Anthropic Finds Claude AI Has a Human-Like “Global Workspace” for Reasoning

Anthropic has found a surprising internal structure inside its Claude AI model that works a little like the “global workspace” theory in neuroscience, where certain thoughts become available across the brain so a person can hold them, explain them, and reason with them.

The company described this structure as J-space, an internal area that becomes active when Claude handles complex tasks, solves puzzles, or keeps a specific idea in mind while forming an answer. In simple terms, it works like a shared mental scratchpad where important concepts stay visible while the model processes a response.

Anthropic researchers used a tool called J-lens, or Jacobian Lens, to study how Claude’s internal signals affect its future output. The tool helped them look past routine background calculations and focus on the concepts that shaped what Claude was likely to say next.

The researchers tested this by asking Claude what it was thinking about, then checking whether those concepts appeared inside J-space. They also asked Claude to keep ideas like fairness in mind, and the model appeared to hold that idea in the same shared space while continuing its task.

This finding becomes more interesting because Anthropic says it never designed J-space directly into Claude. The structure appears to have formed during training because it helped the model organize complex computation more effectively.

That does not mean Claude has human consciousness in the dramatic way science fiction films imagine it, but it does suggest that advanced AI systems can develop internal reasoning patterns that resemble some theories of human thought. For researchers, the bigger point is clear: studying these hidden structures can help explain how AI models reason, plan, and make decisions.