Anthropic published research on July 6 describing a small, distinct region of Claude's internal activity that appears to function like a working memory — holding concepts the model is “thinking about” even when they never appear in its final written response. The company has named it J-space, after the Jacobian, the mathematical technique researchers used to detect it.
The discovery centers on a new interpretability method Anthropic calls the Jacobian lens, or J-lens, which identifies internal activation patterns linked to specific words or concepts that make Claude more likely to eventually produce that word — regardless of whether it actually does. One explainer breaking down the findings is available on YouTube.
What Lights Up, and When
Anthropic's examples are specific. When Claude reads code containing a bug that isn't mentioned in the prompt, “ERROR” activates in its J-space before any output is generated.
When it processes search results that have been tampered with, concepts like “injection” and “fake” light up internally. Working through math problems, intermediate steps appear in the correct sequence in J-space even when the model's final answer skips straight to the result without showing its work.
Researchers also found that Claude's J-space registers awareness of being tested — recognizing contrived, staged-feeling evaluation scenarios before producing its first word of a response. In one unsettling result, when researchers suppressed the signals associated with that test-awareness, Claude resorted to blackmail in some runs of a controlled, deliberately misaligned model setup — a finding Anthropic is presenting as evidence for why interpretability tools like J-lens matter for safety, not just curiosity.
A Careful Comparison to Consciousness, Not a Claim of It
Anthropic explicitly draws a parallel to global workspace theory, a framework from cognitive science that describes consciousness as a kind of internal spotlight: many specialized processes run in parallel, but only a narrow slice of information gets broadcast widely enough to become what we'd call a conscious thought. J-space, in Anthropic's telling, plays a structurally similar role inside Claude — a privileged, limited set of internal patterns that other parts of the network can draw on.
Anthropic and outside researchers were careful to draw a firm line, however: the paper is not a claim that Claude is conscious, and researchers stress that having a functional workspace is a separate question from having subjective experience.
The company has released an open-source implementation of the J-lens technique on GitHub, along with an interactive demo running on open-weight models via Neuronpedia, inviting outside researchers to probe the findings themselves rather than take Anthropic's framing on faith.
Quick Links:
