The Disembodied Generative Model

In the AI Story

The distinction matters because it identifies what the AI cannot supply. Language follows patterns. Reality is one thing that generates those patterns, but not the only thing. Literary convention, argumentative structure, rhetorical expectation, and sheer frequency of co-occurrence all generate patterns too. The model cannot distinguish between patterns that reflect reality and patterns that reflect the structure of language about reality. The two are correlated. The correlation is imperfect. The imperfection is where hallucination lives.

Clark's framework explains why the fix for hallucination is not more data or better training. The problem is structural, not quantitative. A generative model without embodied grounding cannot check its outputs against reality, no matter how vast its training corpus. It can only check its outputs against other linguistic patterns — which is a check against fluency, not accuracy. The fluency of the output, from the brain's perspective, looks like the signature of a mind that has done the kind of careful checking biological cognition performs. The appearance is misleading.

The implication for extended cognition is direct. The human component of the human-plus-AI system brings the embodied grounding that the AI lacks. This is not sentimentality about human uniqueness. It is computational architecture. A generative model without embodied grounding is a model without a tether. Couple it with a grounded model — a brain that lives in the world, acts on the world, suffers consequences when predictions are wrong — and the extended system regains the tethering that the AI component alone cannot provide.

This is what Clark means when he says that "what we have at the moment is something that is close to the limit of passive, non-embodied approaches to AI." Further progress on this specific limitation will require architectures that give AI systems something like embodied engagement with the world — not necessarily bodies in the biological sense, but mechanisms for testing predictions against reality and updating when reality pushes back.

Origin

The concept emerged in Clark's 2024–2025 engagement with generative AI, drawing on his predictive processing framework. The 2024 TIME essay "What Generative AI Reveals About the Human Mind" laid out the parallel and the asymmetry. The 2025 Nature Communications paper deepened the analysis.

The framework converges with independent arguments from embodied cognition researchers, AI safety researchers, and philosophers of mind who have been skeptical of disembodied approaches to intelligence. Clark's contribution is the synthesis: a single framework that explains both why AI works so well at language and why its failures at reality have the specific character they do.

Key Ideas

Two generative models, not one. Brains and language models share a predictive architecture but differ in whether that architecture is tethered to reality.

Language is correlated with reality, not identical to it. Statistical patterns in language reflect literary convention, rhetorical expectation, and frequency of co-occurrence as well as reality.

Hallucination is structural. A model that cannot act on the world cannot check its predictions against the world; fluency is the only signal it can produce.

Embodiment is the fix. The human component of extended cognition brings the tethering that keeps the generative process honest.

More data won't solve it. The limitation is architectural, not quantitative.