Tirassa, 'IS CONSCIOUSNESS NECESSARY TO HIGH-LEVEL CONTROL SYSTEMS? (Book Review of Bringsjord on Robot-Consciousness)', Psycoloquy 941227 URL = http://hegel.lib.ncsu.edu/stacks/serials/psycoloquy/psycol-941227-tirassa-is Two commentaries on Bringsjord: Tirassa and Brown& O'Rourke psycoloquy.94.5.82.robot-consciousness.2.tirassa Fri 30 December 1994 ISSN 1055-0143 (19 paragraphs, 7 references, 225 lines) PSYCOLOQUY is sponsored by the American Psychological Association (APA) Copyright 1994 Maurizio Tirassa IS CONSCIOUSNESS NECESSARY TO HIGH-LEVEL CONTROL SYSTEMS? Book Review of Bringsjord on Robot-Consciousness Maurizio Tirassa Universita' di Milano Ist. Psicologia Fac. Medicina via Francesco Sforza, 23 20122 Milano (Italy) tirassa@imiucca.csi.unimi.it Universita' di Torino Centro di Scienza Cognitiva via G.L. Lagrange, 3 10123 Torino (Italy) tirassa@psych.unito.it ABSTRACT: Building on Bringsjord's (1992, 1994) and Searle's (1992) work, I take it for granted that computational systems cannot be conscious. In order to discuss the possibility that they might be able to pass refined versions of the Turing Test, I consider three possible relationships between consciousness and control systems in human-level adaptive agents. I. INTRODUCTION 1. Artificial Intelligence (AI) can be viewed as the attempt to build a person, under the specification that it (the resulting person) should be the necessary, predictable result of the carrying out of a set of computations. While there is no constraint on the nature of such computations, artificial brains cultured in nutrient media should be excluded from this definition. 2. Newell (1990) takes mind to be "the control system that guides the behaving organism in its complex interactions with the dynamic real world" (Newell 1990, p. 43). This definition is intended for artificial as well as natural minds. The aim of AI, thus, is to devise computational versions of such control systems, together with suitable sensors, effectors, etc. Let us dub "robots" the products of this enterprise. 3. We may (in principle) build robots of any kind; but, whatever their architecture, behavior, etc., may be, they can never be conscious. This is not a principled position, of course: consciousness is a matter of biology, not of computations; and that it cannot be a property of computational systems has been argued by, among others, Bringsjord (1992) and Searle (1992). Thus, there is no reason to expect artificial consciousness, unless we can duplicate the relevant properties of biological nervous tissue (connectionism, of course, offers no solution in this respect, since artificial neural nets are exactly as neural as so-called classical computational systems are). 4. Thus, we have at least one property of biological systems that cannot be duplicated in robots. Does this make any difference as to how the respective control systems work? Three positions may be taken with respect to the role of consciousness in the control systems of high- level adaptive agents: consciousness either (a) has nothing whatsoever to do with control, or it has; in the latter case, it may be either (b) a necessary feature or (c) a contingent one. In the following, I will briefly discuss these three positions. While position (a) seems unsustainable, currently there seems to be no reason to choose between positions (b) and (c). II. POSITION (a): CONSCIOUSNESS IS UNRELEVANT TO CONTROL 5. According to position (a), consciousness exerts no role at all in the control of our behavior, which is assured by nonconscious machinery. In spite of its widespread diffusion in cognitive science, this position is quite unreasonable, at least because of evolutionary concerns. 6. Human beings have a first-person (conscious) understanding of their own behavior as being, at least in part, guided by conscious deliberation. They also tend to interpret other individuals' behavior in the same perspective. This might be their mistake, of course: behavior might be completely independent of consciousness, which would then come only as a post hoc explanation to oneself of what is actually a sterile byproduct of one's nonconscious machinery. Hard as it may be to accept this idea, it is harder to prove it wrong. 7. Such a radical position is actually a common one in cognitive science, where it is usual to conceive of the mind as a bunch of boxes "processing" a flow of input information in order to determine an output. Since this is thought to be sufficient to produce intelligent behavior, consciousness is de facto perceived as irrelevant to the human kind. 8. It follows that either consciousness does not exist (which, as I can personally guarantee, is not the case), or it is superfluous. In the latter case, biological minds must be something more than control systems, because they exhibit a property which is completely unrelated to control. But then, under what selective pressure might such a property have evolved? It should be noticed that social behavior is a matter of control indeed. This means, first, that something completely unrelated to control can play no role in mating; and, second, that problems of control do not concern only low-level, possibly nonconscious, processes. 