The field of consciousness studies has been quite busy lately. There seem to be two major camps on the mind-body problem: those who believe that we already have the tools we need to explain the mind, and those who believe that we don't and perhaps never will. McGinn is in the pessimistic camp, as is Nagel (1975) and others. I'm an optimist.
Some theorists (notably Bernard Baars 1988,1996) focus on explaining the function of consciousness and how this function might be realized by the structures of the brain. Although this is an important area, it won't be my main focus. I agree with Chalmers's assessment (Chalmers 1996) that the "hard problem" of consciousness is to explain how a physical object can have experiences. This is the problem of phenomenal consciousness. It is a hard problem for all theories, but especially for computational ones.
Like everyone else, I can't define "phenomenal consciousness." It's the ability to have experiences. I assume that anyone who can read this knows what it means to experience something (from their own experience!); and that everyone knows that thermostats don't have experiences, even though they can react to temperature differences.
The difficulty of defining consciousness has led some to propose that there is no thing as consciousness. The standard citations are to Churchland (1990) and Rorty (1965). This position is called elimina-tivism. The idea is to replace the concept of consciousness with more refined (more scientific?) concepts, much as happened with concepts like "energy" and "mass" in past scientific revolutions. It seems plain that a full understanding of the mind will involve shifts of this kind. If we ever do achieve fuller understanding (which the pessimists doubt), any book written before the resulting shift, including this one, will no doubt seem laughably quaint. However, we can't simply wait around for this to happen. We have to work on the problems we see now, using the tools at hand. There is clearly a problem of how a thing, a brain or computer, can have experiences, or appear to have what people are strongly tempted to call experiences. To explain how this is conceivable at all must be our goal.
O'Brien and Opie (1999) make a useful distinction between vehicle and process theories of phenomenal consciousness:
Either consciousness is to be explained in terms of the nature of the representational vehicles the brain deploys, or it is to be explained in terms of the computational processes defined over these vehicles. We call versions of these two approaches VEHICLE and PROCESS theories of consciousness, respectively.
A process theory is one that explains experience in terms of things the brain does, especially things it computes. A vehicle theory explains experience as a property of the entity doing the computing. O'Brien and Opie themselves propose a vehicle theory in which experience is identified with
(or correlated with?) stable activation patterns in networks of neurons. Another example, much vaguer, is that of John Searle (1992), who appeals to unknown "causal powers" of brain tissue to explain experience.
Then there is the theory of Stuart Hammeroff (1994; Penrose 1994), which explains consciousness in terms of the quantum-mechanical behavior of brain cells, specifically their microtubules. I don't know how this approach fits into O'Brien and Opie's dichotomy, because the details are so fuzzy. But if the theory is a process theory, it's a theory of an unusual kind, because the processes in question cannot, by hypothesis, be explained mechanistically. Instead, the proposal seems to be that the ability of the mind to arrive at sudden insights is explained in the same way as the sudden collapse of wave functions in quantum mechanics. Grant that, and perhaps you may be willing to grant that phenomenal consciousness gets explained somehow, too.
The problem with all these theories besides their vagueness is that they are vulnerable to subversion by competing process theories (McDermott 1999). This is a flaw they share with dualism. As I explained above, any dualistic or "vehicular" explanation of the mind will have to accommodate all the facts that mundane process models explain, a set of facts that one can expect to grow rapidly. Every time we explain a mental ability using ordinary computational processes, we will have to redraw the boundary between what the vehicle theory accounts for and what the process theory accounts for. There are only two ways for this process to be arrested or reversed: either the vehicle theorists must explain more, or the process theorists must fail to explain very much. In the first case, the vehicle theory must compete with process theory on the process theory's home field, explaining how particular behaviors and competences can result from implementing a computation in one medium rather than another. That's hard to picture. In the second case, the vehicle theory will get a tie by default; no one will have explained consciousness, so the vehicle theory will have explained it as well as anyone.
If process theory doesn't fail, however, vehicle theory will be left in an odd position indeed. Suppose we have two entities E1 and E2 that behave intelligently, converse on many topics, and can tell you about their experiences. Their similarity is explained by the fact that they implement the same computational processes. However, the one implemented using vehicle Vgenuine is, according to the theory, conscious. The other, implemented using vehicle Vbogus, is only apparently conscious. When it talks of its experiences, it's actually making meaningless sounds, which fool only those unfamiliar with the theory. Unfortunately, no matter how elegant the theory is, it won't supply any actual evidence favoring one vehicle over the other. By hypothesis, everything observable about the two systems is explained by the computational process they both instantiate. For instance, if you wonder why E1 is sometimes unconscious of its surroundings when deeply involved in composing a tune, whatever explanation you arrive at will work just fine for E2, because they implement exactly the same computational processes, except that in the case of E2 you'll have to say, "It's 'apparently conscious' of its surroundings most of the time, except when its working on a new tune, when it's not even apparently conscious."
