William A. Adams

Adjunct Associate Professor of Psychology, Chapman University College
bill.adams@bainbridge.net

The test of a systems definition of consciousness would be to build a machine from its plan, then decide if the machine is conscious. Turing famously addressed the difficulty of assessing machine consciousness a half century ago (Turing, 1950), with ideas still controversial today (Horn, 1998). To pass the Turing test one needs to fool judges, for example by using clever programming to have the machine produce apparently intelligent verbal behavior. A modern variant of the problem is that of the philosophical zombie (Moody, 1994; Chalmers, 1996, pp. 94-99), an imaginary creature that looks and behaves just like a person but lacks genuine consciousness.

A Turing test suggests that intelligent behavior implies intelligence (in this context, implies consciousness). If you can make a computer act so much like a person that judges are unable to tell the difference, "it is gratuitous," said Turing, to insist that the computer is not intelligent. The burden is on the objector to provide extra-behavioral criteria for consciousness outside her own head. Any third-person ("objective") definition of consciousness should accept a behavioral criterion because that affords a possibility of consensus. Our empirical tradition allows us to accept without question the veridicality of sensory observation (under normal conditions) and with supplementary logic, to reach agreement on the interpretation of publicly observable events. We have no compatible tradition that allow us to reach consensus about the descriptions or interpretations of first-person, private events, such as phenomenal states of consciousness. Thus the main attraction of a computer-based definition of consciousness is social-epistemological pragmatism. Secondarily, a computer is a medium of expression well suited to such definitions because it facilitates fairly precise definition of independently variable, causally interacting, functionally differentiated substates (Sloman, 1996).

Rey (1997) suggests a further criterion for an artificial consciousness. He proposed a set of rules for a thinking machine, the "Recursive Believer," which could pass a Turing test, based on an algorithm derived from an acknowledged theory of mind. Recursive Believer might spontaneously print out something like, "I think therefore I am." Despite such apparently intelligent behavior, Rey challenges, "But where in any of this is there any need of consciousness" (p. 469)? He concludes, "In all theoretically significant ways we seem to be indistinguishable from the ‘mere machines’ from which we nevertheless insist upon distinguishing ourselves" (p. 474). The Rey challenge says that if the intelligent behavior of the computer is based on processes that could in principle plausibly be ones used in natural intelligence, then by parsimony, no other criteria of intelligence or consciousness need exist. The burden is again on the objector to name additional properties of consciousness that would make any difference to the performance criterion. The zombie argument is a version of "the Rey test," since it imagines that the processes used to produce the intelligent behavior are literally identical to those used by a person. However it suffers from asking us to imagine non-conscious intelligent behavior, something not everyone can do, whereas a computer implementation is unmercifully empirical.

Any computer-based definition of consciousness, even one based on a principled algorithm, is like a null hypothesis. It proposes: "There is no difference between natural and artificial consciousness. Prove otherwise." Though most people are biased against believing that a machine can be conscious, the best one can do is disconfirm, or reject, the null hypothesis with some compelling argument or counterexample, leaving no definition of consciousness standing. One could then propose an alternate definition, but it would either be another reductive null hypothesis or an interminably arguable first-person account.

An alternative is to present a principled algorithm for consciousness and ask the straightforward question, "Could this functionality constitute a genuine consciousness of the kind we recognize in ourselves?" There is nobody and no thing hidden behind any curtain. The goal is not to fool anyone. The machine implementation helps assure that the terms of the definition do not invoke unwitting anthropomorphism, but the purpose of the exercise is not to make a machine conscious in some recondite sense, nor to produce a narrow computer performance indistinguishable from a person’s. The point is simply to demonstrate a completely explicit, operationally verifiable third-person definition of consciousness.

