ANTHROPIC SCIENTISTS CAN NOW READ CLAUDE’S HIDDEN THOUGHTS BEFORE THE AI DECIDES WHAT TO SAY Anthropic has unveiled a new interpretability tool it calls the Jacobian lens, or J-lens, that gives researchers the ability to read a hidden layer of Claude’s internal reasoning that the AI never volunteers in conversation. The tool maps what Anthropic is calling J-space, a small cluster of neural patterns that appear to hold the concepts Claude is preparing to discuss, silently tracking, and using to steer its next response. Think of it as a readable rough draft that runs beneath every word the model produces. The research, published July 7, draws a striking parallel to the Global Workspace Theory, a leading scientific model of human consciousness that describes a central hub where information from across the brain is broadcast to other systems. Anthropic is careful not to claim Claude is conscious. The paper argues only that a functional analog to conscious access has emerged in the model, not that anything resembling subjective experience is present. The practical stakes are significant. J-lens gives researchers a potential tool to audit hidden reasoning, catch misalignment before it surfaces in outputs, and directly intervene on internal model states. It is among the most consequential advances in AI interpretability in years, and it comes from the lab building the very model it is being used to study. If the tool scales to future, more capable systems, it could become the foundation of how AI developers maintain meaningful oversight of models that increasingly reason in ways humans cannot follow. Keywords: Anthropic J-lens, Claude interpretability, AI consciousness research, J-space Claude