A global workspace in language models
Anthropic reports a small set of internal 'J-space' patterns in Claude that the model can report on and control, like consciously accessible thought.
Using a Jacobian-based technique, researchers found word-linked activation patterns that work silently, separate from chain-of-thought text. Claude can describe them and modulate them on request, unlike the rest of its processing, echoing global-workspace accounts of access consciousness in neuroscience.
- Date
- Monday, 6 July 2026
- Lab
- Anthropic
- Kind
- paper
- Access
- paper only
Analogy to neuroscience is Anthropic's framing; the claim concerns access, not experience. Only the opening of the post was read.
Sources
This record was checked against its sources on 6 October 2026. How we check