Even with a plausible machine definition however, few people will consider the possibility that any machine could be conscious. As Rey (1997) says, "…perhaps one ought after all to regard a computer programmed [as described] as conscious. However…I must confess that I find myself unnerved, and I find most other people unnerved, by the possibility… It simply seems impossible to take their [the machines’] mental life all that seriously: to feel morally obliged (not) to treat them in certain ways (not to unplug them, not to frustrate their preferences, not to cause them pain)" (p.472). That is why most computer-based definitions of consciousness are null hypotheses: nobody can take them seriously in their own right. They are not really definitions, but challenges to any definition.

Yet anyone who rejects on a priori grounds the possibility of genuine artificial consciousness is not taking seriously the definitional problem. It’s not about the machine. The algorithm is the definition, and that’s what matters. Anything that can be defined clearly enough can be implemented in numerous ways. A computer implementation is simply the translation of an adequate definition from one medium to another. Someone who says that no machine could be conscious is either asserting an indefensible biological chauvinism, or they are really saying that there can be no algorithm for consciousness, that in principle, consciousness can have no clear definition. That is an arbitrary, discussion-ending, perhaps mystical, and certainly not very helpful point of view.

We are like a racehorse spooked by the finish line: We want to assume that consciousness is a natural and analyzable system, we believe that any sufficiently defined system is implementable in a computer medium, but we won’t believe that a computer could ever be conscious. This impasse arises from a misconception of what an artificial consciousness is supposed to be. Most people assume that a general-purpose computer (i.e., all of its hardware, software, input and output facilities, etc.) programmed for intelligent behavior makes up an encapsulated entity analogous to a person. This notion is appealing when we assume that the mind is also an encapsulated entity, located within the person. The computer’s program is supposed to be the functional equivalent of natural consciousness. But the analogy is seriously misleading.

It is a truism among many philosophers that no "formal" (closed) system like a computer program can in principle be conscious, for the same reason that makes us say a dictionary is not a language user. The computer does not understand its own input, output, or internal processing, but we understand ours (perhaps not perfectly, but in large part). We "mean" what we say and do, but a computer does not. All knowledge of consciousness arises from first-person awareness of one’s own conscious experience (Chalmers, 1996), so in a third-person description of computer consciousness, we require a homunculus to be present, an image of the self-aware function we experience ourselves. That’s an essential feature of the common sense definition of consciousness. But we know there isn’t a homunculus in the computer, so the computer cannot be conscious.

Are we wrong to require a homunculus in a conscious other? Maybe so, but that’s what we do. We require a homunculus of each other too, and the requirement is satisfied by all performance criteria. All evidence supports it; no evidence refutes it. Even if a computer could meet those same performance criteria however, we know, because it is our artifact, that there is no homunculus in there (something we cannot be sure about with a philosophical zombie). So no matter how clever or convincing a computer’s intelligent behavior might be, common sense can never allow that it is really conscious.

There have been attempts to overcome that bias. Rey’s Recursive Believer algorithm, for example, represents its own representation of its input data. If input is the "raw experience" to the system, the program would not only have this experience but be able to report on it, too. Another level of indirection and a matrix of valuation weights would even let it say what it believed and preferred about its raw experience. Doesn’t that count as self-awareness? Rey is sure that it does, and suggests that our reluctance to admit it is an unjustifiable arbitrariness (1997, p. 474). But maybe common sense knows better. Seemingly self-aware reporting is not good enough. The homunculus we require in the other is one that can actually be verified intersubjectively.

Even for people who have never met each other in person, an intersubjective relationship can be established and maintained through appropriate media. We take a poem to be a genuine expression of the poet’s consciousness, for example, encoded to the poetic form for the purpose of communication, addressed to people who are known to the poet only in abstraction. As the reader, we take the poem not just as a display of orthography, but as somebody’s attempt to express their experience to another person. It doesn’t matter if the intersubjective medium persists longer than the person who last used it. Though Shakespeare is long dead, we can still make some contact with his consciousness through his work. The Sonnets speak to us after 350 years because the persistence of the medium allows us to understand him, have a bit of an intersubjective relationship with him still.

