ABSTRACT: Artificial life can take two forms: synthetic and virtual. In principle, the materials and properties of synthetic living systems could differ radically from those of natural living systems yet still resemble them enough to be really alive if they are grounded in the relevant causal interactions with the real world. Virtual (purely computational) "living" systems, in contrast, are just ungrounded symbol systems that are systematically interpretable as if they were alive; in reality they are no more alive than a virtual furnace is hot. Virtual systems are better viewed as "symbolic oracles" that can be used (interpreted) to predict and explain real systems, but not to instantiate them. The vitalistic overinterpretation of virtual life is related to the animistic overinterpretation of virtual minds and is probably based on an implicit (and possibly erroneous) intuition that living things have actual or potential mental lives.
It is for this reason that I believe all positive analogies between the biology of life and the biology of mind are bound to fail: We are invited by some (e.g., Churchland 1984, 1986, 1989) to learn a lesson from how wrong-headed it had been to believe that ordinary physics alone could not explain life -- which has now turned out to be largely a matter of protein chemistry -- and we are enjoined to apply that lesson to current attempts to explain the mind -- whether in terms of physics, biochemistry, or computation -- and to dismiss the conceptual dissatisfaction we continue to feel with such explanations on the grounds that we've been wrong in much the same way before.
But the facts of the matter are actually quite the opposite, I think: Our real mistake had been in inadvertently conflating the problem of life and the problem of mind, for in bracketing mind completely when we consider life, as we do now, we realize there is nothing left that is special about life relative to other physical phenomena -- at least nothing more special than whatever is special about biochemistry relative to other branches of chemistry. But in the case of mind itself, what are we to bracket?
Now let me point out the disanalogy between the special empirical and conceptual situation just described for the case of mind modelling and the parallel case of life-modelling, which involves instead a true, complete description of the kind of physical system that is alive; then, for good measure, this disanalogy will be extended to the third parallel case, matter-modelling, this time involving a true complete description of any physical system at all (e.g., the world of elementary particles or even the universe as a whole): If we were given a true, complete description of the kind of physical system that is alive, could we still ask (a) why the very same description would not be equally true if the system were not alive, but just looked and behaved exactly as if it were alive? And could we go on to ask (b) how we could possibly know that every (or any) system that fit that description was really alive? I suggest that we could not raise these questions (apart from general worries about the truth or completeness of the description itself, which, it is important to note, is not what is at issue here) because in the case of life there just is no further fact -- like the fact of the existence of mental states -- that exists independently of the true complete physical description.
Consider the analogous case of physics: Could we ask whether, say, the planets really move according to relativistic mechanics, or merely look and behave exactly as if they did? I suggest that this kind of question is merely about normal scientific underdetermination -- the uncertainty there will always be about whether any empirical theory we have is indeed true and complete. But if we assume it as given (for the sake of argument) that a theory in physics is true and complete, then there is no longer any room left for any conceptual distinction between the "real" and "as-if" case, because there is no further fact of the matter that distinguishes them. The facts -- all objective, physical ones in the case of both physics and biology -- have been completely exhausted.
Not so in the case of mind, where a subjective fact, and a fact about subjectivity, still remains, and remains unaccounted for (Nagel 1974, 1986). For although the force of the premise that our description is true and complete logically excludes the possibility that any system that fits that description will fail to have a mind, the conceptual possibility is not only there, but amounts to a complete mystery as to why the description should be true and complete at all (even if it is); for there is certainly nothing in the description itself -- which is all objective and physical -- that either entails or even shows why it is probable that mental states should exist at all, let alone be accounted for by the description.
I suggest that it is this elusive extra (mental) fact of the matter that made people wrongly believe that special vital forces, rather than just physics, would be needed to account for life, But, modulo the problem of explaining mind (setting aside, for the time being, that mind too is a province of biology), the problem of explaining life (or matter) has no such extra fact to worry about. So there is no justification for being any more skeptical about present or future theories of life (or matter) than is called for by ordinary considerations of underdetermination, whose resolution is rightly relativized and relegated to the arena of rival theories fighting it out to see which will account for the most data, the most generally (and perhaps the most economically).
