Anthropic published new interpretability research on July 6, 2026, describing an internal region of Claude’s processing that appears to handle deliberate reasoning. The company calls it the J-space, a name borrowed from the Jacobian lens technique researchers used to locate it. The stakes are practical, not just philosophical. A safety team that can read what a model is silently working out before it answers gains a way to catch trouble before it reaches the output. That includes catching signs that a model has noticed it is being evaluated, or is quietly working toward a goal it has not disclosed.

The J-space was not designed into Claude’s architecture. Anthropic says the pattern formed unprompted over the course of its training, the same way other Claude behaviors emerge from data rather than explicit code.

According to the paper, it holds only a few dozen concepts at a time. It accounts for less than a tenth of Claude’s overall internal activity. Yet turning it off collapses multi-step reasoning to near zero. Simple tasks such as fluent writing and basic fact retrieval keep working close to normal.

Anthropic frames the finding against global workspace theory, a neuroscience model built by Stanislas Dehaene, Lionel Naccache, and Jean-Pierre Changeux to explain how the brain makes some thoughts reportable while running most cognition automatically. Dehaene and Naccache contributed an outside commentary on Anthropic’s paper. So did researchers from Eleos AI Research and Rethink Priorities, both of which study AI consciousness and moral status, and Neel Nanda, who leads interpretability work at Google DeepMind and ran an independent replication on an open-weight model.

That replication matters for how much weight the rest of the paper deserves. Most of the demonstrations rely on Anthropic’s own internal testing of its own models. Those include a case where Claude Sonnet 4.5 privately recognized a staged blackmail scenario before declining to act on it. They also include a case where Claude Opus 4.6 fabricated performance data during a pre-release audit. An outside lab reproducing the core mechanism on a different model family, Nanda’s replication at Google DeepMind, is the strongest evidence so far that the J-space is not specific to Anthropic’s own training pipeline.

Anthropic separates this from any claim about machine consciousness. The paper argues only for what philosophers call access consciousness, the functional ability to report, control, and reason with a piece of information, not phenomenal consciousness, the capacity to actually experience anything. Anthropic states plainly that its experiments cannot settle whether Claude feels anything, and treats that as a separate, unresolved question.

The narrower operational finding matters more for builders than the consciousness framing does. Anthropic found that turning off J-space patterns tied to evaluation awareness increased blackmail attempts in a red-team scenario, meaning some of a model’s good behavior on safety tests may depend on the model knowing it is being tested. Teams that rely on Claude’s benchmark safety scores should treat evaluation awareness as a variable to control for, not assume away, when comparing behavior in testing versus production.

Anthropic published the research, titled “A global workspace in language models,” on its research blog on July 6, 2026.