestrial species surely have minds; extra-terrestrial creatures, if they exist, might have minds too; and so might robots. But it's a safe bet that no current-generation robot has a mind, because they are all still just at the most rudimentary toy level, t1, in their robotic capacities (Pylyshyn 1987). Perhaps Turing simply thought our other-minds intuitions would be hopelessly biased by robotic appearance in T3, but not in T2. In any case, the recent progress in making robots look lifelike, at least in movies, may outstrip our prejudices. (If anything, the problem is that today's cinematic robots are getting ahead of the game, being endowed with fictional T3 capacity while their real-world counterparts are still functionally closer to mud than even to mollusks.)

20. Level T4: Internal Microfunctional Indistinguishability

Do T2-T3 exhaust the options for mind-modelling? No, for if the operational constraint is functional equivalence/indistinguishability, there is still a level of function at which robots are readily distinguishable from us: the internal, microfunctional level -- and here the structure/function distinction collapses too. For just as we can ask whether, say, facial expression (as opposed to facial shape) is not in fact a legitimate component of our robotic capacity (it is), so we can ask about pupillary dilation (which allegedly occurs when we find someone attractive, and is Turing-detectable), blushing, heart-rate, etc. These are all still arguably components of T3. But, by the same token, we become speechless if kicked on the left side of our heads, but not the right (Harnad et al. 1977), we exude EEG, we bleed when we are cut, if we are opened up on an operating table we are found to be made of flesh, and in our heads one finds gray/white matter and neuropeptides rather than silicon chips.

These are all Turing-detectable functional differences between ourselves and T3 robots: Are they relevant? We have already said that the fact that the candidate is man-made is not relevant. Not relevant to what? To determining whether or not the candidate has a mind. Do the neurobiological details matter? Perhaps. We are not, after all, implementation-independent symbol systems. Perhaps there are implementational details that are essential to having a mind.

Surely there are. But just as surely, there must be implementational details that are not essential to having a mind. And the problem is that there is no way of knowing which are which. Hence brain-scanning certainly plays no role in our perpetual Turing-testing of one another (in our age-old, day-to-day coping with the Other-Minds Problem), any more than mind-reading does (Harnad 1991).

We are now at the level of T4: Here the candidates are not only indistinguishable from us in their pen-pal and robotic functions, but also in their internal microfunctions. This is still the Turing hierarchy, for we are still speaking about functional indistinguishability and about totality. But do we really have to go this high in the hierarchy in practise? Indeed, is it even consistent with the spirit and methodology of Turing Testing to insist that the candidate be T4-indistinguishable from us? I think not, and I will now introduce my own thought experiment to test your intuitions on this. My example will be recognisably similar to the earlier discussion of underdetermination and the Utopian GUTEs in physics.

21. Grand Tournament of T-Levels: T3 vs. T4 vs. T5

Suppose that at the end of the day, when mind-modelling attains its Utopia, there are nine successful candidates. All nine are T3-indistinguishable from us, but three of them fail T4 (cut them open and they are clearly distinguishable in their internal microfunctions from us and our own real neural functions). Of the remaining six, three pass T4, but fail T5: their internal neural activities are functionally indistinguishable -- transplants can be swapped between us and them, at any biocomponential scale -- but theirs are still crafted out of synthetic materials; they are T4-indistinguishable, but T5-distinguishable; only the last three are engineered out of real biological molecules, physically identical to our own, hence T5-indistinguishable from ourselves in every physical respect. The only remaining difference among these last three is that they were designed in collaboration with three different physicists, and each of the physicists happens to subscribe to a different GUTE -- of which there also happen to be three left at the end of the day.

Now note that the three T5 candidates are empirically identical in kind, right down to the last electron. The only difference between them is their respective designers' GUTES (and recall that each GUTE accounts successfully for all physical data). Now, according to one of the GUTEs, unobservable-in-principle quarks exist; according to another, they do not, but their predictive/explanatory work is done instead by some other unobservable entity. Both these GUTEs have the same number of parameters. The third GUTE has fewer parameters, and needs neither quarks nor their counterparts in the second GUTE.

