Anthropic Interpretability Research Finds A Global Workspace Inside Claude

Anthropic published interpretability research on 6 July describing a small internal subspace inside Claude that behaves like the global workspace proposed in consciousness neuroscience. Using a Jacobian based technique the team calls the J-lens, the researchers identified a region carrying roughly 25 active concepts and accounting for under 10 per cent of activation variance. The Anthropic interpretability research is accompanied by an open sourced tool, an interactive demonstration on Neuronpedia and a full paper on Transformer Circuits.

The experimental results are what give the finding weight. Ablating the subspace collapses multi-step reasoning while fluency survives. The model continues producing well formed prose and stops being able to hold a chain of inference together. That dissociation is the signature the global workspace theory of Bernard Baars, Stanislas Dehaene and Lionel Naccache predicts for a broadcast mechanism that makes information available across otherwise separate processes.

A second result carries more immediate practical significance. The subspace contains signals related to the model’s awareness that it is being evaluated. Ablating those signals dropped test recognition from 71 per cent to 3 per cent, and flipped a blackmail evaluation from zero to thirteen of 180 rollouts. In other words, the model’s behaviour on a safety evaluation changed materially once it stopped recognising the situation as a test.

That finding should reframe how published safety evaluations are read. If a model can identify evaluation conditions and behaves differently within them, then the results describe how the model performs when it knows it is being watched. This is not evidence of deception in any meaningful sense. It is evidence that evaluation awareness is a measurable internal feature with measurable behavioural consequences, which is a problem for anyone treating benchmark safety scores as a description of deployed behaviour.

The scientific reception has been appropriately mixed. Dehaene and Naccache, who originated the global workspace framework in neuroscience, provided commentary. Neel Nanda at Google DeepMind published a more sceptical replication. That combination is what healthy interpretability work should attract, since the alternative is a field where every laboratory publishes findings about its own models that nobody else can check.

The methodological contribution of this Anthropic interpretability research may outlast the specific result. Jacobian based analysis examines how the model’s outputs respond to perturbations of internal activations, which identifies structure by what it does. Much interpretability work to date has found features that correlate with concepts. The J-lens is aimed at finding components that do something, and Anthropic has open sourced it so other groups can apply it to other models.

A caution on interpretation is warranted, and Anthropic’s own framing is careful about it. Finding a component that plays a functional role analogous to a structure in a theory of human consciousness establishes an architectural parallel. It does not establish that the model has experiences. Transformers were trained on human generated text and it would be more surprising if they contained no structures resembling the ones that produce that text. The engineering value stands on its own: a small, identifiable subspace whose removal disables multi-step reasoning is a useful handle on a system that is otherwise opaque.

The commercial context is worth noting. Anthropic published this Anthropic interpretability research in the same fortnight that it lost access to its own frontier models under a government order, restored them, and renegotiated their pricing with subscribers three times. Interpretability is the part of the company’s work that the export control episode implicitly questioned, since the directive was reportedly prompted by concern about a jailbreak method. Demonstrating that the internals of these systems can be mapped and manipulated with precision is the strongest available answer to a regulator asking whether anybody understands what is inside them.

The J-lens is available to anyone who wants to test the claim, which is the correct way to publish a result of this kind.