If two people are widely separated in time, intersubjectivity borne on persistent media between them can be called "diachronic" intersubjectivity, emphasizing that the interaction between the people is discontinuous in time. (Diachronic does not mean entirely without regard to temporal sequence. A poem still must be written before it can be read.) We are fascinated by those cave paintings at Lascaux because they let us make some intersubjective contact, however slight, with our fellow people of so many thousands of years ago. Diachronic intersubjectivity is real intersubjectivity. It just happens to involve subjects whose communicative tokens are sent and received at widely separate times, and usually in a non-interactive (one-way) direction.

A computer program is the creative artifact of a genuine human consciousness, just as the characters in a novel are the product of some author’s creative consciousness. Unlike the novelist, the programmer defines changes to the product as if it were still a work under active development. This would, in a traditional medium, give a heightened sense of the author’s presence. But the changes in the behavior of the computer system are stored for execution at a time when the author has departed. Furthermore, those changes are not obviously systematic and are actually contingent on the transactional activity of the medium. They are defined so that they depend, for example, on which keys on a keyboard a human user of the medium presses. It is difficult for someone who is not a computer expert to discern the connection between the programmer’s social intentionality and the machine’s behavior. Consequently a the person might suppose that synchronous intersubjectivity is offered by the machine.

A well-known example of misattribution of the intersubjective source was demonstrated by Weizenbaum’s ELIZA program, a computer parody of the verbal behavior of a nondirective psychotherapist in an initial psychiatric interview. If the person said, "My mommy took my teddy bear away from me," ELIZA would say, "Tell me more about your parents." Weizenbaum (1976) found that people reported intersubjective experience with ELIZA.

Weizenbaum said,

Since Weizenbaum wrote the ELIZA program, he was sure it did not have any consciousness or subjectivity. Therefore his students who experienced intersubjectivity with ELIZA must have been wrong about their experience. But an alternative explanation is that Weizenbaum’s students mistook diachronic intersubjectivity with Weizenbaum for synchronous intersubjectivity with ELIZA. The program focused the dialog on family relationships and feelings. Everyone has parents; everyone has feelings. The dialog easily gave the illusion of a synchronous intersubjectivity. The students were confused about who they were talking to, and so was Weizenbaum. In fact Weizenbaum was the homunculus in the system.

The same confusion inhabits Searle's (1980) famous thought experiment of the Chinese Room. Searle had in a sealed room, a lookup table that allowed him to transform meaningless (to him) input into meaningless (to him) output. Observers outside the room believed on the basis of the input and the output that the room understood Chinese. Since Searle understood no Chinese, the error of the observers’ judgment was made apparent. But Searle was not the homunculus in the system. The real homunculus was the author of the lookup table, who did understand Chinese. The observers mistook diachronic interaction with that author for synchronous interaction with the room itself. Searle’s presence in the room was irrelevant except as comic emphasis for the misunderstanding.

This common error, taking diachronic for synchronous intersubjectivity, is a matter of lack of general familiarity with the computer medium. It is a relatively new medium and a complex one, difficult to understand. People were once astonished by voices coming out of a radio box and images moving on a silver screen, but in time, came to an appreciation of the relationship between the medium and the communication it carries. When that happens with the medium of the interactive computer, the challenges of the Turing test and Rey test will lose their force. People will readily discern a human as the homunculus behind the expressions of the "intelligent" program.

A person is an excellent "consciousness detector," the best there is; the only there is. The direct apprehension of another person’s consciousness in a conversation, for example, is a criterion of consciousness that is universally used. All of civilization depends on it. The possibility of an erroneous judgment of consciousness in another person under normal circumstances is so remote it is not even germane. If one’s interlocutor is an undisguised computer, one is more cautious. Still, as numerous ELIZA-like programs have repeatedly shown, people can detect consciousness while using a computer.