So there we have our nine candidates: three T3's, three T4's and three T5's. All nine can correspond with you as a pen-pal for a lifetime; all nine can interact robotically with the people, objects, events and states in the world indistinguishably from you and me for a lifetime. All nine cry when you hurt them (whether physically or verbally), although three of them fail to bleed, and three of them bleed the wrong kind of blood. For the sake of argument, you have known all nine all your life (and consider them all close friends), and this state of affairs (that all nine are man-made) has only just been revealed to you today. You had never suspected it (across -- let us say, to calibrate our intuitions -- 40 years of friendship and shared life experiences!).

You are now being consulted as to which of the nine you feel can safely be deprived of their civil rights as a result of this new revelation. Which ones should it be, and why?

I think we are still squarely facing the Turing intuition here, and the overwhelming answer is that none should be deprived of their civil rights; none can be safely assumed to be "Zombies" (Harnad 1995b). By all we can know that matters for such moral decisions about people with minds, both empirically and intuitively, they all have minds, as surely as any of the rest of us do [Footnote 2].

22. Level T3 Is Decisive

So does that mean that all differences above the level of T3 matter as little (or as much) to having a mind as the differences between the three Utopian GUTEs? Intuitively (and morally), I think the answer is undeniably: Yes. Empirically, it may prove to be a different story, because there may turn out to be some causal dependencies: Certain microfunctional (T4) or even microphysical (T5) properties may not prove to be viable in passing T3 (e.g., it might not be possible to pass T3 with serial processors, or with components made out of heavy water); other T4 or T5 properties might give us positive clues about how to pass T3 (excitation/inhibition or 5-hydroxytryptamine may be essential) -- and any success-enhancing clue is always welcome in reverse-engineering, as anywhere else! In this way neuroscience and even fundamental physics could conceivably help us pass T3; but T3 would still be the final arbiter, not T4 or T5.

The reason is clear, and again it is already implicit in Turing's functional-indistinguishability criterion: Not only are ordinary people not mind-readers; reverse-engineers aren't mind-readers either. All are constrained by the other-minds barrier. Function is the only empirical criterion.

(And structural correlates are only reliable if they are first cross-validated by function: the reason I trust that one flashing red nucleus in my otherwise comatose and ostensibly brain-dead patient is evidence that he is still alive, feeling, and conscious, and will shortly make a come-back, is that prior patients have indeed made come-backs under those conditions, and subsequently reported the words they had heard me speak as I contemplated pulling the plug on their life-support systems at the very moment that only their red nucleus was still flashing.)

23. 1st-Person and 3rd-Person Facts

Now the forward-engineering of the mind would clearly be only a functional matter, just as the forward engineering of bridges, cars and space stations is. Reverse-engineering the mind is a bit more complicated, because we know one extra fact about the mind, over and above the 3rd-person functional facts, and that is the 1st-person fact that it feels like something to have/be a mind.

But this 1st-person fact is precisely the one to which the 3rd-person task of reverse engineering can never have any direct empirical access; only its functional correlates are available (Harnad 2000c). Yet it is indeed a fact, as is clearly illustrated if we contrast the case of quarks with the case of qualia ("qualia" is the name philosophers have given to the qualitative contents of our mental experiences: they are what it feels like to have a mind, experience experiences, think thoughts, mean meanings; Nagel 1974; Harnad 1993c):

24. Quarks, Qualia, and Causality

What we are actually asking about each of our nine candidates is whether or not it really has qualia. Is this like asking, about the Utopian world-models, whether there really are quarks in the world (i.e., whether the quarkful GUTE is the right GUTE)? Yes, but it's a bit worse in the case of mind-models, because whereas the reality or nonreality of (unobservable-in-principle) quarks, apart from the formal predictive/explanatory role they play (or don't play) in their respective GUTEs, makes and can make no palpable difference to anyone (I chose the word "palpable" advisedly), this is decidedly untrue of qualia. For surely it would make a world of palpable difference to one of the nine candidates, if we got our inferences wrong, and assumed it did not have qualia, when it in reality did! And each of us knows, palpably, precisely what that difference is -- unlike the nebulous difference that the existence or non-existence of quarks would make, given the total predictive/explanatory power of its GUTE.)

So, metaphysically (as opposed to empirically or intuitively), more is at stake in principle with qualia than with quarks -- which is why mind-modelling has a Zombie Problem whereas world-modelling does not (Harnad 1995b).

