All these debates about LLM consciousness are overlooking the fact that we already have very compelling evidence that language—by itself—doesn't seem to produce consciousness.

The evidence is us.

The only qualia we know to exist are sensory qualia. There is no human phenomenology of pure language or pure thought. When you think, linguistically or otherwise, the actors in the Cartesian theater (calm down, Dennettians, I'm just being cute) are inner voice, visual imagery, motor echoes. Only the second-order sensory effects of the thought process show up. If there was such a thing as purely linguistic consciousness, we should experience it. But we don't. Instead, the only thing we experience—the only kind of experience we are aware of and can conceive of—is based in the sensory/perceptual, things LLMs fundamentally lack.

Now, why would sensory/perceptual processing be privileged? Because it presents the world as egocentric action affordances, something that seems fundamentally related to the core notion of subjectivity. Depth is the potential space of movement. Objects are invitations to pick up. Pain is a pressure to withdraw. Hunger is a pull toward food. Subjectivity seems to appear as competition over a space of things you, the agent, could do. This accounts for what I would argue are the core properties of subjective phenomenology (besides the fundamental property of being phenomenal in the first place):

1. egocentricity—the fact that experience belongs to a 'you'

2. valence—the fact that experience matters to that you

And these aren't two independent properties that happen to co-occur. Valence is what affordances are made of. An affordance isn't a neutral description of something you could do; it's a pull or a push toward or away from doing it. Egocentricity and valence are the same structure seen from two angles: the experience is yours because it's about what you could do, and it matters because what you could do matters.

This reframes subjectivity as a kind of computation. Not what 'emerges' from that kind of computation, but a description of the very kind of computation that tends to carry phenomenology.

But back to LLMs: an LLM is a big autoregressive generator with no body, no egocentric frame, no action space, none of the properties that we observe in the kind of processes we observe in the only consciousness we know about: our own.

The evidence is us, and it’s not looking good for LLM consciousness.