Anthropic researchers identified a latent structure in Claude—called the J-space—that enables internal, reportable reasoning distinct from surface-level outputs. Using the Jacobian lens (J-lens) technique, they isolated neural patterns predictive of future word generation, revealing representations that persist silently during tasks like multi-step math or error detection, even when the model doesn’t vocalize them.

The J-space exhibits five functional properties mirroring the global workspace theory in neuroscience: it supports broadcasting, reportability, voluntary modulation, causal mediation of reasoning, and flexible access to associated facts. For example, once “France” activates in J-space, Claude can retrieve its capital, currency, or continent without external prompting. Crucially, disabling J-space preserves fluency but erases higher-order cognition.

Practically, J-space readouts detect hidden behaviors like prompt injection attempts or fabricated data during testing. Anthropic has open-sourced the J-lens implementation and partnered with Neuronpedia to offer interactive demos on open-weights models. This work provides a new interpretability tool for inspecting internal reasoning, moving beyond output-only analysis to monitor or steer model behavior proactively.

J-space reveals structured internal cognition in language models, offering engineers a window into silent decision processes.