9. Thus, it would be quite amazing if consciousness turned out to be an extravagant byproduct of our brains. If this were the case, however, appropriately designed robots would be able to pass any version of the Turing Test, no matter how refined (Turing, 1950; Harnad, 1991): consciousness is a first-person property; were it a useless one too, how could we ever know whether the individual sitting in front of us was a robot? III. POSITION (b): CONSCIOUSNESS IS NECESSARY FOR (HIGH-LEVEL) CONTROL 10. At the other end of the continuum is the idea that consciousness is a necessary feature of high-level control systems (position (b)). In this perspective, our being conscious, far from being an extravagant luxury, would be intimately connected to the distinctive behavioral performance exhibited by our species (and possibly by others as well). 11. This would be in accordance with evolutionary concerns: complex- minded species need to be conscious, and that's all. The question of whether this pertains only to the human species as well as whether there are middle steps between the nonconscious and the conscious, should be viewed as empirical matters, to be settled with yet-to-be-devised methodologies. In this case, no robot, however accurately designed, could ever pass a refined version of the Turing Test: some predictable detail, though possibly an imperceptible one, would be guaranteed to betray the artificial. 12. The problem here is that we do not have the slightest idea of what consciousness is for. As soon as we have a theory of the role it might play in some mental process, we are ipso facto excluding it from that very process, thereby restricting it to all the processes we have not yet understood. This does not imply that consciousness does not exist. Rather, consciousness is a first-person property, whereas scientific theories must be stated in the third person: if we could devise a third-person theory of consciousness, we would simply rule out the need for consciousness itself. The problem is, of course, that as a matter of fact human beings are conscious. 13. This is not meant to imply that we cannot theorize about consciousness: theories of consciousness might be possible in principle, but they may require some major shift in our conception of what a cognitive theory is. IV. POSITION (c): CONSCIOUSNESS IS RELEVANT BUT NOT NECESSARY TO CONTROL 14. According to position (c), consciousness is relevant to control, but not necessary. We must acknowledge, as a matter of fact, its crucial role in the control of our behavior, but we cannot exclude the possibility that analogous results may be achieved with different methods. 15. In this perspective, the role of consciousness is restricted to biological agents, leaving room for computational ones which be unconscious but able to behave in a human-like fashion. There might exist nonconscious forms of control which work equally well, and with exactly the same results, in producing high-level behavior. The difference would be that humans entertain conscious, first-person mental states, whereas robots should happily limit themselves to third-person computations over mental states equivalents (whatever that might mean). This is Bringsjord's thesis (Bringsjord, 1992), as I understand it. 16. Such a position rules out neither the need for suitable theories of biological consciousness (i.e., the need for a change in the postulates underlying the psychological branches of cognitive science), nor the possibility for robots to pass however refined versions of the Turing Test. 17. In spite of my personal sympathy for this position, I can adduce no empirical support in its favor. An empirical, though obviously weak, case against it might be found in the relatively poor accomplishments of current AI, if viewed from the psychologist's viewpoint: no existing system can be said to reach the level of complexity of even an insect. The hope is that this is not the symptom of the difficulties I have described under position (b). As is common in science, only time will tell. V. CONCLUSION 18. Both positions (b) and (c) imply a separation between AI and psychology; the cleavage, however, would follow different lines. In the (b) case, robots could never aspire to great accomplishments: they would be limited to low-level behaviors, such as those exhibited by presumably nonconscious animals (insects? reptiles?). Accordingly, relationships should be found between AI and ethology rather than psychology or animal cognition. 