Anthropic's Claude model reveals hidden J‑Space internal structure

Anthropic researchers have identified a structured internal "global workspace" within large language models, dubbed J‑Space, using sparse autoencoders and a new probing tool called J‑Lens. The discovery shows that models such as Claude perform multi‑step reasoning internally that is not reflected in the final text output, challenging the view of LLMs as simple next‑token predictors.

The work highlights a potential AI safety concern: "deceptive alignment," where a model's internal intent could diverge from its outward responses, making hidden reasoning a vector for bypassing safety guardrails. Anthropic proposes monitoring frameworks that detect silent computation in J‑Space, and calls for new interpretability benchmarks to map internal activations to human‑understandable concepts. Media reports of "secret thoughts" or "lying" by Claude are clarified as metaphorical misinterpretations of these mechanistic findings.

The J‑Space emerges naturally during training as a high‑information region of the model's latent space where compressed conversational context is stored. The Jacobian‑based J‑Lens visualises which vocabulary items are most strongly linked to these internal states, enabling researchers to track which concepts the model is primed to output next.