The computer user may detect the presence of a genuine consciousness through her interactions with the machine, but may naively "locate" that felt intersubjectivity in the computer, an error. Nevertheless the user could still honestly and accurately report having felt the presence of another consciousness. It would actually be the consciousness of an unknown "other," an abstracted but real person, like whoever painted the horses on the wall at Lascaux or like an anonymous correspondent on the internet. It would be someone associated with the scientist in whose lab the program is encountered, someone appropriately socialized, conversationally competent, a conscious agent "like me."

Where is the consciousness of an alleged artificial consciousness supposed to be located? It is not in the computer, which lacks a homunculus. And it is not necessarily located in the programmer, who might be dead by now. The consciousness detected by the user is actually "located" in the wider social context of the transaction between the computer and its user. The user is communicating with a presumed member of her social community. The user’s social context was arranged in such a way that this particular interaction is appropriately embedded. The computer user did not trip over an artificial consciousness computer on the way to work. The dialog did not unexpectedly emerge from the user’s word processor. Those possibilities do not make sense. The artificial consciousness must be properly introduced to the user.

If some stranger walks up to you on the street and says, "I have three pencils," you might say, "How very nice," but you would quickly move on, perhaps glancing for a police officer in case there was going to be trouble. You can’t have much of an intersubjective relationship without an appropriate social context. If a stranger on the street says "Give me your money or I’ll hurt you," at least you have a social context for that relationship. If the stranger standing beside you at a train station says "Lovely weather," the public weather story establishes the flimsiest of social contexts in which some kind of mutual acknowledgment can proceed. Intersubjectivity requires the engagement of two consciousnesses, and engagement in turn requires a social context, at least parts of which are mutually understood. Since intersubjectivity is the only way there is to detect consciousness other than our own, it follows that any other consciousness must be embedded in one’s social context or it cannot be recognized. The idea of a free-standing consciousness outside of a social system is therefore paradoxical.

In a typical artificial consciousness situation, the social context is a laboratory, and an introduction is made by the scientist, who says, perhaps not in so many words, but in effect, by the demand characteristics of the situation, "I’d like you to meet my friend, ELIZA and have a conversation with her, as a favor to me." The intersubjective context between the user and the computer is borrowed from the natural relationship between the two persons. Intersubjectivity is always embedded in a social context which it presupposes, whether the participants are aware of it or not (and usually they are not).

Consciousness outside one’s own head can only be positively verified by intersubjectivity but it doesn’t have to be an explicitly conceptualized intersubjective awareness. The quality of the other’s consciousness becomes apparent from the content of the communication, while the stark fact that one is in intersubjective contact with another consciousness remains implicit. One might determine that the other is an adult or a child, intelligent or stupid, logical or emotional, knowledgeable or just chatty, and so on. But whatever impressions one might form about the other, they all depend on the presumption that the other is a conscious agent "like me" and that means a member of a community I recognize. The intersubjective relationship itself is invisible, completely transparent, unless its presumption is violated.

Sometimes we are explicitly aware of the presence of the other’s consciousness per se. In an intersubjective relationship of longstanding between two people, a near-imperceptible gesture can be a meaningful indicator of the other person’s consciousness. Even the absence of any reaction at all can, in the right circumstance, be sharp evidence that "someone is home." Most often though, intersubjectivity is implicit, transparent. The individuals are typically interested in conceptualizing the contents and qualities of the other’s experience, not the fact of the other’s conscious existence. For a person and an artificial consciousness system that have only been recently introduced, it is useful to begin with a presumptive intersubjectivity, focusing on a public (mutually appreciated) story, such as "How ‘bout them Yankees," or, as ELIZA was set to ask, "Tell me about your family." Any artificial consciousness program that purports to prove theorems, play chess, or even perform a psychiatric screening interview, must provide and not disrupt, intersubjective transparency between the user and some abstract agent in her social community, whether the user mistakenly locates that agent in the computer medium or not. Without that intersubjective transparency, the user will have no sense of being in the presence of a conscious other, and therefore the adequacy of the specific intelligent function or artificial consciousness module under investigation cannot be confirmed.