Vehicle theories are thus likely to be a dead end. This is not to say that the explanation of consciousness may not require new mechanisms. The point is, though, that if they are required they will still be mechanisms. That is, they will explain observable events. Phenomenal consciousness is not a secret mystery that is forever behind a veil. When I taste something sour, I purse my lips and complain. A theory must explain why things have tastes, but it must also explain why my lips move in those ways, and the two explanations had better be linked.
Many critics of computational theories reject the idea that phenomenal consciousness can be explained by explaining certain behaviors, such as lip pursing, or utterances such as "Whew, that's sour." But even those who believe that explaining such behavior is insufficient must surely grant that it is necessary. We can call this the Principle of the Necessity of Behavioral Explanation: No theory is satisfactory unless it can explain why someone having certain experiences behaves in certain characteristic ways. Naturally, process theories tend to explain behavior well and experience not so well, whereas vehicle theories tend to have the opposite problem.
Process theories tend to fall into two groups, called first-order and higher-order theories. The former are those in which in some contexts the processing of sensory information is "experience-like" in a way that allows us to say that in those contexts the processing is experience. Higherorder theories, on the other hand, say that to be an experience a piece of sensory information must itself be the object of perception. First-order theories include those of Kirk (1994) and Tye (1995).
Shapiro (2000) sketches what might be considered the "primordial" first-order theory:
Why not think of our perceptual experiences as sometimes entering a channel that makes us phenomenally aware of what they represent and other times bypassing this channel? When I'm driving around town thinking about a lecture I need to prepare, my perceptions of the scenery bypass the phenomenal awareness channel. When, on the other hand, I need to attend more closely to the world, to heighten my awareness of the world, I elect (perhaps unconsciously) to make my perceptions of the world conscious. Accordingly, my perceptions of the world get funneled through the phenomenal awareness channel. Phenomenal awareness, on this picture, requires no higher-order representational capacities.
He says this model is pure speculation, so I don't suppose he would defend it if pressed, but it makes a convenient target in that its obvious flaw is shared by all first-order theories. The flaw is that nothing is said about what makes events in one channel conscious and those in another channel unconscious. In Kirk's version, sensory data that are presented to the "main assessment processes" (1994, p. 146) are experienced, whereas other sensory data are not. Why not, exactly? Just as for dualist theories and vehicle theories, first-order computational theories suffer from a disconnection between the hypothesized process and its visible effects. The theories tend to treat "conscious" as a label stuck onto some signals in the brain, without any explanation of how this labeling causes people to be able to report on those signals (while being unable to report on those without the label).
Second-order theories do not have this flaw, because they say exactly what the labels consist in: computational events whose topic is the sensory processes being labeled. Here the labels are as concrete as data structures in a computer, so there is no difficulty following the chain of causality. The difficulty is convincing anyone that the resulting system has real experiences.
I will be pursuing a second-order theory in the rest of the book. Chapter 2 is a survey of the present state of research in artificial intelligence. Chapter 3 is a detailed explanation of my theory of consciousness. Chapter 4 deals with various objections, including those alluded to in the previous paragraph. The most serious objections are based on the observation that a computational theory of mind can't be correct because concepts such as "computer" and "symbol" are ill defined. Chapter 5 deals with this issue. Chapter 6 deals with various consequences of the theory, including the impact on religion and ethics.
One feature that will strike many readers is how little I appeal to neuroscience, unlike a great many recent theories of consciousness (Flanagen 1992; Crick and Koch 1998; Churchland 1986, 1995). One reason neuroscience is so popular is that it has produced some detailed, interesting proposals for how the brain might work. Many people won't be satisfied with an explanation of thinking and consciousness that doesn't ultimately appeal to the facts of neurophysiology.
But another reason is a prejudice that the basic case of phenomenal consciousness is the quivering bit of protoplasm in contact with the cruel world. The brain feels, in the end, because it is made of living, feeling parts. I have the opposite intuition: that feeling has nothing to do with being alive. The great majority of living things never feel anything. When evolution invented feeling, it stumbled onto a phenomenon that can be elicited from a living system, but not just from living systems. A theory of phenomenal consciousness must reflect this neutrality.
Another way to put it is this: Different organisms sense vastly different things. We rely primarily on visual inputs and so we receive completely different data about the world than a bat would at the same point in space and time. But gathering data and having experiences are two different things, and it may be that adding phenomenal consciousness to data processing is always a matter of adding the same simple twist to the system. If we ever find life on other planets, we will probably find that the data they gather from their environment, and the anatomical structures they use to process them, are specific to that environment and their needs; but if they're conscious it will be because of the same trick that our brains use. What they're conscious of will be different, but the way they're conscious may be the same. The same conclusions would apply to conscious robots.
