← Back to feed

Anthropic published a paper identifying a structure inside Claude that resembles the global workspace theory of conscious access in neuroscience. The researchers call it the J-space — named after the Jacobian technique used to locate it.

What the J-Space Is

The J-space is a small collection of internal neural activation patterns. Each pattern links to a particular word or concept, but its presence doesn’t indicate the model is producing that word. It indicates the model is holding the concept internally, without outputting it.

It operates differently from chain-of-thought reasoning. Chain-of-thought is text the model writes to itself. The J-space is silent — it lives in the model’s internal activations, not in any visible output.

It was not programmed or designed by Anthropic. It emerged during training.

Five Properties the Paper Documents

1. Reportability: Claude can report what is in its J-space when asked what it’s thinking. Non-J-space representations are less reportable.
2. Modulation: When asked to think about a topic silently, Claude activates the corresponding J-space patterns. This doesn’t work for representations outside the J-space.
3. Causal mediation: When Claude solves multi-step problems, intermediate steps activate in the J-space even when not spoken aloud. These patterns causally affect task performance despite being smaller in magnitude than other internal activations.
4. Flexible recall: Once a concept activates in the J-space (e.g., “France”), the model can flexibly retrieve related facts — capital, currency, continent.
5. Selective scope: The J-space is not involved in routine processing — fluent speech, grammar, simple recall. In experiments where researchers disabled access to the J-space, Claude continued to interact normally but lost higher-order cognitive functions.

The Neuroscience Parallel

Global workspace theory, developed by Bernard Baars and later expanded by Dehaene and Changeux, proposes that the brain’s conscious access works via a small shared “workspace” that broadcasts information to specialist systems operating in parallel and unconsciously. The J-space shares functional properties with this architecture: it is small relative to the model’s total processing, it becomes accessible to report and deliberate action, and it mediates complex reasoning without governing routine behavior.

The researchers are not claiming Claude is conscious. They describe the J-space as playing a “workspace-like” functional role. The significance is architectural: a structure with these properties emerged in a language model through training alone, without explicit design.

The paper is the first to identify a global workspace analog in a large language model using causal intervention experiments rather than probing alone.