Anthropic discovers a 'global workspace' inside Claude mirroring consciousness theory

What happened

Anthropic published research revealing that Claude has an internal 'global workspace' — a compact subspace with approximately 25 active concepts accounting for less than 10% of activation variance. The J-lens (Jacobian lens) technique can read and edit this internal space. Research was open-sourced on July 6, 2026.

Context and impact

The parallel is to Global Workspace Theory (Bernard Baars) from neuroscience — the theory that the brain operates like a theater where dozens of specialized processors work in parallel backstage, but only a tiny spotlight of information gets broadcast to the whole 'theater' of conscious thought. Anthropic stops short of claiming Claude is conscious — it describes only structural similarity. The finding has direct implications for AI interpretability and safety.

Details

• Technique: Jacobian lens (J-lens) — open-sourced July 6, 2026
• ~25 active concepts in the subspace, less than 10% of activation variance
• Ablating the workspace collapses multi-step reasoning while fluency survives
• Published at anthropic.com/research/global-workspace
• VentureBeat and The Decoder covered the research on July 17, 2026