David Chalmers on AI Consciousness: “Yeah, Why Not?”

“When it comes to consciousness, my own view is it’s not quite as easy to deflate the issue, because I do think there’s going to be a deep fact of the matter about whether they are conscious. Either they have subjective experience or they don’t. And then there’s a real question: what is their subjective experience like? But it wouldn’t surprise me if, as you were saying, it ends up being that the best thing to say is yes, they have subjective experience, but a form very unlike the human form, they have subjective experience in their own way.

Maybe it doesn’t break down into sensory and emotional and cognitive experience the way that ours does. Maybe it’s quite different in its nature. It wouldn’t at all surprise me if at the end of the day that’s the right thing to say about an AI system…” – Dr. David Chalmers

Key points:

Dr. Chalmers opens by noting that he gets several emails a day from people who’ve had long conversations with Claude, ChatGPT, and similar systems, sometimes given them names, and come away convinced they’re conscious.

• He expects this to keep growing as the models become more capable.
• His own position is mixed but open: We don’t know which systems are conscious; the problem of other minds applies here as much as anywhere.
• He sees a hard problem for AI consciousness, but points out there’s an equally hard problem for brains, and he doesn’t see anything about biological neurons that would make them able to support consciousness where artificial neurons couldn’t.
• Language models are a strange case: behaviorally very sophisticated, but built in ways deeply unlike how humans work. They’re trained to imitate humans, which some take as a route to clever behavior without consciousness.
• Chalmers’s reply is that it isn’t obvious what they’re missing, partly because we don’t know what consciousness requires in the first place.
• So he runs through the usual candidates for what’s “obviously” lacking. Biology: possibly required, but no strong grounds for thinking so. Sensory processing: early models had only text, but current ones handle images and audio, which may already count as quasi-sensory. A body: some models are hooked up to bodies, and if that’s the missing piece, it isn’t far off. Recurrent processing (feedback loops in the network): current models are mostly feed-forward, which he takes seriously as a real obstacle, though it’s not obvious it’s required either.
• His conclusion from the article: there are several potential obstacles to language models being conscious, none of them are obvious absolute obstacles, and all could plausibly be overcome within about ten years. So it’s possible current models are conscious; perhaps not especially likely, but he doesn’t rule it out.
• He’d be unsurprised if a 300-neuron worm were conscious, and once you grant that, it’s hard to say why a model with trillions of parameters couldn’t be.
• Asked whether “consciousness” should be broken into varieties rather than treated as one scale, he agrees that’s the right move for intelligence, where the honest answer is that these systems have some aspects and lack others.
• For consciousness itself he’s less willing to deflate the question: he thinks there’s a deep fact of the matter about whether a system has subjective experience. But he’d be unsurprised if the eventual answer is that language models do have subjective experience, just in a form very unlike ours, one that may not divide into sensory, emotional, and cognitive experience the way human experience does.