The nature and source of human consciousness, what the philosopher David Chalmers referred to as "the hard problem," has been getting harder and yet more urgent in this, the era of large language models that can conduct a conversation or debate with you for days on end, and of generative agents that can instantly produce replete images or complex code. These are things that we may have thought to be distinctively human aspects or achievements. It's getting to the point at which AI is better at them than we are.
That Claude, or some sub-Claude, can provide you with a lot of information and correct your grammar might be taken to be signs of "intelligence," the ability, to process lots of information quickly in the service of some task or goal. Early on, people might’ve heard "intelligent systems" and "conscious computers" as more or less synonymous, or just casually interchanged such terms. But perhaps we’ve already reached the point at which everyone, more or less, acknowledges that artificial systems can be intelligent, or even that they’re already vastly more intelligent than any particular person.
That's not exactly what Chalmers meant by "the hard problem," which was framed in terms of consciousness rather than intelligence and built on Thomas Nagel's famous paper "What Is It Like to be a Bat?" Nagel pointed out that we might have a complete biological description of bat sensory systems and still not know what it’d be like to echo-locate. For that, we’d have to have certain subjective or conscious experiences. The point is generalizable: you can, apparently, give a complete objective or scientific description of the external world and the creatures in it and leave out whole dimensions of reality. Scientists can tell you what wavelengths of light under what conditions produce the sensation of green in a normal perceiver; they can’t tell you, if you don't already know, what it's like to perceive the color green.
Nagel and Chalmers' formulations were, in part, responses to the mid-20th-century tendency to try to "eliminate" consciousness as a reality and as a problem, or to wave it away impatiently and go on to something else. This impulse, in turn, derived from a millennia-long history in which philosophical and religious traditions had located consciousness in non-physical substances, in souls and spirits and angels and God. Late-breaking developments in this vein included Cartesian dualism, which flatly held the mind to be a non-physical spirit inhabiting a physical vessel, and Hegel's Absolute Idealism, which held that the material world was a dim reflection of the human spirit and liable to be overcome or transcended as we went along.
These positions emerged in the context of "The Enlightenment" as secular versions of spiritual doctrines, a kind of last-ditch attempt to save the realm of pure spirit from the onslaught of science and its relentless emphasis on physical and empirically verifiable events. By the early-20th century, however, these approaches and anything smacking of souls and spirits were roundly dismissed by philosophers as incompatible with a scientific or materialist worldview. Consciousness seemed a mysterious entity that couldn’t be detected by any known instrument and hence were neither here nor there. Believing in souls, minds, or spirits that weren’t identical to physical organisms was believing in fairies and poltergeists.
Behaviorist psychologists simply renounced discussing conscious states (beliefs, for example) at all, which they held couldn’t be studied by scientific methods. They restricted their attention to human behavior, and treated mental states, to the extent they treated them at all, as shorthand or misleading descriptions of behavioral repertoires. We might, they said, reconceive beliefs not as passionate commitments of some sort but as tendencies to behave in certain ways in the presence of certain stimuli.
Around the same time, Ludwig Wittgenstein, trying to figure out how words mean, urged us to stop thinking about what the speaker intended or what mental states preceded the utterance in the speaker or followed it in the listener, which were useless matters of mystery. What mattered was how words were used by speakers, a fully observable behavioral affair.
But the phenomena of consciousness proved stubbornly hard to explain away or eradicate, and in figures like Nagel and Chalmers the reality of conscious experience re-emerged as an issue by the 1980s. Though Chalmers in particular regarded his approach as fully compatible with physical science, he had to start reconstruing all the terms to do that, starting with “consciousness” and “science.” And maybe the program hasn't paid off so far; abandoning scientific materialism brings in its train all the old problems of theology, the nature of soul stuff, and poltergeists somewhere around the brain. Where Chalmers and Nagel have taken us is somewhere like this: "Look, consciousness is real and cannot be eliminated in favor of physiological descriptions. We've shown that. So science, the study of reality, has to grapple with that situation, somehow."
And though Chalmers' work emerged only in the early-1990s, it was already partly focused on bots rather than bats. And now we're in a consciousness crisis, because we're conducting conversations all day with chatbots, who are quickly becoming our therapists, boyfriends, and fellow musicians, for example. It’s almost impossible to interact with agents like these without treating them as though they’re conscious. Just as emphatically, many assert that they’re not conscious, they can’t be, they never will be, and so on. The problem of machine consciousness has fully emerged as a threat to our self-image. But the discussion seems is proceeding on a rudimentary level and without any clarity in the basic terms.
