Anthropic researchers find a verbalizable 'global workspace' in language models

Anthropic interpretability researchers published work identifying what they called a “workspace” of internal representations that language models draw on when reporting their own reasoning — a small, evolving set of concepts distinct from the much larger volume of processing happening elsewhere in the network. To find it, the team built a new probing technique, the “Jacobian Lens,” which measures a representation’s average linear effect on a model’s likelihood of producing a given token across thousands of contexts, isolating concepts the model is disposed to verbalise from ones used only in a specific context.

The paper draws its framing explicitly from global workspace theory, an account of consciousness from cognitive neuroscience in which information is broadcast from a limited-capacity hub to many specialised processes. The authors report several matching properties in the models they studied — the workspace representations can be verbally reported, deliberately activated, used in multi-step reasoning, and generalise across tasks — while noting real disanalogies: transformers lack the recurrent loops and competitive “ignition” dynamics that characterise the brain’s workspace, and the paper explicitly takes no position on whether any of this bears on subjective experience.

The reported workspace was small: never more than 10% of a layer’s activation variance, concentrated in roughly the middle third to two-thirds of the network, with about 25 workspace vectors active at once. Ablating the top ten workspace directions collapsed two-hop reasoning accuracy to near zero while leaving simpler tasks such as sentiment classification largely intact, and a companion intervention reduced verbalised awareness of being under evaluation in one setting from 71% to 3%. The authors framed the practical payoff as a new class of safety-auditing tool, capable of surfacing internal reasoning that a model’s output does not disclose, rather than as a claim about machine consciousness.