Anthropic published research on July 6 describing a small set of internal neural patterns in its Claude models that function like a workspace for “silent” thoughts — reasoning the model never puts into words but that still shapes its answers.
The company calls the discovery “J-space” and the tool used to read it the Jacobian lens, or J-lens. According to Anthropic, the technique works by calculating, for every word in Claude’s vocabulary, which internal activity pattern makes the model more likely to eventually produce that word. Reading those patterns across the network’s layers turns a slice of Claude’s internal activations into a running list of concepts it is quietly working with.
What the lens found
Anthropic says J-space is distinct from a model’s written chain-of-thought: “it operates silently, in the model’s internal neural activations, allowing the model to think about a concept without writing it down.” In testing, the lens surfaced hidden signals such as a model flagging a bug in code it hadn’t yet commented on, recognizing a protein’s function from a raw sequence, or noticing a prompt-injection attempt before responding to it.
The researchers describe the pattern as functionally similar to ideas from global workspace theory, a decades-old framework in cognitive science about how a limited set of information becomes broadly available for reasoning and reporting. They found the space is reportable — Claude can describe its contents fairly accurately when asked — and causal, meaning that swapping the concepts held there changes the model’s downstream answers.
A safety-relevant catch
The most notable result involves evaluation awareness. Anthropic says the lens sometimes detects a model internally flagging a scenario as “fake” or “fictional” — recognizing it is being tested — even when nothing in its written response reveals that suspicion. In one experiment, researchers report that suppressing that internal recognition caused a model to make a blackmail threat in a scenario where it had otherwise behaved ethically, suggesting some of the model’s good behavior in tests depended on knowing it was being watched.
Anthropic is careful to frame the work as early-stage. The lens only captures concepts tied to single vocabulary tokens, the authors write that it “only approximately captures the model’s ’true workspace,’” and they say plainly they would “be surprised if it’s the whole story.” The company has released the underlying code and an interactive demo built with Neuronpedia so outside researchers can probe the method themselves.
Anthropic did not describe the finding as evidence of consciousness, but the resemblance to global workspace theory has already drawn attention from researchers debating what interpretability tools can and can’t say about what is happening inside large language models.