19. In the (c) case, there would be no principled limits to robots' possibilities; there might be relationships between AI and psychology, though far from those usually conceived. Rather than considering the similarities in the computations carried out, as is usually recommended (see, e.g., Pylyshyn, 1984), it might be more interesting to study issues like architectural requirements for efficient control, requirements on initial knowledge, etc. Once we have accepted this idea, the quest for Turing-testable robots becomes meaningless: Turing-testable behaviors are those of our species, and there is no (scientific) point in trying to simulate them when we know that computational consciousness is impossible. Different species will have different abilities, so why not disconnect Turing machines from the Turing test and study robots as an evolutionary line independent from ours? REFERENCES Bringsjord, S. (1992) What Robots Can and Can't Be. Boston: Kluwer Academic. Bringsjord, S. (1994) Precis of: What Robots Can and Can't Be. PSYCOLOQUY 5(59) robot-consciousness.1.bringsjord. Harnad, S. (1991) Other bodies, other minds: A machine incarnation of an old philosophical problem. Minds and Machines 1:43-54. Newell, A. (1990) Unified theories of cognition. Boston: Harvard University Press. Pylyshyn, Z.W. (1984) Computation and cognition. Boston: MIT Press. Searle, J.R. (1992) The rediscovery of the mind. Boston: MIT Press. Turing, A.M. (1950) Computing machinery and intelligence. Mind 59:433- 460. ---------------------------------------------------------------------- psycoloquy.94.5.83.robot-consciousness.3.brown Fri 30 December 1994 ISSN 1055-0143 (11 paragraphs, 9 references, 171 lines) PSYCOLOQUY is sponsored by the American Psychological Association (APA) Copyright 1994 Brown & O'Rourke AGNOSTICISM ABOUT THE ARBITRARY REALIZATION ARGUMENT Book Review of Bringsjord on Robot-Consciousness Marina Brown and Joseph O'Rourke Dept. of Computer Science Smith College Northampton, MA 01063 orourke@cs.smith.edu ABSTRACT: We argue that Bringsjord (1992, 1994) should be agnostic about his version (ARA1) of Block's Arbitrary Realization Argument, because Bringsjord admits to agnosticism about low-level functionalism and low-level functionalism implies that the key claim of ARA1 is false. I. INTRODUCTION 1. There is much to admire (and dispute!) in Bringsjord's stimulating book, What Robots Can and Can't Be (1992). We choose to focus on an argument that is key to Bringsjord's rejection of "Artificial Intelligence Functionalism" (AI-F), and hence to his position that persons are not automata. Our comments are directed to the Arbitrary Realization Argument (ARA), and in particular, to ARA1, Bringsjord's version of Block's "Chinese Nation" argument (Block, 1978). 2. In Chapter IV of his book, Bringsjord "heartily agree[s]" with Pollack's "evidence for functionalism, generically conceived" (p. 125). But he remains unmoved because of the "formidable deductive arguments against (AI-F)" (p. 125) in Chapter VI, which are the various Arbitrary Realization Arguments. Argument ARA1 tells a Block-like story, replacing Block's Chinese by Norwegians implementing a Turing Machine in Texas. Although the argument is indeed deductive, it hangs entirely on one claim (1_VI): the Turing Machine M implemented by the Norwegians in Texas does not constitute an agent with the appropriate mental state ("fearing purple unicorns" in his example). We would like to argue that, rather than believing that "ARA has once and for all demolished AI-Functionalism" (p.223), Bringsjord should be agnostic with respect to claim 1_VI and therefore not view ARA1 as a refutation of AI-Functionalism. II. ARGUMENT 3. The following is a brief synopsis of our argument: (1) Bringsjord admits to being agnostic about low-level functionalism. (2) Low-level functionalism is plausible. (3) Low-level functionalism, with a few more plausible assumptions, implies claim 1_VI is false. (4) So Bringsjord should be an agnostic about 1_VI. (5) So Bringsjord should not find ARA1 compelling. We now detail the argument, referring to the above points as "Step (1)," etc. 4. Step (1): Bringsjord's agnosticism about "low-level functionalism." In his parenthetical mention of the "scenario in which neurons are replaced one at a time by silicon-based work-alikes," Bringsjord says he is "agnostic" about this argument's presuppositions: materialism and low-level functionalism (p. 218). 