Technology
Inside Claude, Researchers Found a Hidden Workspace for Deliberate Thought
Interpretability researchers at Anthropic have identified a small, privileged region inside Claude that functions like a workspace for conscious-style reasoning. The team calls it J-space: a limited set of internal neural representations the model can report on, hold in mind, and use to reason about deliberate problems, sitting on top of a much larger volume of processing that it cannot directly access.
What a global workspace is
The idea comes from a long-standing theory in neuroscience and cognitive science. A global workspace is a bottleneck through which information becomes available to multiple parts of a system at once, enabling focused, goal-directed reasoning rather than automatic reaction. Claude's J-space shows a strikingly similar structure: most of the model's computation runs in a broad background, but a narrow set of representations gets promoted into the workspace where concepts are actively manipulated.
How the discovery was verified
The researchers did not just observe correlations. They programmatically altered a concept represented in J-space—swapping, for example, one sport for another—and watched the model's final output change to match the edit. That causal link is what separates the finding from a statistical guess: the workspace representation is not merely tracking an answer, it is participating in producing it. The workspace also appeared to emerge on its own during training rather than being explicitly designed.
Why it matters for AI development
Locating a workspace-like structure inside a language model gives engineers a concrete handle on how the model reaches decisions. It supports more reliable steering, clearer explanations of model behavior, and better ways to check whether a model is genuinely reasoning rather than guessing. If such workspaces become a standard design feature, future models could be built with transparency and controllability baked in from the start.
Knowledge takeaway: J-space is a small set of internal representations in Claude that act as a global workspace for deliberate reasoning. Researchers verified its role causally by editing a stored concept and seeing the output follow. Finding workspace-like structures in LLMs could make future AI systems easier to explain, steer, and trust.