Review - Machine Consciousness
by Owen Holland (Editor)
Imprint Academic, 2003
Review by Catherine Legg, Ph.D.
Sep 1st 2004 (Volume 8, Issue 36)

Consciousness is a perennial subject of fascination to human inquiry, and whether it can be reproduced in a human artifact − the question of "Artificial Consciousness" − has recently distinguished itself from the question of "Artificial Intelligence" as a research area in its own right, enabling one to pose the question of whether "AC" will turn out to be a necessary precondition for AI (now arguably one of AI's most intriguing questions). This collection arises from a workshop entitled, "Can a Machine be Conscious?" which took place at the Banbury Center, Long Island, in 2001 (though some of the published essays were not presented at the workshop but solicited later). Methodologically, the collection ranges across a wide spectrum: from armchair discussions from a range of philosophical perspectives, to technical reports from actual robotics research projects. The result is a stimulating resource for advanced undergraduates or the interested layperson.

Four papers sit on the more purely theoretic or philosophical end of the spectrum. First of all, Stevan Harnad ("Can a Machine be Conscious? How?") re-argues the classic Cartesian line that consciousness is forever closed to mechanical reproduction because of the epistemic inaccessibility in principle of qualia, also known as "the other minds problem" ("…our forward- and reverse-engineering can only explain how it is that we can do things, not how it is that we can feel things. And that is why the ghost in the machine is destined to continue to haunt us even after all cognitive science's empirical work is done" p. 75).

A more original paper by Susan Blackmore argues that seeking to artificially create consciousness is actually somewhat of a red herring for robotics research, since, "ordinary human consciousness is an illusion". To be more specific, it is an illusion "created by memes for their own propagation" (p. 20). Obviously following Dawkins, she claims that we are the only species which supports memetic evolution. For this to happen we must be able to copy each others' memes, which requires us to believe in "other selves", (here Blackmore replaces the good old English word "self" with the neologism "memeplex"), but since this is the only function selves perform, they are entirely fictive. All this is not to say, she acknowledges, that it would be impossible to create a machine subject to the same illusions, but what would be the point? This is a fascinatingly subversive argument, though it's worth noting that, rather than drawing from it Blackmore's eliminativist conclusion regarding the memeplex, one might equally posit selves as themselves memes (and in the case of a powerful personality − Socrates and Buddha come to mind – even some of the most potentially enduring).

Jesse Prinz, in a paper entitled "Level-Headed Mysterianism and Artificial Experience", explicates his favored mysterianism as the view that we can give necessary conditions for consciousness and sufficient conditions, but never both. This however, surprisingly, does not mean we cannot formulate "good, concrete hypotheses about the material basis of consciousness". Prinz sketches his favoured "science of consciousness", which consists of a number of explanatory levels, building on the pioneering work of Marr (namely, "psychological profile", the "algorithmic level", then a number of levels of "neuronal implementation"). Such a multilayered story means, however, that "we can describe the key systems involved in consciousness at varying degrees of abstraction" (p. 120), and it is difficult to isolate exactly which levels matter for consciousness, and how, in the way required for necessary and sufficient conditions. He also allows that consciousness could alter without affecting behavior, so that although we can make predictions about which of the machines we build are conscious, we can never confirm them. (Here a certain assumption regarding Cartesian privacy of the mental strikes again.)

Finally, the editors show a nice eclecticism by including the provocatively entitled "Borg or Borges" by cultural critic William Irwin Thompson. This stimulating piece launches a poetic attack on the very idea of 'artificial intelligence' ("It is a paradox of the work of Artificial Intelligence that in order to grant consciousness to machines, the engineers first labour to subtract it from humans…", p. 187), and on the attempt to provide a technological solution to what Thompson claims are essentially spiritual questions ("Technologists are closer to paranoids than they are to mystics in the sense that they are literalists given to perceptions of misplaced concreteness; they always see spiritual experiences as the products of technology…Mystics flip this literalism over to see technology as a system of externalized metaphors…" p. 188).

On the engineering side, the collection contains, as mentioned, a number of technical reports on robotic consciousness projects, though, disappointingly, none seem to show many real results as yet. Firstly, Rodney Cotterill (Danish Technical University) reports on a project called CyberChild which "aims to search for the neural correlates of consciousness through computer simulation" (p. 29). This system simulates two senses (hearing and touch), and also typical baby-discomforts such as hunger and wetness. The hope is that by allowing the system to evolve, it will "ontogenetically acquire novel reflexes", though nothing like this seems to have happened so far, and Cotterill concludes with the disappointingly weak, "The project is still in its very early stages, and although no suggestion of consciousness has yet emerged, there appears to be no fundamental reason why consciousness could not ultimately develop and be observed."(p. 29).