This book is mainly about philosophical questions, but I confess that I do not always feel comfortable using the usual philosophical tools to approach them. Usually this is owing to differences in training and disposition. I hope those who would feel more at home in a discussion conducted in the pure philosophical style will nevertheless bear with me, in spite of my neglect of some of the problems and issues that philosophers focus on.
For example, philosophers spend a lot of time arguing about functional-ism. This term has several meanings. Some people treat it as synonymous with computationalism, the doctrine that the mind can be explained entirely in terms of computation. Since I'm defending a version of computa-tionalism, to that extent I'm defending functionalism, too. However, there are also other meanings assigned to the term, which reduces its utility. One version may be summarized thus: what mental terms and predicates refer to is whatever fills certain causal roles in a functional description of the organism involved. For example, a statement such as "Fred is in pain" is supposed to mean, "Fred is in state X, where X is a state with the properties pain is supposed to have in a worked-out functional description of Fred, e.g., the property of causing Fred to avoid what he believes causes X." Actually, stating the full meaning requires replacing "believe" with a functionally described Y, and so forth.
The purpose of this version of functionalism is to show that, in principle, mental terms can be defined so that they can be applied to systems without making any assumptions about what those systems are made of. If pain can be defined "functionally," then we won't be tempted to define it in terms of particular physical, chemical, or neurological states. So when we find an alien staggering from its crashed spaceship and hypothesize that it is in pain, the claim won't be refutable by observing that it is composed of silicon instead of carbon.
I am obviously in sympathy with the motivation behind this project. I agree with its proponents that the being staggering from the spaceship might be in pain in spite of being totally unlike earthling animals. The question is whether we gain anything by clarifying the definitions of terms. We have plenty of clearcut mental states to study, and can save the borderline cases for later. Suppose one had demanded of Van Loewenhook and his contemporaries that they provide a similar sort of definition for the concept of life and its subconcepts, such as respiration and reproduction. It would have been a complete waste of time, because what Van Loewenhook wanted to know, and what we are now figuring out, is how life works. We know there are borderline cases, such as viruses, but we don't care exactly where the border lies, because our understanding encompasses both sides. The only progress we have made in defining "life" is to realize that it doesn't need to be defined. Similarly, what we want to know about minds is how they work. My guess is that we will figure that out, and realize that mental terms are useful and meaningful, but impossible to define precisely.
In practice people adopt what Dennett (1978a) calls the "intentional stance" toward creatures that seem to think and feel. That is, they simply assume that cats, dogs, and babies have beliefs, desires, and feelings roughly similar to theirs as long as the assumption accounts for their behavior better than any other hypothesis can. If there ever are intelligent robots, people will no doubt adopt the intentional stance toward them, too, regardless of what philosophers or computer scientists say about the robots' true mental states. Unlike Dennett, I don't think the success of the intentional stance settles the matter. If a system seems to act intentionally, we have to explain why it seems that way using evidence besides the fact that a majority of people agree that it does. People are right when they suppose babies have mental states and are wrong when they suppose the stars do.
So I apologize for not spending more time on issues such as the structure of reductionism, the difference between epistemological and metaphysical necessity, and the varieties of supervenience. I am sure that much of what I say could be said (and has been said) more elegantly using those terms, but I lack the requisite skill and patience. My main use of the philosophical literature is the various ingenious thought experiments ("intuition pumps," in Dennett's phrase) that philosophers have used in arguments. These thought experiments tend to have vivid, intuitively compelling consequences; that's their whole purpose. In addition, the apparent consequences are often completely wrong. I believe in those cases it is easy to show that they are wrong without appeal to subtle distinctions; if those familiar with the philosophical intricacies are not satisfied, there are plenty of other sources where they can find the arguments refuted in the correct style. In particular, Daniel Dennett (1991), David Rosenthal (1986,1993), Thomas Metzinger (1995a), and William Lycan (1987, 1996) defend positions close to mine in philosophers' terms, though they each disagree with me on several points.
Several other nonphilosophers have proposed second-order theories of consciousness. Marvin Minsky is especially explicit about the role of self-models in consciousness (Minsky 1968). Douglas Hofstadter's proposals (Hofstadter 1979; Hofstadter and Dennett 1981) are less detailed, but in many ways more vivid and convincing. Michael Gazzaniga (1998) bases his ideas on neuroscience and psychology.
Adding another voice to this chorus may just confuse matters; but perhaps it may persuade a few more people, and perhaps even clarify the issues a bit.
Was this article helpful?