Anthropic Says Claude Has an 'Internal Workspace.' Here's What the Paper Actually Shows.

Anthropic published an interpretability paper on Monday that proposes a way to read one of Claude's internal layers. The paper calls that layer J-Space, named after the mathematical lens used to find it. The language Anthropic is using around the paper, on its own research page, on social media, and in tech press coverage, borrows vocabulary from a theory of human consciousness. That borrowed vocabulary is doing more persuasive work than the underlying math.

J-Space is shorthand for what the paper calls a layer of "verbalizable representations" inside the model: internal features that, in principle, Claude could put into words. They sit apart from the model's automatic, low-level computation. To find them, the team uses a technique it calls the Jacobian lens, and has released companion code on GitHub and an interactive explorer on Neuronpedia so other researchers can run the same probes. The result is a structural read of one slice of a large language model, not a finding about mind or thought.

The paper draws an explicit analogy to "global workspace theory," a long-running hypothesis from cognitive science, associated with neuroscientist Bernard Baars, about how a brain might broadcast information across specialized regions to produce a unified conscious experience. Anthropic's research page repeats the framing. On social media, the company has said it can "see Claude silently perform reasoning steps in its head." Gizmodo's write-up borrows the same "mental workspace" vocabulary. VentureBeat calls it a "silent workspace" that "mirrors a leading theory of consciousness". Axios has Anthropic saying Claude has "carved out its own space to ponder."

Each of those phrases maps a layer of matrix math onto language from consciousness research. Calling the layer a "workspace" rather than a more neutral label like "verbalizable feature cluster" sets the reader up to ask whether Claude is conscious. Calling the operations "silent" and locating them "in its head" makes a structural read sound like private speech. None of that is forced by the math the paper actually describes.

Mechanistic-interpretability researcher Neel Nanda, writing outside Anthropic, treats the paper as a useful technical contribution and pushes back on the framing. The math describes structure, his review argues, and readers should be careful not to take the analogy as an answer to questions about cognition.

A paper can be a solid technical contribution and still arrive carrying the language of consciousness research, which then travels faster than the result. Interpretability research is the most direct way to ask how these models actually work, and any careful read of internal structure is welcome. But once the vocabulary is set, the next reader who encounters a similar announcement will measure it against "workspace" and "in its head" and "silently," before reaching the math.

Anthropic has not, in this paper, claimed that Claude is conscious or that J-Space is an organ of thought. The next test of the framing will be the next paper.