The special case I am referring to is Searle's (1980) infamous Chinese Room Argument against computationalism (or the computational theory of mind). I am one of the tiny minority (possibly as few as two) who think Searle's Argument is absolutely right. Computationalism (Dietrich 1990) holds that mental states are just computational states -- not just any computational states, of course, only certain special ones, in particular, those that are sufficient to pass the Turing Test (TT), which calls for the candidate computational system to be able to correspond with us as a pen pal for a lifetime, indistinguishable from a real pen pal. The critical property of computationalism that makes it susceptible to empirical refutation by Searle's thought experiment is the very property that had made it attractive to mind-modellers: its implementation-independence. According to computationalism, all the physical details of the implementation of the right computational states are irrelevant, just as long as they are implementations of the right computational states; for then each and every implementation will have the right mental states. It is this property of implementation-independence -- a property that has seemed to some theorists (e.g. Pylyshyn 1984) to be the kind of dissociation from the physical that might even represent a solution to the mind/body problem -- that Searle exploits in his Chinese Room Argument. He points out that he too could become an implementation of the TT-passing computer program -- after all, the only thing a computer program is really doing is manipulating symbols on the basis of their shapes -- by memorizing and executing all the symbol manipulation rules himself. He could do this even for the (hypothetical) computer program that could pass the TT in Chinese, yet he obviously would not be understanding Chinese under these conditions; hence, by transitivity of implementation-independent properties (or their absence), neither would the (hypothetical) computer implementation be understanding Chinese (or anything). So much for the computational theory of mind.
It's important to understand what Searle's argument does and does not show. It does not show that the TT-passing computer cannot possibly be understanding Chinese. Nothing could show that, because of the other-minds problem (Harnad 1984, 1991): The only way to know that for sure would be to be the computer. The computer might be understanding Chinese, but if so, then that could only be because of some details of its particular physical implementation (that its parts were made of silicon, maybe?), which would flatly contradict the implementation-independence premise of computationalism (and leave us wondering about what's special about silicon). Searle's argument also does not show that Searle himself could not possibly be understanding Chinese under those conditions. But -- unless we are prepared to believe either in the possibility that (1) memorizing and manipulating a bunch of meaningless symbols could induce multiple personality disorder (a condition ordinarily caused only by early child abuse), giving rise to a second, Chinese-understanding mind in Searle, an understanding of which he was not consciously aware (Dyer 1990, Harnad 1990c, Hayes et al 1992), or, even more far-fetched, that (2) memorizing and manipulating a bunch of meaningless symbols could render them consciously understandable to Searle -- the emergence of either form of understanding under such conditions is, by ordinary inductive standards, about as likely as the emergence of clairvoyance (which is likewise not impossible).
So I take it that, short of sci-fi special pleading, Searle's Argument that there is nobody home in there is valid for the very circumscribed special case of any system that is purported to have a mind purely in virtue of being an implementation-independent implementation of a TT-passing computer program. The reason his argument is valid is also clear: Computation is just the manipulation of physically implemented symbols on the basis of their shapes (which are arbitrary in relation to what they can be interpreted to mean); the symbols and symbol manipulations will indeed bear the weight of a systematic interpretation, but that interpretation is not intrinsic to the system (any more than the interpretation of the symbols in a book is intrinsic to the book). It is projected onto the system by an interpreter with a mind (as in the case of the real Chinese pen-pal of the TT-passing computer); the symbols don't mean anything to the system, because a symbol-manipulating system is not the kind of thing that anything means anything to.
There remains, however, the remarkable property of computation that makes it so valuable, namely, that the right computational system will indeed bear the weight of a systematic semantic interpretation (perhaps even the TT). Such a property is not to be sneezed at, and Searle does not sneeze at it. He calls it "Weak Artificial Intelligence" (to contrast it with "Strong Artificial Intelligence," which is the form of computationalism his argument has, I suggest, refuted). The practitioners of Weak AI would be studying the mind -- perhaps even arriving at a true, complete description of it -- using computer models; they could simply never claim that their computer models actually had minds.
This is also the correct way to think of a (hypothetical) TT-passing computer program. It would really just be a symbol system that was systematically interpretable as if it were a pen-pal corresponding with someone. This virtual pen-pal may be able to predict correctly all the words and thoughts of the real pen-pal it was simulating, oracularly, till doomsday, but in doing so it is no more thinking or understanding than the virtual solar system is moving. The erroneous idea that there is any fundamental difference between these two cases (the mental model and the planetary model) is, I suggest, based purely on the incidental fact that thinking is unobservable, whereas moving is observable; but gravity is not observable either, and quarks and superstrings even less so; yet none of those would be present in a virtual universe either. For the virtual universe, like the virtual mind, would really just be a bunch of meaningless symbols -- squiggles and squoggles -- that were syntactically manipulated in such a way as to be systematically interpretable as if they thought or moved, respectively. This is certainly stirring testimony to the power of computation to describe and predict physical phenomena and to the validity of the Church/Turing Thesis (Davis 1958; Kleene 1969), but it is no more than that. It certainly is not evidence that thinking is just a form of computation.
So computation is a powerful, indeed oracular, tool for modelling, predicting and explaining planetary motion, life and mind, but it is not a powerful enough tool for actually implementing planetary motion, life or mind, because planetary motion, life and mind are not mere implementation-independent computational phenomena. I will shortly return to the question of what might be powerful enough in its place, but first let me try to tie these considerations closer to the immediate concerns of this conference on artificial life.
Well the answer is quite simple, as long as we don't let the power of hermeneutics loosen our grip on one crucial distinction: the distinction between objects and symbolic descriptions of them. There may be a one-to-one correspondence between object and symbolic description, a description that is as fine-grained as you like, perhaps even as fine-grained as can be. But the correspondence is between properties that, in the case of the real object, are what they are intrinsically, without the mediation of any interpretation, whereas in the case of the symbol description, the only "objects" are the physical symbol tokens and their syntactically constrained interactions; the rest is just our interpretation of the symbols and interactions as if they were the properties of the objects they describe. If the description is complete and correct, it will always bear the full weight of that interpretation, but that still does not make the corresponding properties in the object and the description identical properties; they are merely computationally equivalent under the mediation of the interpretation. This should be only slightly harder to see in the case of a dynamic simulation of the universe than it is in the case of a book full of static sentences about the universe.
In the case of real heat and virtual heat or real motion and virtual motion the distinction between identity and formal equivalence is clear. A computer simulation of a fire is not really hot and nothing is really moving in a computer simulation of planetary motion. In the case of thinking I have already argued that we have no justification for claiming an exception merely because thinking is unobservable (and besides, Searle's "periscope" shows us that we wouldn't find thinking in there even if we became the simulation and observed for ourselves what is unobservable to everyone else). What about life? But we have already seen that life -- once we bracket the mind/body problem -- is not different from any other physical phenomenon, so virtual life is no more alive than virtual planetary motion moves or virtual gravity attracts. Chris Langton's virtual biosphere, in other words, would be just another symbolic oracle: yet another bunch of systematically interpretable squiggles and squoggles: Shall we call this "Weak Artificial Life"? (Sober 1992, as I learned after writing this paper, had already made this suggestion at the Artificial Life II meeting in 1990).
I did promise to return to the question of what more might be needed to implement life if computation, for all its power and universality, is not strong enough. Again, we return to the Chinese Room Argument for a hint, but this time we are interested in what kind of system Searle's argument cannot successfully show to be mindless: We do not have to look far. Even an optical transducer is immune to Searle, for if someone claimed to have a system that could really see (as opposed to being merely interpretable as if it could see), Searle's "periscope" would already fail (Harnad 1989). For if Searle tried to implement the putative "seeing" system without seeing (as he had implemented the putative "understanding" system without understanding), he would have only two choices. One would be to implement only the output of the transducer (if we can assume that its output would be symbols) and whatever symbol manipulations were to be done on that output, but then it would not be surprising if Searle reported that he could not see, for he would not be implementing the whole system, just a part of it. All bets are off when only parts of systems can be implemented. Searle's other alternative would be to actually look at the system's scene or screen in implementing it, but then, alas, he would be seeing. Either way, the Chinese Room strategy does not work. Why? Because mere optical transduction is an instance of the many things there are under the sun -- including touch, motion, heat, growth, metabolism, photosynthesis, and countless other "analog" functions -- that are not just implementations of implementation-independent symbol manipulations. And only the latter are vulnerable to Searle's Argument.
This immediately suggests a more exacting variant of the Turing Test -- what I've called the Total Turing Test (TTT) -- which, unlike the TT, is not only immune to Searle's Argument but reduces the level of underdetermination of mind-modelling to the normal level of underdetermination of scientific theory (Harnad 1989, 1991). The TT clearly has too many degrees of freedom, for we all know perfectly well that there's a lot more that people can do than be pen-pals. The TT draws on our linguistic capacity, but what about all our robotic capacities, our capacity to discriminate, identify and manipulate the objects, events and states of affairs in the world we live in? Every one of us can do that; so can animals (Harnad 1987; Harnad et al. 1991). Why should we have thought that a system deserved to be assumed to have a mind if all it could generate was our pen-pal capacity? Not to mention that there is good reason to believe that our linguistic capacities are grounded in our robotic capacities. We don't just trade pen-pal symbols with one another; we can each also identify and describe the objects we see, hear and touch, and there is a systematic coherence between how we interact with them robotically and what we say about them linguistically.
My own diagnosis is that the problem with purely computational models is that they are ungrounded. There may be symbols in there that are systematically interpretable as meaning "cat," "mat" and "the cat is on the mat," but in reality they are just meaningless squiggles and squoggles apart from the interpretation we project onto them. In an earlier conference in New Mexico (Harnad 1990a) I suggested that the symbols in a symbol system are ungrounded in much the same way the symbols in a Chinese-Chinese dictionary are ungrounded for a non-speaker of Chinese: He could cycle through it endlessly without arriving at meanings unless he already had grounded meanings to begin with (as provided by a Chinese-English dictionary, for an English speaker). Indeed, translation is precisely what we are doing when we interpret symbol systems, whether static or dynamic ones, and that's fine, as long as we are using them only as oracles, to help us predict and explain things. For that, their systematic interpretability is quite sufficient. But if we actually want them to implement the things they predict and explain, they must do a good deal more. At the very least, the meanings of the symbols must somehow be grounded in a way that is independent of our projected interpretations and in no way mediated by them.
The TTT accordingly requires the candidate system, which is now a robot rather than just a computer, to be able to interact robotically with (i.e., to discriminate, identify, manipulate and describe) the objects, events and states of affairs that its symbols are systematically interpretable as denoting, and it must be able to do it so its symbolic performance coheres systematically with its robotic performance. In other words, not only must it be capable, as any pen-pal would be, of talking about cats, mats, and cats on mats indistinguishably from the way we do, but it must also be capable of discriminating, identifying, and manipulating cats, mats, and cats on mats exactly as any of us can; and its symbolic performance must square fully with its robotic performance, just as ours does (Harnad 1992).
The TT was a tall order; the TTT is an even taller one. But note that whatever candidate successfully fills the order cannot be just a symbol system. Transduction and other forms of analog performance and processing will be essential components in its functional capacity, and subtracting them will amount to reducing what's left to those mindless squiggles and squoggles we've kept coming up against repeatedly in this discussion. Nor is this added performance capacity an arbitrary demand. The TTT is just normal empiricism. Why should we settle for a candidate that has less than our full performance capacities? The proper time to scale the model down to capture our handicaps and deficits is only after we are sure we've captured our total positive capacity (with the accompanying hope that a mind will piggy-back on it) -- otherwise we would be like automotive (reverse) engineers (i.e., theoretical engineers who don't yet know how cars work but have real cars to study) who were prepared to settle for a functional model that has only the performance capacities of a car without moving parts, or a car without gas: The degrees of freedom of such "handicapped" modelling would be too great; one could conceivably go on theory-building forever without ever converging on real automobile performance capacity that way.
A similar methodological problem unfortunately also affects the TTT modelling of lower organisms: If we knew enough about them ecologically and psychologically to be able to say with any confidence what their respective TTT capacities were, and whether we had captured them TTT-indistinguishably, lower organisms would be the ideal place to start; but unfortunately we do not know enough, either ecologically or psychologically (although attempts to approximate the TTT capacities of lower organisms will probably still have to precede or at least proceed apace with our attempts to capture human TTT capacity).
The world of objects and the physics of transducing energy from them provide the requisite constraints for mind-modelling, and every solution that manages to generate our TTT capacity within those constraints has (by my lights) equal claim to our faith that it has mental states -- I don't really see the one that is TTTT-indistinguishable from us as significantly outshining the rest. My reasons for believing this are simple: We are blind to Turing indistinguishable differences (that's why there's an other-minds problem and a mind/body problem). By precisely the same token, the Blind Watchmaker is likewise blind to such differences. There cannot have been independent selection pressure for having a mind, since selection pressure can operate directly only on TTT capacity.
Yet a case might be made for the TTTT if the capacity to survive, reproduce and propagate one's genes is an essential part of our TTT capacity, for that narrows down the range of eligible transducers still further, and these are differences that evolution is not blind to. In this case, the TTTT might pick out microstructural features that are too subtle to be reflected in individual behavioral capacity, and the TTT look-alikes lacking them might indeed lack a mind (cf. Morowitz 1992).
My own hunch is nevertheless that the TTT is strong enough on its own (although neuroscience could conceivably give us some clues as to how to pass it), and I'm prepared to extend human rights to any synthetic candidate that passes it, because the TTT provides the requisite constraints for grounding symbols, and that's strong enough grounds for me. I doubt, however, that TTT capacity can be second-guessed a priori, even with the help of symbolic oracles. Perhaps this is the point where we should stop pretending that mind-modellers can bracket life and that life-modelers can bracket mind. So far, only living creatures seem to have minds. Perhaps the constraints on the creation of synthetic life will be relevant to the constraints on creating synthetic minds.
These last considerations amount only to fanciful speculation, however; the only lesson Artificial Life might take from this paper is that it is a mistake to be too taken with symbol systems that display the formal properties of living things. It is not that only natural life is possible; perhaps there can be synthetic life too, made of radically different materials, operating on radically different functional principles. The only thing that is ruled out is "virtual" or purely computational life, because life (like mind and matter) is not just a matter of interpretation.
Churchland, P. M. (1989) A neurocomputational perspective: the nature of mind and the structure of science Cambridge, MA: MIT Press, 1989.
Churchland, P. S. (1986) Neurophilosophy: toward a unified science of the mind-brain Cambridge, Mass.: MIT Press, 1986.
Davis, M. (1958) Computability and unsolvability. Manchester: McGraw-Hill.
Dietrich, E. (1990) Computationalism. Social Epistemology 4: 135 - 154.
Dyer, M. G. Intentionality and Computationalism: Minds, Machines, Searle and Harnad. Journal of Experimental and Theoretical Artificial Intelligence, Vol. 2, No. 4, 1990.
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.
Harnad, S. (1982b) Consciousness: An afterthought. Cognition and Brain Theory 5: 29 - 47.
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. (1987) The induction and representation of categories. In: Harnad, S. (ed.) (1987) Categorical Perception: The Groundwork of Cognition. New York: Cambridge University Press.
Harnad, S. (1989) Minds, Machines and Searle. Journal of Theoretical and Experimental Artificial Intelligence 1: 5-25.
Harnad, S. (1990a) The Symbol Grounding Problem. Physica D 42: 335-346.
Harnad, S. (1990b) Against Computational Hermeneutics. (Invited commentary on Eric Dietrich's Computationalism) Social Epistemology 4: 167-172.
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.
Harnad, S. (1991) Other bodies, Other minds: A machine incarnation of an old philosophical problem. Minds and Machines 1: 43-54.
Harnad, S. (1992) Connecting Object to Symbol in Modeling Cognition. In: A. Clarke and R. Lutz (Eds) Connectionism in Context Springer Verlag.
Harnad, S. (1993) Grounding Symbols in the Analog World with Neural Nets. Think 2: 12 - 78 (Special Issue on "Connectionism versus Symbolism" D.M.W. Powers & P.A. Flach, eds.).
Harnad, S., Hanson, S.J. & Lubin, J. (1991) Categorical Perception and the Evolution of Supervised Learning in Neural Nets. In: Working Papers of the AAAI Spring Symposium on Machine Learning of Natural Language and Ontology (DW Powers & L Reeker, Eds.) pp. 65-74. Presented at Symposium on Symbol Grounding: Problems and Practice, Stanford University, March 1991; also reprinted as Document D91-09, Deutsches Forschungszentrum fur Kuenstliche Intelligenz GmbH Kaiserslautern FRG.
Hayes, P., Harnad, S., Perlis, D. & Block, N. (1992) Virtual Symposium on the Virtual Mind. Minds and Machines (in press)
Kleene, S. C. (1969) Formalized recursive functionals and formalized realizability. Providence. American Mathematical Society.
MacLennan, B. J. (1987) Technology independent design of neurocomputers: The universal field computer. In M. Caudill & C. Butler (Eds.), Proceedings, IEEE First International Conference on Neural Networks (Vol. 3, pp. 39-49). New York, NY: Institute of Electrical and Electronic Engineers.
MacLennan, B. J. (1988) Logic for the new AI. In J. H. Fetzer (Ed.), Aspects of Artificial Intelligence (pp. 163-192). Dordrecht: Kluwer.
MacLennan, B. J. (in press-a) Continuous symbol systems: The logic of connectionism. In Daniel S. Levine and Manuel Aparicio IV (Eds.), Neural Networks for Knowledge Representation and Inference. Hillsdale, NJ: Lawrence Erlbaum.
MacLennan, B. J. (in press-b) Characteristics of connectionist knowledge representation. Information Sciences, to appear.
MacLennan, B. J. (1993) Grounding Analog Computers. Think 2: 48-51.
Morowitz, H. (1992) Beginning of Cellular Life. Yale University Press.
Nagel, T. (1974) What is it like to be a bat? Philosophical Review 83: 435 - 451.
Nagel, T. (1986) The view from nowhere. New York: Oxford University Press.
Newell, A. (1980) Physical Symbol Systems. Cognitive Science 4: 135 - 83.
Pylyshyn, Z. W. (1984) Computation and cognition. Cambridge MA: Bradford Books
Searle, J. R. (1980) Minds, brains and programs. Behavioral and Brain Sciences 3: 417-424.
Sober, E. (1992) Learning from functionalism: Prospects for strong AL. In: C.G Langton (Ed.) Artificial Life II. Redwood City, Calif.: Addison-Wesley
Turing, A. M. (1964) Computing machinery and intelligence. In: Minds and machines, A . Anderson (ed.), Engelwood Cliffs NJ: Prentice Hall.