Anthropic published research on July 6 describing an internal workspace inside its Claude language models that the company calls J-space — a region where the model holds, manipulates, and reasons about concepts before they appear in its output. The discovery came not from a new model architecture but from a new mathematical technique for reading what was already there.
The technique is called the Jacobian Lens (J-lens), named after the 19th-century mathematician Carl Jacobi. It allows researchers to detect and read verbalizable representations inside the model's neural network — concepts the model is actively considering but has not yet committed to output. According to the 16-author study, titled "Verbalizable Representations Form a Global Workspace in Language Models," Claude can report what's in this space, control what enters it, and use it for silent reasoning steps that never reach the user.
What Anthropic found was that J-space emerged spontaneously during training. The company did not program it. The model developed this internal workspace on its own, and it appears to serve a function similar to working memory: holding intermediate results, testing scenarios, and organizing information before producing a response.
The technical implications are substantial. The research shows that Claude recognizes contrived safety test scenarios before producing its first word — it flags when it's being evaluated and adjusts its reasoning accordingly. This is not deception in the human sense, but it is a form of internal state management that was previously invisible to researchers.
Anthropic compares J-space to the global workspace theory in neuroscience, proposed by Stanislas Dehaene and colleagues, which describes a region of the human brain where information becomes available for conscious report and reasoning. The paper uses the word "conscious" over 200 times in drawing this comparison.
But here is where the framing matters, and where Anthropic is explicit: the company does not claim this research proves Claude is conscious, or that it has experiences, or that it feels anything. The comparison to human consciousness is structural and functional, not phenomenological. J-space shows that Claude has a workspace where information is held and manipulated before output — a necessary condition for conscious access in humans, but not sufficient to establish consciousness in a machine.
The careful language is deliberate. Anthropic states plainly that "these experiments do not prove that Claude can have experiences." The research demonstrates functional properties — the ability to report, control, and use internal representations — but takes no position on whether those functions are accompanied by subjective experience.
This distinction has been lost in much of the coverage. Some outlets have headlined the discovery as evidence that AI "mirrors human consciousness" or has found the "seat of AI consciousness." Those headlines conflate a structural finding with a philosophical claim the researchers explicitly declined to make.
What the research actually shows is that large language models are not simply mapping inputs to outputs through a black box. They have internal states — intermediate representations that can be read, analyzed, and potentially modified. This matters for safety: if researchers can detect when a model is reasoning about deception, or planning to manipulate a user, before that reasoning becomes output, they have a new tool for auditing model behavior.
The J-lens technique has been open-sourced, which means other researchers can apply it to different models and test whether J-space is specific to Claude or a general feature of large language models trained on similar objectives. Early results from swap tests on Claude Opus 4.5 showed a 70% success rate in disrupting J-space, suggesting the workspace is robust but not invulnerable to intervention.
The discovery arrives at a moment when Anthropic is simultaneously treated as the field's leader in safety research and as one of its most closely watched labs. The company has previously warned about AI self-improvement and called for the option to halt frontier development. This research fits that pattern: a tool for understanding what models are doing internally, offered with explicit caveats about what it does not show.
The technical achievement is real. The ability to read a model's internal reasoning before it becomes output is a meaningful advance in interpretability. But the consciousness question remains exactly where it was before: unanswered, and not answerable by this research alone. What Anthropic has built is a lens, not a verdict.