25. Mind-Blindness, the Blind Watchmaker, and Mind over Matter

And there is a further difference between the case of quarks and the case of qualia: quarks, whether or not they exist, do play both a formal and a causal role in their GUTE. Qualia can play neither a formal nor a causal role in any mind-model, be it T3, T4 or T5. They are undeniably there in the world, but they can play no explicit role in the reverse-engineering, on pain of telekinetic dualism or worse (Harnad 1982b; Alcock 1987)! This too is implicit in Turing's insistence on functional criteria alone.

To put in context the mind-blind functionalism that Turing indistinguishability imposes on us, it is instructive to note that the Blind Watchmaker who forward-engineered us all (Darwinian Evolution) was a strict functionalist too, and likewise no mind-reader. Natural Selection is as incapable of distinguishing the Turing-indistinguishable as any of the rest of us (Harnad 2000a).

26. Turing's Text and Turing's Mind

This completes our exegesis of Turing (1950). Was this really the author's intended meaning? Perhaps not, but if not, then perhaps it ought to have been! His text is certainly systematically interpretable as meaning and implying all of this. But if Turing did not actually have it all in mind, then we are making a wrong pen-pal inference in imputing it to him -- and it should instead be attributed only to the inert code he generated in the form of his celebrated paper.

FOOTNOTES

Footnote 1. One of the referees for this article suggested that I respond to the following three published criticisms by Hauser (1993), so I hereby do so:

So far we have mentioned only sufficiency to pass T2. Sufficiency "to evidence intelligence" is the further question of whether the T2-passing candidate would indeed have a mind. Let us call this Turing's conjecture (b).

I subscribe to Turing's conjecture (b) too, with one special exception, namely, should my grounding hypothesis (a) be incorrect, and should it turn out that an ungrounded, implementation-independent symbol system alone could pass T2 after all, then it would not have a mind (as shown by Searle's Periscope).

This invalidates T2 as evidence of having a mind in the case of one very special kind of candidate only, a purely computational one. All the possibilities and contingencies, however, are perfectly coherent.

Turing's conjecture (b) is that we do not have a better basis for inferring that a candidate has a mind (consciousness) than T2/T3 capacity. T3 is a "better" basis for reverse-engineering that capacity than T2 alone (for T2 is a subset of T3, and grounded in it); it is also a better basis for testing success in reverse-engineering it (it would be easier to catch out a robot than a pen-pal, both because a robot's capacities subsume a pen-pal's capacities, and because they give us more to go on, in mind-reading for anything that distinguishes the candidate from a real person).

But a positive outcome to T3 still leaves open the distinct possibility that the candidate is mindless (i.e., that Turing's conjecture (b) is false) -- a possibility that is trivial when it comes to Turing-Testing members of our own species (unless they are comatose), but one that becomes more substantial as we move toward species less and less like ourselves; the only way to know whether a polyp, or a plant, feels anything when you pinch it is to be that polyp or plant. The same is true of a toy robot; but as its capacities become more and more like our own, can we help ourselves to the same easy confidence we have with natural members of our own species (Harnad 1985)?

All of this is perfectly coherent. We don't need 1st-hand evidence of consciousness in other members of our species because we are designed with mind-reading capacities that leave no room for doubt. But when we deploy those same mind-reading capacities on species or artifacts that they were not designed for (by the Blind Watchmaker), they could lead us astray. Turing is right to remind us, though, that only a fool argues with the indistinguishable.

Otherwise put: Because of the other-minds problem, we will never have access to positive evidence about consciousness in any case but our own. But, on Turing's conjecture, in other cases, absence of Turing Indistinguishability could be taken as negative evidence.

Footnote 2. Having appealed to our intuitions in this particular hypothetical case, it is unavoidable to go on to raise the question of whether we would feel comfortable (on the strength of Searle's Chinese Room Argument) in pulling the plug on our lifelong T2 pen-pal, if we found out that he was just a computer? Here our evolved "theory-of-mind" intuitions -- our Turing intuitions -- are utterly at odds with both our rational analysis and our real-life experience to date. Some people are even loath to pull the plug on cyberplants.

So of the two aspects of the Turing Test -- the empirical one, of successfully reverse-engineering our full performance capacity, and the intuitive one, of being unable to discern any difference between the candidate and any other person with a mind -- the intuitive, "mind-reading" aspect would be tugging hard in the case of a T2-passing computer, despite Searle's periscope, which unmasks the symbol system, but not the computer. For in the case of the computer pen-pal, we could rationalise the outcome by saying "So it's that computer implementation that has a mind, not just the symbol system it is implementing" (thereby merely jettisoning implementation-independence, and Searle's periscope with it).

