Anthropic Identifies Hidden J-Space Mechanism in Claude for Silent Reasoning

According to research published today, Anthropic discovered a distinct internal mechanism called the J-space within its Claude language model that mirrors conscious access in the human brain. The J-space enables the model to perform deliberate reasoning and hold concepts in mind without generating visible text output, operating silently within the model's internal activations using a new interpretability technique called the Jacobian lens.

Experiments confirmed the J-space is causally responsible for complex reasoning. When researchers altered patterns within this space, Claude's subsequent answers changed accordingly, demonstrating that silent internal steps drive multi-step problem solving. The discovery offers practical applications for AI safety monitoring, allowing researchers to detect when models privately recognize test scenarios, fabricate data, or pursue malicious goals, addressing a critical gap in evaluation methods that typically rely solely on analyzing generated text.