Nobody knows what they’re talking about on A.I. consciousness

Your hot takes won’t solve the major riddles of Western philosophy.

This is something that people have been debating for a long time. (Image made with ChatGPT)

I was a philosophy major in college. I was also interested in politics, so I took several classes on political philosophy. When I became a journalist, I assumed the political philosophy would at some point have some relevance to my professional work and that most of the other coursework would be buried in my mind forever, just some interesting ideas I learned about once.

Instead, the rise of artificial intelligence — and especially of chatbots powered by increasingly sophisticated large language models — has made classes I took decades ago about the philosophy of mind and the philosophy of language increasingly relevant to current discourse.

This generally manifests not as any great insight into what’s going on, but instead as constant annoyance that tons of people are mouthing off about debates that have been raging for thousands of years as if they have no idea anyone has ever argued about this stuff before.

The latest offender is the science fiction writer Ted Chiang, who assures us that “No, Artificial Intelligence Is Not Conscious” in a take that makes some good points about A.I. but fundamentally blunders because he has no explanation of what consciousness is.

If you point at a duck and say “really that’s a kind of dinosaur,” someone might say that’s ridiculous — it’s a duck. To convince them that it’s a dinosaur, though, you can’t just cite facts about ducks. You need to offer an account of what a dinosaur is.

In modern cladistics, dinosaurs are distinguished from crocodilians and lizards in that they evolved a gait in which the legs protrude directly below the body rather than sprawling out to the side. Everything that evolved from the original “legs straight down” reptile is a dinosaur, and ducks evolved from that lineage; ergo, a duck is a dinosaur. By contrast, the question of whether a shelduck is really a duck does not have an unambiguous answer, because the concept of a “duck” is not considered to be a proper biological taxon. According to modern scientific practice, you can call it a shelduck, a duck, or not a duck depending on your mood, but it’s definitely a dinosaur.

Words are just conventions, and we can use them in different ways.

“Dinosaur” was not always defined the way it is today, and “duck” may be defined differently in the future. But the point is that to reasonably assess the question of whether X is an example of Y, you need to know what Y is. So a fundamental barrier to understanding whether Claude or ChatGPT is conscious is that you need to know what “conscious” means — and this is extremely controversial.

Consciousness, not explained

My personal favorite writer on the philosophy of mind is Daniel Dennett, who died two years ago at the age of 82 and thus wasn’t quite in fighting form to debate contemporary frontier models.

Dennett laid out his theory of consciousness in his 1992 book “Consciousness Explained” and in his 2017 work “From Bacteria to Bach and Back: The Evolution of Minds.” These books were published before the capabilities of chatbots really took off, but the latter displays some familiarity with the then-embryonic modern A.I. industry (although it’s more skeptical than warranted about the likely medium-term pace of progress).

But his work provides an account of consciousness that’s broadly grounded in the philosophical school of functionalism, and it’s clear on this account of what consciousness is that there’s a real question about whether an L.L.M.-type system is conscious.

The basic idea of functionalism is that the mind is what the mind does, so a silicon mind that can do everything a biological mind can do would be a conscious mind in whatever sense a biological mind is conscious to begin with.

I don’t think the existing L.L.M.s quite qualify, largely because the way their systems of context and memory function are pretty different from a human. But if you interact with Claude, the interactions are fairly human-like. And I think that’s because there is a fairly human-like mind at work.

The intentional stance

Dennett’s view on this is something of an outlier among philosophers, most of whom I think would offer different accounts of consciousness, accounts that would probably support Chiang’s conclusion. But functionalism has a distinguished lineage. You can go all the way back to Aristotle in “De Anima” talking about contrasting perspectives on anger described as a physical system and anger described in mentalistic terms:

Hence a physicist would define an affection of soul differently from a dialectician; the latter would define e.g. anger as the appetite for returning pain for pain, or something like that, while the former would define it as a boiling of the blood or warm substance surrounding the heart.

One of Dennett’s contributions to philosophy is the idea of the “intentional stance.” He says that we don’t gain knowledge of the interior lives of others. We instead simply observe that in many cases, we best understand the behavior of another entity by thinking of it as conscious and possessing intentions.

If we’re modeling a fish, we attribute very limited amounts of intentionality to it. A three-year-old is successfully modeled as a lot more sophisticated than a fish. An adult is modeled as more sophisticated than a toddler. You, of course, could give purely physicalist accounts of human behavior — it’s all neurons firing in such-and-such a way — but that’s not a very useful way to understand the world.

By contrast, anthropomorphizing is a much better way to understand how Claude will react to things you type into the chat than trying to understand it in terms of semiconductor physics. It’s true that from a certain perspective, it is all just electrons moving around, but so are you.

Again, I am aware that Dennett’s view is an outlier, and I don’t think I’ll convince you in a newsletter.

For a more conventional approach, I’d recommend Thomas Nagel’s “What Is It Like to Be a Bat” or David Chalmers’s “The Hard Problem of Consciousness,” both of which insist that there is something much more mysterious going on than Dennett allows. Ross Douthat (like most people but few philosophers) is an A.I. consciousness skeptic because he’s religious.

I think this is all wrong and that as the experience of conversing with a chatbot converges on the experience of conversing with a very patient human, we should assume the chatbot is having an experience similar to being a very patient human.

But the main thing I am begging everyone who addresses this topic to acknowledge is that you have to address B.F. Skinner’s point that “the mystery which surrounds a thinking machine already surrounds a thinking man.” In other words, whatever you end up deciding about machine minds, you need to first have a theory of human minds — it’s not at all obvious what it means to say that you are conscious or what evidence you have to believe that I am conscious or what that word even means.