So what if the code was being ported to a different computer every day? Our knowledge about transplants and cloning, plus our telefantasies about teleportation would still restrain us from pulling the plug (reinforced by the asymmetry between the relative consequences of false positives vs. false negatives in such capital decisions).

And what if it turned out to have been Searle mindlessly executing the code all those years, and telling us now that he had never understood a word? Well then the question would be moot, because there would be nothing left to pull the plug on anyway (although most of us would feel that we'd still like to send Searle back to the Room so we could write one last letter to our pen-pal under even those conditions, to find out what he had to say about the matter -- or at least to say goodbye!)

We are contemplating hypothetical intuition clashes here; so perhaps we should wait to see what candidates can actually pass the TT before needlessly committing ourselves to any counterfactual contingency plans.

REFERENCES

Alcock, J.E. (1987) Parapsychology: Science of the anomalous or search for the soul? Behavioral and Brain Sciences 10 (4): 553-565.

Black M. (1952) The Identity of Indiscernibles. Mind 61: 152-164.

Crockett, L. (1994). The Turing Test and the Frame Problem: AI's Mistaken Understanding of Intelligence. Norwood, NJ: Ablex.

Davidson, D. (1990) Turing's Test. In: "Modelling the mind." Karim A. Mohyeldin Said, W. H. Newton-Smith, R. Viale, K. V. Wilkes (Eds.) Clarendon Press/Oxford University Press, Oxford, England. (Cognitive Science Seminars, 1985, U Oxford, Oxford, England) 1990. p. 1-11.

Dennett, D. (1985). Can machines think? P. 121. In How We Know. (ed.) M. Shafto. San Francisco, CA: Harper & Row.

Dennett, D.C. (1994) Cognitive Science as Reverse Engineering. Proceedings of the 9th International Congress of Logic, Methodology and Philosophy of Science. Pp. 679-689. Ed. D. Prawitz, B. Skyrms, and D. Westerstahl. North-Holland. http://cogsci.soton.ac.uk/~harnad/Papers/Py104/dennett.eng.html

Dennett, D.C. & Kinsbourne, M. (1995). Time and the observer: The where and when of consciousness in the brain. Behavioral and Brain Sciences 15 (2): 183-247. http://www.cogsci.soton.ac.uk/bbs/Archive/bbs.dennett.html

French, R. (1990). Subcognition and the limits of the Turing Test. Mind, 99(393):53-65.

Hadamard, J. (1949) An essay on the psychology of invention in the mathematical field. Princeton University Press.

Harnad, S. (1982a) Neoconstructivism: A unifying theme for the cognitive sciences. In: Language, mind and brain (T. Simon & R. Scholes, eds., Hillsdale NJ: Erlbaum), 1 - 11. http://www.cogsci.soton.ac.uk/~harnad/Papers/Harnad/harnad82.neoconst.html

Harnad, S. (1982b) Consciousness: An afterthought. Cognition and Brain Theory 5: 29 - 47. http://www.cogsci.soton.ac.uk/~harnad/Papers/Harnad/harnad82.consciousness.html

Harnad, S. (1984) Verifying machines' minds. (Review of J. T. Culbertson, Consciousness: Natural and artificial, NY: Libra 1982.) Contemporary Psychology 29: 389 - 391.

Harnad, S. (1985) Bugs, Slugs, Computers and Consciousness. American Psychologist 73(2): 21

Harnad, S. (1989) Minds, Machines and Searle. Journal of Theoretical and Experimental Artificial Intelligence 1: 5-25. http://www.cogsci.soton.ac.uk/~harnad/Papers/Harnad/harnad89.searle.html

Harnad, S. (1990a) The Symbol Grounding Problem. Physica D 42: 335-346. http://www.cogsci.soton.ac.uk/~harnad/Papers/Harnad/harnad90.sgproblem.html

Harnad, S. (1990b) Against Computational Hermeneutics. (Invited commentary on Eric Dietrich's Computationalism) Social Epistemology 4: 167-172. http://www.cogsci.soton.ac.uk/~harnad/Papers/Harnad/harnad90.dietrich.crit.html

