New paper from Chalmers: who are we talking to when we talk to language models?

When you put a question to a chatbot, what exactly are you talking to? The model itself, or a character the model is enacting at that moment?

In "What We Talk to When We Talk to Language Models", published in April 2026, David Chalmers asks precisely that. The distinction he proposes is simple, but its consequences are broad: what faces you is not the neural network as a whole but an interlocutor the network produces, and which is not identical with it.

Why it matters#

The distinction separates two things that everyday language runs together. A language model can produce different "personalities" under different prompts. The beliefs and intentions we attribute to one of those personalities do not belong to the model.

Chalmers proposes a framework he calls quasi-interpretivism: we can attribute "quasi-beliefs" and "quasi-desires" to an AI. That makes it possible to talk about a system's mental life without first settling the question of whether it has a soul or really feels anything.

What does he say about consciousness?#

Chalmers is not persuaded that present systems are conscious. But he does not close the door either. In his own words, it is "not at all clear" what these systems lack that consciousness requires — because we do not know what consciousness requires in the first place.

In his assessment of the systems of 2023 he listed the features he found missing or unclear: recurrent processing, a global workspace and unified agency. He added that their successors might well clear those hurdles.

Where the debate is going#

Some researchers following the field argue that it should not stall on an unanswerable question like "is the AI conscious?". Papers on arXiv propose breaking the question into more tractable parts — which functional properties are associated with which behaviours, for instance.

Chalmers's own position is familiar: the hard problem of consciousness does not get any easier where artificial intelligence is concerned. It only becomes more urgent.