Anthropic says Claude has a silent workspace inside it.
The new research is called "A global workspace in language models." Anthropic says Claude has developed a small set of internal neural patterns it calls J-space.
This is not chain-of-thought. It is not text the model writes to itself. It is activity inside the model where concepts can show up before, or without, appearing in the final answer.
That matters because the method gives researchers a possible way to inspect what a model is privately tracking. Anthropic says J-space can reveal silent reasoning steps, signs that Claude noticed it was being evaluated, and even internal signals around fabricated data or hidden goals in test scenarios.
The important part is not "Claude is conscious." Anthropic explicitly does not claim that. The important part is that model auditing is starting to move from reading outputs to reading internals.
Takeaway: the next safety frontier is not just what models say. It is what they are computing before they say anything.
Read more: