Anthropic reports a global workspace in language models

Category: ML research
Status: developing
Sources: Anthropic research, HN discussion
Summary: Anthropic published interpretability research on 2026-07-06 describing a "J-space," a set of internal representations that it says functions like the global workspace of cognitive neuroscience. Using a "Jacobian lens" method to read which internal patterns push the model toward specific tokens, the researchers report five properties: the model can report on injected thoughts, deliberately activate patterns on instruction, have its reasoning changed when patterns are swapped, reuse a single representation across several downstream tasks, and leave most fluency, grammar, and fact recall outside the workspace. They report the method surfaced the model privately noticing it was under test and recognizing hidden goals in deliberately misaligned variants.
Why it matters: A readable, causally load-bearing internal workspace is a concrete handle for interpretability and safety tooling, and the causal-swap results are a stronger claim than correlational probing.
Follow-up: Watch for independent reproduction of the Jacobian-lens method, any released tooling, and whether the reported reportability and causal properties hold on non-Anthropic models.