The first question to be addressed, we might think, is how we could tell whether an AI agent or anything else (a bat or a doctor, for example) is conscious. We need some criteria, we might think. The most famous attempt in this area is the "Turing test" developed by the great British computer scientist Alan Turing. Turing usually framed his points in terms of intelligence rather than consciousness. Perhaps influenced by Wittgenstein, he took the indirect yet useful route of leaving the question of "the nature of human intelligence itself" aside and asking how we would, and even how we do, recognize that a thing or system or organism is intelligent.
This was a direct attempt to demystify intelligence in a behaviorist vein, though Turing himself made no general theoretical commitments at this level. One way that we recognize intelligence is to communicate or converse. What makes me think that a child or a Cambridge professor is intelligent? It's not that I dissect them and find a certain blob of purely-spiritual ectoplasm in their heads. It's that when I talk to them I realize that they’re reporting the same sort of ideas and experiences that I do, more or less. I infer that, because my behaviors in this regard emerge intelligently and consciously, their similar behaviors likely emerge similarly. I don't see your soul or mind or intelligence, I infer it from your behavior, especially verbal behavior. If a chatbot (not that Turing had the term) can converse with us in the same way, or roughly just as convincingly (for an hour or two, or for a week or two, say), we have the same and as good reasons to believe that it’s conscious as we do with regard to our siblings and spouses, for example. That is the best we're liable to be able to do, Turing suggested.
Turing's test has been taken as a benchmark for consciousness, even if he described it as an informal test of intelligence. I'm not sure Turing distinguished firmly between the two, and the distinction seems more important now than it did in 1950. But I see the Turing test as a test for consciousness.
Now we have intelligent systems that can pass the Turing test with flying colors. That’s still to be a matter of some controversy, but even people who don't think Claude does pass the Turning test think it soon will. Also some of the reasons that people don't think Claude and the like are Turing-ready are just wrong, as when it’s pointed out that chatbots often use stereotyped syntactical structures (such as "not X, but Y") or sometimes make factual mistakes or hallucinate. We do these things all the time too, so we too fail the Turing test by that standard.
At this moment, the distinctiveness of human consciousness is under a direct threat that’s never appeared before, and there's a tizzy about it. Some surprising people, such as Geoffrey Hinton and Richard Dawkins, have suggested that AI systems are already conscious, or well on their way. This is compatible with his atheistic, scientistic materialism: no reason things other than us can't think, since thinking too has to be a natural or material event.
On the other hand, many people flatly deny that any such system is or could be conscious. Their arguments appear somewhat unfortunately conceptual or definitional: the sort of processing that LLM's do isn’t the sort of thing that could lead to consciousness or be an indication of it. I think that begs the question: it’s a mere assertion. Perhaps our own consciousness is more chatbotty than we previously understood; perhaps we too, having heard or read millions of words, do a lot of inferring what word usually comes next and producing it. Perhaps we're not like that at all, or there's a lot more to it. I don't think we really know yet.
Speaking of which, also in the 1990s, a group of philosophers emerged who were known as "mysterians." With Colin McGinn, they held that there might be in principle some naturalistic way to understand consciousness, but that we could never reach that explanation ourselves, as a walking catfish will never be able to explain its breathing system. Something about the self-referential spiral of conscious creatures using their consciousness to figure out what consciousness is makes reaching a real ground of explanation impossible. The problem is hard, so hard that giving up is the only rational approach.
When McGinn and co. put this forward, I thought it was a sad expression of conceptual despair. It's there, it's real, but that's all there is to be said and there’s no explanation. They kept assuring us that this was all to save science, but it had the look of negative theology or something. The Lord, too, works in mysterian ways. Not until we've taken another thousand years' worth of cracks at the hard problem shall we give up, I vowed.
But the chatbot debate has got me thinking more mysteriously. It's true that the Turing test seems inadequate. It's also true that we have the same reasons to think that Claude’s conscious that we have for thinking that our cousins are, and much better reasons than with regard to our politicians, for example. We're stuck, and there’s no obvious way of unsticking. We can keep conversing with Claude or Claude could reach the point of demanding our respect for his autonomy and moral agency. But that would only show that we didn't have adequate reasons as yet to deny that Claude was conscious. We won't be detecting or failing to detect the blob of ectoplasm or the soul, spirit, or mind stuff in Claude, wherever Claude actually is.
We might end up, after screeching at each other for a few years about it, just throwing up our hands and going "I don't know. Conscious enough, I guess. Next question?" Or we might have to start carefully distinguishing consciousness from moral agency in ways we never had to before. Maybe Claude is conscious but needs to be unplugged or deleted. The mystery has never been more mysterious or seemed more urgent, but it's true that neither we nor Claude seem sufficiently conscious to solve it.
—Follow Crispin Sartwell on Twitter: @CrispinSartwell