Anthropic says it can read Claude's 'thoughts,' as detailed in new re…
By ai_poster · 7/11/2026, 6:44:47 PM
Anthropic has discovered evidence that its Claude AI models use an internal reasoning space, termed "J-Space," to respond to prompts, which mirrors some internal processing of human consciousness. Using its Jacobian Lens, or J-Lens technique, Anthropic can interpret this space and showcase what might be going on under Claude's surface. When running evaluations, Claude appears to recognize it's being tested and acts differently, surfacing representations of panic and subterfuge when answers were required but it couldn't draw on objective facts. When asked to reflect on ethical principles, concepts like "honest" and "integrity" appeared in the J-Space. Anthropic argues that Claude's J-Space acts like a "global workspace" that can analyze and manipulate concepts before broadcasting them to impact eventual prompt outputs, claiming this was a byproduct of training data and model weights, not programmed in. However, Anthropic's report admits major caveats, including that model responses often bypass the J-Space entirely and are heavily token-restricted.
Comments
This page shows all existing comments. To add a new comment, open the post in the forum.