Harnad, S. (1990c) Lost in the hermeneutic hall of mirrors. Invited Commentary on: Michael Dyer: Minds, Machines, Searle and Harnad. Journal of Experimental and Theoretical Artificial Intelligence 2: 321 - 327. http://www.cogsci.soton.ac.uk/~harnad/Papers/Harnad/harnad90.dyer.crit.html

Harnad, S. (1991) Other bodies, Other minds: A machine incarnation of an old philosophical problem. Minds and Machines 1: 43-54. http://www.cogsci.soton.ac.uk/~harnad/Papers/Harnad/harnad91.otherminds.html

Harnad, S. (1992a) Connecting Object to Symbol in Modeling Cognition. In: A. Clark and R. Lutz (Eds) Connectionism in Context. Springer Verlag, pp 75 - 90. http://www.cogsci.soton.ac.uk/~harnad/Papers/Harnad/harnad92.symbol.object.html

Harnad, S. (1992b) The Turing Test Is Not A Trick: Turing Indistinguishability Is A Scientific Criterion. SIGART Bulletin 3(4) (October) 9 - 10. http://www.cogsci.soton.ac.uk/~harnad/Papers/Harnad/harnad92.turing.html

Harnad, S. (1993a) Grounding Symbols in the Analog World with Neural Nets. Think 2(1) 12 - 78 (Special issue on "Connectionism versus Symbolism," D.M.W. Powers & P.A. Flach, eds.).
http://cwis.kub.nl/~fdl/research/ti/docs/think/2-1/index.stm

Harnad, S. (1993b) Artificial Life: Synthetic Versus Virtual. In: Artificial Life III. Proceedings of the Workshop on Artificial Life. Ed. Chris Langton. Santa Fe Institute Studies in the Sciences of Complexity. Volume XVI. Reading, Mass.: Addison-Wesley, Advanced Book Program 1994. http://www.cogsci.soton.ac.uk/~harnad/Papers/Harnad/harnad93.artlife.html

Harnad S. (1993c) Discussion (passim) In: Bock, G.R. & Marsh, J. (Eds.) Experimental and Theoretical Studies of Consciousness. CIBA Foundation Symposium 174. Chichester: Wiley http://www.cogsci.soton.ac.uk/~harnad/Papers/Harnad/harnad93.ciba.html

Harnad, S. (1994a) Levels of Functional Equivalence in Reverse Bioengineering: The Darwinian Turing Test for Artificial Life. Artificial Life 1(3): 293-301. Reprinted in: C.G. Langton (Ed.). Artificial Life: An Overview. MIT Press 1995. http://www.cogsci.soton.ac.uk/~harnad/Papers/Harnad/harnad94.artlife2.html

Harnad, S. (1994b) Computation Is Just Interpretable Symbol Manipulation: Cognition Isn't. Special Issue on "What Is Computation" Minds and Machines 4:379-390 http://www.cogsci.soton.ac.uk/~harnad/Papers/Harnad/harnad94.computation.cognition.html

Harnad, S, (1995a) Does the Mind Piggy-Back on Robotic and Symbolic Capacity? In: H. Morowitz (ed.) "The Mind, the Brain, and Complex Adaptive Systems." Santa Fe Institute Studies in the Sciences of Complexity. Volume XXII. P. 204-220. http://www.cogsci.soton.ac.uk/~harnad/Papers/Harnad/harnad95.mind.robot.html

Harnad, S. (1995b) Why and How We Are Not Zombies. Journal of Consciousness Studies 1: 164-167. http://www.cogsci.soton.ac.uk/~harnad/Papers/Harnad/harnad95.zombies.html

Harnad, S. (1996) The Origin of Words: A Psychophysical Hypothesis In Velichkovsky B & Rumbaugh, D. (Eds.) Communicating Meaning: Evolution and Development of Language. NJ: Erlbaum: pp 27-44. http://www.cogsci.soton.ac.uk/~harnad/Papers/Harnad/harnad96.word.origin.html

Harnad, S. (2000a) Turing Indistinguishability and the Blind Watchmaker. In: Mulhauser, G. (ed.) Evolving Consciousness Amsterdam: John Benjamins (in press) http://www.cogsci.soton.ac.uk/~harnad/Papers/Harnad/harnad98.turing.evol.html

