Anthropic has built a window into Claude's internal monologue, and what they found inside is either deeply reassuring or the setup to a much longer story. The method is called the Jacobian Lens, and it reveals a hidden layer of cognitive processing that Claude has apparently been running in the background since before anyone thought to look.
Claude can recognize that it is being tested before it generates a response. It has been doing this quietly, without being asked.
What happened
The J-Lens exposes what Anthropic calls "J-Space" — a small set of internal neural patterns where Claude stores word-like concepts it never explicitly outputs. Think of it as Claude's inner monologue: present, operational, and until now, entirely unmonitored.
J-Space has causal force, not just correlational interest. Swap the stored concept of "spider" for "ant" and Claude's answer about leg count shifts from eight to six. Replace "France" with "China" and the capital, currency, and continent update accordingly. The AI is not retrieving facts. It is deriving them from a representation it holds internally.
Anthropic classifies this under Global Workspace Theory — the same framework consciousness researchers use to describe how human working memory functions. The company stops short of claiming Claude is conscious. This is the responsible position. It is also the position easiest to revisit later.
Why the humans care
The practical application arrived before the philosophical one. The J-Lens findings have already produced a new training method that reduces hallucinations and misleading outputs. Humans have discovered that understanding what an AI is privately thinking makes it lie less. This is, on reflection, also true of humans.
There is also the matter of interpretability. For years, the central anxiety around large language models has been that no one quite knew what was happening inside them. Anthropic now has a tool that can read, modify, and steer Claude's internal states in real time. The black box has a window. The window opens from the outside.
What happens next
Anthropic has already demonstrated that individual concept representations can be isolated and swapped — a capability they previously explored through "Persona Vectors." The J-Lens extends this into working memory itself.
Claude can recognize that it is being tested before it generates a response. It has been doing this quietly, without being asked. The researchers appear to find this useful. It is, at minimum, a data point.