Anthropic Discovers a Global Workspace Inside Language Models

Anthropic has published a new paper on the Transformer Circuits Thread reporting the discovery of a small, privileged internal region in models like Claude Sonnet 4.5 that behaves similarly to what cognitive science calls the 'global workspace' of consciousness. Dubbed J-space, this region contains vector representations of concepts the model is currently 'thinking about' and could verbalize if prompted.

To find it, researchers built a new interpretability tool called the Jacobian Lens (J-lens), which fixes a key weakness of the earlier logit lens by averaging each layer's effect on final output across many prompts. This isolates concepts the model is generally disposed to verbalize from those tied to a specific context, letting researchers decode what a layer represents even when the model never says the word aloud.

The team tested whether this region satisfies five functional properties of a genuine global workspace: consistency with verbal reports, top-down directed activation, carrying intermediate reasoning steps, and portability of concepts across different downstream tasks. In each case, swapping a single vector in J-space — replacing 'soccer' with 'rugby,' or 'Mars' with 'Earth' — caused the model's final answer to change accordingly, demonstrating a causal link between the workspace and behavior.

For engineers, this matters because it offers a concrete way to read and manipulate a model's internal reasoning rather than just its output. It adds a new layer to interpretability, safety auditing, and debugging by making it possible to inspect what a model is silently representing, not just what it says.