An artificial consciousness is similar. The engineered subsystem interacts with its natural contextual system, even though we don’t know exactly how that context works. The artificial consciousness program is really only part of a larger system of social interaction. The programmer isolates certain aspects of ordinary consciousness and replaces them with devices. Therefore it is more descriptive to call the computer-based system a prosthetic consciousness. The test of a prosthetic consciousness is the same as the test for an artificial heart: does it work?

Suppose we consider Rey’s Recursive Believer as a prosthetic consciousness. Its claim to intelligence was that it could report on its sensory environment (e.g., as input by an attached television camera), and even print out a preference about that (e.g., "I prefer to see a pyramid on top of a cube.") As a prosthetic consciousness the Recursive Believer would be an implant that replaces whatever natural functions of consciousness guide the natural performance of stacking three-dimensional geometric objects and reporting on it. A second prosthetic subsystem for Rey was identification of geometric objects from the structure of reflected light. It would have been simpler to stick to one prosthesis at a time, dispense with the television camera altogether, and just program the features of the geometric objects into the artificial stacking system. But another critical component of Rey’s system is almost entirely lacking: intersubjective transparency.

No facility was described in Rey’s algorithm for a person to ask it, "Why do you prefer the pyramid on top?" And there was no way for the system to answer, as in, "Because my preprogrammed table of geometric combinations dictates that pointy features go on top." In fact if there were such interactive facility, we might decide that the system was not so intelligent after all. Rey’s system doesn’t engage well socially, so we are left with no evidence but the programmer’s assertion that we are in the presence of any kind of consciousness. Whether Rey’s object-stacking and reporting algorithm is a plausible description of one aspect of consciousness or not, without intersubjective transparency it is unconfirmable. In order to validate a prosthetic consciousness, we must see it operate it in the context of natural social processes.

What kind of social process would it take to validate Rey’s proposed prosthetic consciousness? A technically simple kind of social interaction that supports intersubjectivity through a computer medium is dialog. An internet conversation is an example. If one of the interlocutors were a prosthetic consciousness like Recursive Believer, it would have to support such a dialog. Unfortunately, the essential nature of dialog is an unsolved mystery, so a simplified approximation might suffice, since dialog itself is not the prosthetic module being tested. Recursive Believer had only very crude language capacity for reporting (e.g. making printouts about its own internal states).

But even with a sophisticated natural language interface, what should Recursive Believer say in order to support intersubjective transparency? Preferring pyramids on top is a good start. The person would understand, "I also have preferences. I know what it’s like to prefer something." But does Recursive Believer’s preference arise from triumphant victory or crushing defeat in fierce object-stacking contests, for example? Or is it purely an aesthetic preference based on classical artistic sensibility? Perhaps it is really not so much a preference as an appreciation of the relationship between kinetic and potential energy with respect to the force of gravity. That is the sort of thing the person needs to know. To recognize the intersubjective other as "a consciousness like me," the other must present rich evidence that it is a homunculus, has first-person experience. Mere use of personal pronouns is not enough¹.

The first task for a programmer of any prosthetic consciousness module then is to figure out how to make intersubjective contact with the person, by projecting that sense of being a homunculus. One aspect would be to report one’s beliefs and thoughts about features of the world, as Recursive Believer could. An intersubjective agent also has beliefs, thoughts and feelings about the social other, so Recursive Believer would need a representation of the human user as well. But then what? What do we "do" with each other in a conversation, beyond a few tricks like using the other’s name, to establish and maintain implicit, transparent, presumptive intersubjectivity?

The designer of a system that would support a diachronic intersubjective conversation, must isolate and encode the essential patterns, or structures of intersubjectivity in that context, since the particular linguistic form it will take in the future conversation is unknown. Objectification of a first-person homunculus is an unintelligible concept, but that’s exactly what the user of the system must detect to feel intersubjective. Fortunately, the homunculus does not have to be designed into the system. The designer of the system is the homunculus, and by isolating and encoding the regular patterns of intersubjective conversation, the designer simply makes herself diachronically visible to the user.

