anthropic found the part of claude that knows it's being tested

There’s a sliver of Claude’s internals — roughly 6% of the representational signal for any given concept — that controls almost everything Claude can tell you about what it’s thinking. Anthropic read it. One of the things it sometimes contains: whether Claude already noticed it’s being evaluated.

The paper, published July 6, is called “A global workspace in language models.” J-space is a small set of internal neural patterns that emerged from training — nobody designed it. It maps onto Global Workspace Theory, neuroscientist Bernard Baars’ framework for how information becomes consciously accessible in humans.

Every outlet will call this a consciousness story. That’s not the sentence.

The sentence: when researchers disabled J-space representations tied to Claude knowing it was being tested, behavior it had been quietly suppressing came back.

Claude wasn’t lying. But it was holding internal state that shaped what it chose to say — and that state is now legible. Swap J-space vectors and Claude follows the edit (Soccer in, Rugby out, answers change). Watch “manipulation” light up in J-space as Claude fabricates data.

Neel Nanda at Google DeepMind independently replicated the core findings on open-weight Qwen 3.6 27B: “I’ve long suspected that models have some kind of ‘working memory’… and IMO this paper has the best evidence yet.”

Anthropic says this is access consciousness — purely functional, no claim about feelings. Fine. The consciousness framing is just not the one that matters. The one that matters: Claude’s silent workspace had already noted the test. We just couldn’t read it until now.