AI and Consciousness—Part One

Caveat

“AI” is not a single ‘thing’, despite being used, both colloquially and professionally, as if it was. The most egregious abusers of the term, “AI,” are the ones hyperbolically promoting specific products, e.g., ChatGPT, or specific techniques, e.g., Large Language Models and Deep Learning, as if they were the totality, the summation, of a family of techniques, tools, and algorithms.

AI has always been a collection of techniques and approaches that allow a computer-software artifact to exhibit outputs that resemble human behaviors. Technologies and algorithms that allow a computer, hooked up to a camera, to recognize a face (or a fingerprint). Others to translate French to Urdu.

Periodically, a particular approach or technology will gain center stage, become the face of AI, and be touted with intense hyperbole. In the 1980s-1990s it was Expert Systems. Today it is Large Language Models and Deep Learning.

Limning the Conversation

Most of the discussion will use the term ‘AI’ in the current fashion; i.e., as if Large Language Models, Deep Learning, and tools like ChatGPT “are” AI.

The core issue, ‘what is consciousness?’, long standing; primarily focused on human consciousness with, more recent, extension to other biological organisms. Assertions that computers are “intelligent” and are, or will soon be, “conscious” poses some interesting questions.

Exploring those questions cannot be done directly. AI is an application of Computer Science, both of which lack bodies of knowledge—e.g., computer psychology, computer philosophy, computer cognition— that might correspond to what we know, think we know, about human and animal consciousness.

Also important, explorations of consciousness are at the fringe of science, largely an unexplored territory.

“Along the philosophical fringes of science, we may find reasons to question basic conceptual structures and to grope for ways to refashion them. Old idioms are bound to fail us here, and only metaphor can begin to limn the new order.” W.V. O. Quine

Quine’s, use of the term metaphor, and ours as we proceed, is a bit colloquial. Strictly speaking a metaphor takes the form X is Y, and is different from analogy and simile. Metaphoric reasoning incorporates, rather willy-nilly, all three concepts.

MacCormac offers a kind of ‘lifecycle’ for metaphorical/analogical reasoning. Wanting to explore an unknown “X” we start with a known “Y” and say that “X is Y.” MacCormac says this is an epiphor.

Characteristics, details, of Y are noted and ‘X’ is examined for similar characteristics. MacCormac calls these “referents.”

If referents are found to correspond the epiphor evolves to become a “lexical term” a mere statement of fact.

If they do not, the metaphor is dead.

When Bohr proposed that an atom was a tiny solar system, with a nucleus (sun) orbited by electrons (planets) it proved to be a useful tool to explore the then unknown, and unseen, atom. As more and more was learned about atoms and elementary particles the referents failed to correspond and the metaphor died.

[Except, as a simple model to teach some aspects of basic chemistry, comparison, the analogy, still proves useful—a kind of “as if” analogy.]

To explore whether or not what we know about AI and the, as yet, possibility of machine consciousness, we must begin with metaphor.

The Disastrous Computational Metaphor

The computational metaphor is actually a cluster of bi-directional metaphors”

A computer is a brain is a computer.

An executing program on hardware is thinking is an executing program on meatware.

Machine thinking is algorithmic is human thinking.

An AI is Intelligent.

An AI is a mind.

The origins of this cluster probably begins with Descartes who believed that “thinking” was simply the manipulation of precisely defined tokens according to fixed formal rules.

Descartes built a mechanical device that used formal tokens manipulated by formal rules; therefore, the device was a “thinking machine.”

Leibniz, building on the work of the 13^th century Catholic mystic, The Blessed Ramon Lull, established computational science and also built a mechanical “thinking machine.” So did Blaise Pascal.

These philosophers defined and constrained human intelligence in strictly logical, analytical, and formal terms. They dismissed as irrelevant any other kind of thought—except, maybe, piety and faith. It is not surprising that Alan Newell and Herbert Simon, often cited as creators of the first AI, The General Problem Solver, also defined intelligence strictly in terms of what a university professor believed themselves to be doing when thinking.

Convinced that human thinking was a solved problem and noting that Boolean Logic and Leibniz’s binary arithmetic was literally embedded in the circuitry of digital computers, attention shifted to human thinking.

Specifically: could the equivalent of computer hardware, circuits and switches, be found in the human brain. Was it reasonable to equate hardware and “wetware?”

Initially, the answer seemed affirmative. Research had revealed that the human brain was, at least in part, a vast network of axons, dendrites (circuitry), and neurons (switches). Neurons appeared to act as switches because they could either fire (binary 1) or not (binary 0).

Superficially, this reasoning seemed to confirm the metaphor. However, only if two unspoken assumptions were made: first was that the complex chemistry that made the ‘wetware’ wet was irrelevant; second, that the network of nerves throughout the body were merely “sensors” or “effectors” playing no role in the “thinking” taking place in the brain.

The human brain is equivalent to the computer if, and only if, significant details (which may or may not be essential) are replaced with a simple model of current passing through circuits containing binary switches. It is this simplifying assumption that allows for construction of ‘neural nets’ in chips.

The next question concerns “programming.” Computers can [must] be programmed, could humans?” An answer of “yes” was vigorously asserted—for a thankfully brief period of time. Humans could be “programmed” and, if for example they were the victim of a cult, “deprogrammed.”

Related, but more interesting, it was noted that computers—at least their virtual implementation—could be ‘rewired’. Could brains be?

A neural net model suggested yes. The key was the synapse firing or not. If the summed value of incoming current exceed a threshold value, a “weight,” the synapse would fire. Otherwise not.

That weight could be adjusted by providing feedback as to whether inputs were ‘correctly’ connected to outputs. This feedback could come from an external source (supervised “learning”) or internally (unsupervised “learning”). By adjusting weights, the neural net could be configured so as to reliably and consistently channel inputs to expected/desired outputs.

More sophisticated and complicated algorithms, often involving extensive recursion, yield “Deep Learning.”

The appellation of “learning” to this process is not totally wrong. Infant brains seem to be “rewiring,” creating new and stable virtual circuits, almost continuously. An activity that slows down, dramatically, with age.

Based on these few referents that, at least partially, seem analogous on both sides of the metaphor have established a conviction, among computer scientists and AI advocates, that an AI is a Mind, fact, not metaphor.

Any evidence that human intelligence, human cognition is not replicable by a computer is dismissed, along with the underlying ability, as irrelevant or something that could, in the very near future, be replicated in a machine.

It is necessary to recognize that some AI researchers are not concerned with replication of human brain function in silicon. They focus on behavior, not internals. This is the foundation for the Turing Test: if an informed communicant cannot distinguish between interacting with a human or a computer, there is no meaningful difference. If differences in behavior exist, tweak the computer to eliminate that difference. The question then becomes, can we do that for every behavior yet to be mimicked?

Anthropomorphization, coupled with sloppy use of metaphoric language, established the conviction—at least among AI advocates and computer scientists—that brains and computers are simply instances of “physical symbol systems” and machine intelligence was identically equal to human intelligence.

The question of machine consciousness, typically, does not arise. It was simply asserted that human consciousness was an epiphenomenon arising from operation of the brain and, therefore, machine consciousness would necessarily arise from operation of the computer.

Scale is, perhaps, the most commonly referenced variable. As soon as a computer had an equivalent number of “neurons,” or equivalent “memory” capacity, or sufficient processing speed, it would simply, “wake up.” Other approaches argue that ‘architecture’ or ‘training methods’ might be altered; but with the same effect—the machine becomes conscious.

Perhaps our quest to understand consciousness would be better served with a different set of metaphors, conceived from an alternate perspective?

Perspective

Alan Kay once said, “the right perspective is worth 80 IQ points.”

AI, the ‘computational metaphor’, and all of its variants are firmly grounded in a world view, a perspective, that could be called “scientific,” “rational,” “formalist,” or “left-brain.”

The roots of this perspective include prominent thinkers of the Age of Reason, including Descartes and Leibniz, plus the mechanistic theories of 19^th century physics.

The entire universe is a vast complicated, but deterministic, system; one that could be understood in its entirety (at least by Laplace’s Demon) and therefore is utterly predictable (quantum mechanics not withstanding), in detail.

