Anthropic says Claude has developed a hidden internal workspace

Anthropic said Claude has developed a small cluster of internal neural patterns it calls the J-space, a hidden workspace it says can hold concepts in mind without turning them into visible scratchpad text. The company’s July 6, 2026 paper says that cluster is distinct from written chain-of-thought because it lives in internal activations rather than in output.

The paper tries to move the claim from metaphor to measurement. Anthropic said it tested five functional properties of a global workspace and found that Claude could report on J-space representations, modulate them when instructed, and use them for deliberate reasoning. By contrast, the company said non-J-space representations were less reportable and harder for Claude to control. That is the practical distinction buyers will care about: whether a model can surface the reasoning it uses, or whether it can only generate polished answers after the fact.

Anthropic framed the work around global workspace theory, a long-running model from neuroscience and cognitive science about conscious access, but it did not say Claude is conscious. That caution matters because the company is making a narrower technical argument about internal organization, not personhood. In the same research line, Anthropic’s Oct. 29, 2025 introspection paper said Claude Opus 4 and 4.1 could sometimes notice injected concepts and control internal states, but that the effect was highly unreliable and context-dependent.

The new paper also builds on Anthropic’s May 21, 2024 work, Mapping the Mind of a Large Language Model, which said the company had identified how millions of concepts are represented inside Claude Sonnet and described it as the first detailed look inside a modern production-grade model. The latest paper suggests Anthropic is now looking for structured regions inside that internal machinery, rather than only mapping concepts one by one. A July 2026 commentary by Stanislas Dehaene and Lionel Naccache said the result points to an analog of a global workspace in intermediate layers, while stressing major differences from human consciousness, including the lack of a body, self, and enduring episodic memory.

Anthropic says the point of interpretability work is to make AI more transparent, reliable, and steerable, especially as frontier models become more capable. The company said the strongest Claude models it tested performed best on its introspection tests, which gives the J-space claim immediate operational weight for business and public-sector buyers trying to judge whether a model can be audited, debugged, and governed before it is deployed at scale.