The second real-world project described is IDA, a US Navy computer program designed to take over the task of scheduling sailors' work rosters from traditional 'detailers'. Detailing is a task requiring an interesting set of skills sophisticated both purely computationally and regarding the maintenance of human relationships. The IDA developers use global workspace theory to conceptualise consciousness, and there is an interesting discussion of this. Once again, though, the results seem rather minimal, evidenced, among other things, in a coy use of scare-quotes around the term 'conscious' whenever it is attributed to IDA. For instance, the authors write, "Though IDA does not, as yet, engage in non-routine problem-solving, work on adding that capability is in progress. She uses her 'consciousness' module to handle routine problems with novel content. All this together makes a strong case…for functional consciousness" p. 63).

Holland and Goodman, in their "Robots With Internal Models", take perhaps the most explicitly engineering approach to achieving machine consciousness. The key to consciousness, they claim, is a robot's ability to include itself in its model of the world. (This is not a new idea, of course.) The authors outline a plan to build a succession of robots that function in the world by building and exploiting internal models of ever-increasing sophistication. However once again it would have been good to see the researchers actually implement this plan, but they appear merely to have worked with a set of ('ARAVQ') simulations at CIT.

As well as the philosophical and the exclusively engineering approaches to machine consciousness, a number of papers attempt to bridge the two. Sloman and Chrisley (University of Birmingham), in "Virtual Machines and Consciousness", begin by arguing that the concept of consciousness is a cluster concept subject to a Babel of confused claims. It therefore needs a scientific precisification (as happened with 'warmth' when it became 'temperature', in this way their claim is a version of the pragmatism of Charles Peirce, though the authors appear unaware of this antecedent). The authors suggest (like Peirce) that such scientific precisification is best performed a posteriori, by designing and building virtual-machine architectures which capture various features of consciousness.

In a clear-eyed discussion with potential to throw new light on the highly worked-over 'supervenience' issue in analytic metaphysics, the authors claim that virtual machines, though they are based on or realized in physical mechanisms, are not necessarily describable in the language of the physical sciences (consider, for example, a chess-playing 'virtual machine'). At the same time, bridging laws between virtual machine and physical mechanisms are neither analytic (since they cannot be provided a priori) yet neither are they empirical (since they are possessed of a form of necessity once grasped). 'Virtual machine functionalism' (p. 148), they urge, allows for multiple coexisting, independently varying, causally interacting states and processes − as opposed to atomic state functionalism which allows an organism only one mental state at a time.

Drawing originally from mathematics, they note that the conscious / non-conscious distinction may be neither a dichotomy nor a continuum. A third formal possibility is a conceptual space with many discontinuities. They cash out this insight with the claim that just as architecture-space contains many niches, correspondingly many different varieties of mentality are possible, even sketching an 'architecture-schema' in which one might map out some of this space. Their overall aim is to show that many of our pre-theoretic concepts of mind can be reconstrued as architectures, rather than (as GOFAI would have it) as algorithms. This 'structural' interpretation of mentality, unlike that of many other papers in this volume, transcends simple Cartesian dualism. (Even qualia too are given architectural explication, and regarding the now well-worn comeback about the logical possibility of zombies, the authors bite the bullet – "it is not clear that anything intelligible is left over" p. 169.). This is an original and very interesting paper, which, if taken seriously, has the potential to change the methodology of much philosophy of mind, from seeking to find 'correct' conceptual analyses of inherently indeterminate folk mental concepts, to exploring and experimentally testing spaces of more determinate concepts discovered a posteriori.

Finally, Aleksander and Dunmall (Imperial College, London) set themselves the task of crafting a system of axioms to guide and structure any conceivable test of a machine for 'minimal' consciousness. Examples include: Depiction: an agent has perceptual states that depict parts of S (a sensorily-accessible world). Imagination: an agent has internal imaginational states that recall parts of S or fabricate S-like sensations.

A well-written introduction traces links between the papers in terms of shared themes and/or contrasting takes on the same questions. One opportunity missed in a book entitled "Machine Consciousness" is the lack of any systematic interrogation of the word "machine". The solution to the problem of machine consciousness has been assumed by all parties represented in this collection to lie with the concept of consciousness, which in this book is prodded, pushed and pulled in all directions – many illuminating.

© 2004 Catherine Legg

Catherine Legg, Ph.D., Department of Philosophy, University of Waikato, New Zealand