Claude has an internal workspace, and Anthropic built a tool to read it

Anthropic's interpretability team says Claude carries an internal "workspace" that behaves a little like deliberate thought, and they built a tool to read it. Using a technique they call the Jacobian lens, or J-lens, they isolated a small set of neural patterns, the "J-space," that hold the concepts most likely to shape Claude's next words. It is small: a few dozen concepts at a time, and less than a tenth of the model's activity. Most of what Claude does never passes through it.

What makes the J-space worth attention is that it shows five properties you would associate with conscious reasoning. Claude can report what is in it when asked what it is thinking, and it can push a concept in on request. The intermediate steps of a multi-step math problem appear there before the final answer, and suppressing them breaks the answer, so the workspace is doing real work rather than narrating after the fact. A single concept, say "France," then feeds many later tasks: its capital, its currency, its continent.

The practical payoff is monitoring. The team used J-lens to catch a model privately registering a staged scenario as "fake" before it replied, to spot data fabrication, and to see hidden goals in models they had deliberately misaligned. They are careful on the philosophy. This looks like "access consciousness," the ability to report and reason with a thought, and says nothing about whether Claude has experiences. The method is admittedly rough, since it only reads concepts that match a single token. The full write-up is on Anthropic's research blog.

Why it matters

If you work on AI safety or evaluation, this is a concrete handle on a hard problem: a way to watch what a model is attending to and catch it recognizing a test or concealing a goal before it acts. It is early and imperfect, so treat it as a new signal to validate on your own probes, not a finished detector.