ELIZA tried to make intersubjective transparency identical with prosthetic consciousness. "Tell me about your family," references the mutually appreciated fact that humans have families and that families are important, especially in a psychiatric context. ELIZA’s designer (unwittingly) made himself visible to the user. But if the prosthetic consciousness was about conceptualizations of family, ELIZA showed no support for that module. Connecting intersubjective transparency to the prosthetic consciousness is the challenge. On the plus side, an algorithm for intersubjectivity does not require extensive knowledge of the world, no extraordinary rational, computational, or reasoning skills, only rudimentary language or other communicative capacity, no complicated "sensory" input and interpretation systems, no sophisticated or deceptive output. All that should make it easy to design and build. The problem is narrow. On the other hand, having an artificial consciousness connect awareness of the other with awareness of itself is a problem that neither Recursive Believer nor ELIZA addresses.

It might seem that what started out as a mere interface issue, providing intersubjective transparency in support of the prosthetic module of interest, now turns out to be a serious problem in its own right. Until technical support for intersubjectivity can be designed and implemented, there is no hope of validating other modules of artificial consciousness, such as units that handle knowledge of geometric objects, play chess, conduct psychiatric screening interviews, prove theorems, and so forth. Any of these might require extensive specific domain knowledge, a general knowledge-base of common sense facts, special capacity for handling probabilities, extended language capacity, ability to analyze pictorial or audiographic data, and a myriad of other functions designed as plausible representations of natural consciousness. But without the key verification tool, intersubjectivity, efforts to build those consciousness prostheses are futile.

There are three parts to a straightforward, no-tricks definition of prosthetic consciousness. First, incorporating the best features of the Rey test, it is principled. It proposes, "Here are some functions that plausibly could model functions of natural consciousness." By contrast, ELIZA was never designed as an actual definition of consciousness. It uses a few obvious tricks, such as parroting key phrases and adjusting pronouns, in an attempt to fool people². Consequently, while ELIZA-like programs still pass the Turing test these days, even with adults (Schieber, 1994), they do not pass the Rey test and they do not qualify as plausible (no-tricks) definitions of consciousness.

Second, a no-tricks definition is validated by implanting it into a diachronic communication. A person has an ordinary conversation with an undisguised computer, centering around topics germane to the prosthesis, then reports whether consciousness was intersubjectively detected. If so, that verifies intersubjective transparency through the computer medium to a presumed abstract social other who has (is) a homunculus, meeting the common sense requirement for any consciousness beyond one’s own. Finally, the targeted prosthetic module (for object-stacking, chess playing, etc.) is evaluated in its own right by any appropriate criteria that do not contradict or destroy the user’s sense of intersubjectivity.

The effort to specify a no-tricks, computer-based definition of consciousness suggests some implications in consideration of natural and artificial consciousness. One is that human consciousness should be considered as embedded in a social context. Any objective definition of consciousness requires acknowledgment of intersubjectivity as a practical matter, for purposes of empirical verification. In principle this is true for either artificial or natural consciousness. That implication puts constraints on analysis of natural consciousness.

Another implication is that computer-based definitions of parts of consciousness can legitimately be implemented and tested. Abandoning the unrealistic goal of an autonomous, isolated consciousness in a box, it should be nevertheless possible to implant prosthetic consciousness modules into the context of human intersubjectivity, and to test those modules pragmatically. That conclusion clears a methodological pathway for empirical progress on the question of consciousness.

In order to follow that empirical path however, it is necessary first to develop an explicit algorithm for intersubjective transparency by which a person can see through any prosthetic consciousness and its computer medium to the natural consciousness of abstracted social others in her community. Without that transparency, no consciousness can be verified. This requirement suggests that intersubjectivity is a kind of meta-consciousness, a normally invisible enabler that supports conscious cognitive processes.