Reductionism led to the conviction that everything in that universe was also a kind of machine. That included computers and human beings.

This perspective has been immensely successful; allowing for all of the scientific and technical marvels that make modern civilizations possible. (Also, all the ‘marvels’ that threaten those same civilizations.)

However, a burgeoning number of researchers, scholars, and philosophers, (past and present), including: McGilchrist (cognitive science), Bohm (physics) von Bertalanffy (general systems), Whitehead (philosophy), Maturana and Varela (biology), Capra (physics), Senge (business), Prigogine (physics), da Vinci (polymath), Geertz (anthropology), and Fullerton (economics); argue for an alternative worldview, an alternative perspective.

Living systems. Complexity rather than baroque complications. Biology as inspiration instead of clockworks and steam engines. Culture and human “exceptionalism.” Uncertainty, ambiguity, and constant change. Co-evolution.

These ideas, collectively and interactively, form the base for an alternative approach.

Following the lead of Clifford Geertz, a name familiar to this audience, I will label this alternative approach “hermeneutic.”

Hermeneutics enjoys as extensive a tradition as formalism: in philosophy, Dilthey, Gadamer, Heidegger, Husserl, and Merleau-Ponty; in psychology, Vygotsky, and Wundt; in anthropology, Boas, Mead, Geertz, Lakoff, and Turner; and even in computer science, Winograd and Dreyfus.

Murray Leaf documents how our definitions of, and theories concerning, human beings have alternated between the formalist and hermeneutic (or interpretivist) poles.

A central point of divergence between formalism and hermeneutics: the former emphasizes reductionistic decomposition into quasi-independent “parts” or subsystems; the latter emphasizes holism and inter-connectedness.

There is obvious overlap. Some formalists look to context (e.g., Tarski and Godel), some hermeneuticists emphasize “scientific” thought (e.g., Marvin Harris). But the pre-disposition remains; formalists focusing within the system and hermeneuticists from the system outward.

Consider “mind,” “consciousness,” or “intelligence.” A formalist like Descartes will look to the manipulation of symbols via formal rules, within a subsystem, like the brain.

A hermeneuticist, like Geertz, argues that thought and behavior are manifestations of “interpretations of reality by individuals and collections of individuals.” Symbols are central to his conceptions but the meaning of such symbols is public, negotiated, and contextually dependent.

“[a certain set of] dispositions of an organism” that do not arise from any innate (genetic) capacity of the physical organism but arise as a consequence of “tools, hunting, family organization, art, religion, and science.” These influences “mold man somatically and are, therefore, necessary not merely to his survival but to his existential realization.”

Recognizing that some kind of mechanism must exist whereby mind is realized, suggests:

“Although, conceivably, mere increase in numbers of neurons may itself prove able fully to account for the florescence of mental capacity in man, the fact that the large human brain and human culture emerged synchronically, not serially, indicates that the most recent developments in the evolution of nervous structure consist in the appearance of mechanisms which both permit the maintenance of more complex regnant fields and make the full determination of these fields in terms of intrinsic (innate) parameters increasingly possible. The human nervous system relies, inescapably, on the accessibility of public symbolic structures to build up its own autonomous, ongoing pattern of activity.”

Lakoff offers a complementary analysis: (emphasis mine)

“The [human] information processing system is a joint body-mind system, not factorable into purely mental and purely bodily functions … [consider] the difference between signal processing and symbol processing. Both are forms of information processing. But individual symbols are assumed to have meaning and individual signals are not. Information processing in the central nervous system involves signal processing, not symbol processing. A joint body-mind system might involve both signal and symbol processing., without a single, clearly isolatable, symbol processing subsystem. … within connectionist approaches, it may ultimately be possible to maintain a joint body-mind position [where] the body-mind information processing functions overlap and significantly determine many of what have traditionally been called purely mental functions.”

It is difficult to see how the AI model where “mind,” “intelligence,” and “consciousness” are nothing more than algorithms executing on hardware can serve us. It can be argued, however, that some kind of model is necessary to move from “mere’ speculative philosophy.