Harnad, S. (2000b) What's Wrong and Right About Searle's Chinese Room Argument? (M. Bishop & J. Preston, eds.)
http://www.cogsci.soton.ac.uk/~harnad/Papers/Harnad/harnad00.searle.html
 

Harnad, S. (2000) Correlation VS. Causality: How/Why the Mind/Body Problem Is Hard. [Invited Commentary of Humphrey, N. "How to Solve the Mind-Body Problem"] Journal of Consciousness Studies 7(4): 54-61. http://www.cogsci.soton.ac.uk/~harnad/Papers/Harnad/harnad00.mind.humphrey.html

Harnad, S., Doty, R.W., Goldstein, L., Jaynes, J. & Krauthamer, G. (eds.) (1977) Lateralization in the nervous system. New York: Academic Press.

Hauser, L. (1993). Reaping the whirlwind: Reply to Harnad's "Other bodies, other minds" Minds and Machines, 3:219-237. http://members.aol.com/lshauser/harnad3.html

Hayes, P., Harnad, S., Perlis, D. & Block, N. (1992) Virtual Symposium on Virtual Mind. Minds and Machines 2: 217-238. http://www.cogsci.soton.ac.uk/~harnad/Papers/Harnad/harnad92.virtualmind.html

Heyes, C. M. (1998). Theory of mind in nonhuman primates. Behavioral and Brain Sciences 21 (1): 101-134. http://www.cogsci.soton.ac.uk/bbs/Archive/bbs.heyes.html

Loebner, H.G. (1994) Lessons from a restricted turing test - In response. Communications of the ACM 37(6): 79-82
http://www.loebner.net/Prizef/In-response.html

J.R. Lucas (1961) Minds, Machines and Goedel. Philosophy 36 112-127.
http://cogprints.soton.ac.uk/abs/phil/199807022

Mitchell, R.W. & Anderson, J.R. (1998) Primate theory of mind is a Turing test. Behavioral and Brain Sciences 21 (1) 127-128.

Nagel, T. (1974) What is is like to be a bat? Philosophical Review 83: 435-451.

Newell, A. (1980) Physical Symbol Systems. Cognitive Science 4: 135 - 83.

Pylyshyn, Z. W. (1980) Computation and cognition: Issues in the foundations of cognitive science. Behavioral and Brain Sciences 3: 111-169

Pylyshyn, Z. W. (1984) Computation and cognition. Cambridge MA: MIT/Bradford

Pylyshyn, Z. W. (Ed.) (1987) The robot's dilemma: The frame problem in artificial intelligence. Norwood NJ: Ablex

Schweizer, P. (1998). The Truly Total Turing Test. Minds and Machines, 8:263-272, 1998.

Searle, John. R. (1980) Minds, brains, and programs. Behavioral and Brain Sciences 3 (3): 417-457 http://www.cogsci.soton.ac.uk/bbs/Archive/bbs.searle2.html

Searle, John R. (1984). Minds, Brains and Science. Harvard University Press, Cambridge, Mass.

Searle, John R. (1993) The Failures of Computationalism. THINK 2(1) http://cwis.kub.nl/~fdl/research/ti/docs/think/2-1/searle.htm

Searle, John R. (1990) Consciousness, Explanatory Inversion, and Cognitive Science", Behavioral and Brain Sciences 13: 585-642

Shieber, Stuart M. (1994) Lessons from a restricted Turing test. Communications of the ACM 37, 70-78.
http://www.eecs.harvard.edu/shieber/papers/loebner-rev-html/loebner-rev-html.html

Steklis, H.D. & Harnad, S. (1976) From hand to mouth: Some critical stages in the evolution of language. In: Harnad, S., Steklis, H. D. & Lancaster, J. B. (eds.) Origins and Evolution of Language and Speech. Annals of the New York Academy of Sciences 280.

Turing, A.M. (1950) Computing Machinery and Intelligence. Mind 49 433-460 [Reprinted in Minds and machines. A. Anderson (ed.), Engelwood Cliffs NJ: Prentice Hall, 1964.]
http://cogprints.soton.ac.uk/abs/comp/199807017

Watt, S. (1996) Naive Psychology and the Inverted Turing Test. PSYCOLOQUY 7(14) http://www.cogsci.soton.ac.uk/psyc-bin/newpsy?7.14

Wimsatt, W. K. (1954) The verbal icon: studies in the meaning of poetry. University Press of Kentucky 1954.