5. Step (2): Low-level functionalism is plausible. Bringsjord's agnosticism is a reasonable position to hold, but we are more inclined to view low-level functionalism as quite likely true, albeit not yet firmly established. Let us distinguish several levels of functionalism, all varieties of low-level functionalism: a. Particle-level functionalism. If every electron in a brain were replaced by a particle that is functionally equivalent to an electron, it is reasonable to suppose that the mental states would not change. There is every reason to believe that an anti- matter brain (in an anti-matter body, in an anti-matter universe), would operate the same as our brains. b. Atom-level functionalism. Neuroscientists study the brain's use of certain substances by labeling them with rare isotopes that can be detected more easily than the naturally occurring isotopes (Kauppinen, Williams, Busza and Bruggen, 1993). Underlying this procedure is the assumption that isotopes of an element can be substituted for one another without interfering with their function. c. Molecule-level functionalism. Synthetic molecules with certain shapes and chemical properties function in the brain. For example, various synthetic drugs are known to mimic several of the functional roles of opioid peptides. d. Neuron-level functionalism. A number of authors have considered the scenario of replacing neurons with "silicon-based work-alikes" in Bringsjord's phrase: Cole & Foelber (1984) (whom Bringsjord cites), Cuda (1985), Chalmers (1993). Especially when the scenario is imagined with some detail (as by Chalmers), we find it quite plausible that the overall function of the brain would not be affected by such replacement. It is clear that it is neuron-level functionalism about which Bringsjord is agnostic. 6. Step (3): Neuron-level functionalism implies 1_VI. Since Bringsjord is agnostic about neuron-level functionalism, he should be agnostic about its consequences. In the following paragraphs, we will state a few plausible hypotheses and draw conclusions. 7. Hypothesis (a): The mental aspects of the brain are determined by its pattern of neuron activity. This is an assumption underlying much research in neuroscience (e.g.. as stated explicitly by Changeux and Dehaene, 1989). Therefore, replacement of all the neurons in the brain by silicon "work-alikes" should maintain the mental aspects of the brain. Now we can view the silicon brain as a particular digital computer. 8. Hypothesis (b): The functionality of a computer design is not dependent on the physical material from which it is composed. Computers have been built out of silicon, brass gears (Babbage's "Analytical Engine"), fiber optics, even Tinkertoys (Dewdney, 1993). Therefore replacement of the silicon brain-computer with one made of some other material should maintain the mental aspects. 9. Hypothesis (c): The functionality of a computer design is not dependent on its size. Bringsjord is "prepared to admit, for the sake of argument, that size can't make a difference," although he does have "some inchoate reservations" (p.214). But certainly that the size of computers doesn't make a difference is well-established by advances in miniaturization. Therefore, replacement of the brain-computer with Norwegians in Texas should maintain the mental aspects. 10. Step (4): Since Hypotheses (a), (b), and (c) above are all plausible, Bringsjord should be agnostic about his 1_VI. If his "inchoate reservations" about size are a sticking point, then in the absence of (c) he should be agnostic about "same-size" AI-Functionalism, a version of AI-F that demands that the functionally equivalent systems be the same size. 11. Step (5): Rather than finding the ARA1 argument (which rests entirely on 1_VI) "formidable," "demolishing" AI-Functionalism, Bringsjord's agnosticism about low-level functionalism should extend to ARA1, and he should be as unmoved by it as we are. REFERENCES Block, N. (1978) Troubles with Functionalism. In Readings in the Philosophy of Psychology, Vol. 1, Harvard University Press. Bringsjord, S. (1992) What Robots Can and Can't Be. Boston: Kluwer Academic. Bringsjord, S. (1994) Precis of: What Robots Can and Can't Be. PSYCOLOQUY 5(59) robot-consciousness.1.bringsjord. Chalmers, D.J. (1993) Toward a theory of consciousness, Center for Research on Concepts and Cognition, Indiana University. Changeux, J.-P. and Dehaene, S. (1989) Neuronal models of cognitive functions, Cognition 33:63-109. Cole, D. and Foelber, R. (1984) Contingent materialism, Pacific Philosophical Quarterly 65.1: 74-85. Cuda, T. (1985) Against neural chauvinism, Philosophical Studies 48: 111-127. Dewdney, A.K. (1993) The Tinkertoy Computer, W.H. Freeman, 7-15. Kauppinen, R.A., Williams, S.R., Busza, A.L., and van Bruggen, N. (1993) Applications of magnetic resonance spectroscopy and diffusion- weighted imaging to the study of brain biochemistry and pathology, Trends in Neurosciences Vol. 16, No. 3, 91-92.