Anthropic Discovers 'Global Workspace' in Claude's Internal Processing
Original: A global workspace in language models
Why This Matters
Findings suggest large language models develop structured internal reasoning systems, with implications for AI interpretability and alignment research.
Anthropic published research on July 6, 2026, revealing that Claude has developed an internal neural structure called 'J-space'—a small set of patterns that act like a silent scratchpad for reasoning, emerging spontaneously during training without explicit programming.
Anthropic's new paper identifies a distinct internal structure in Claude called the 'J-space,' named after the mathematical concept of the Jacobian used to detect it. Unlike the model's chain-of-thought output, J-space operates silently within neural activations, allowing Claude to 'think' about concepts without writing them down. Researchers found that J-space representations have four key properties: Claude can report on them when asked what it is thinking; it can modulate them on request (e.g., 'think about this silently'); they mediate multi-step reasoning, with intermediate steps lighting up in J-space even when not verbalized; and they are flexibly reusable across tasks—activating 'France' enables recall of its capital, currency, and continent. Crucially, J-space was not designed by Anthropic but emerged during training. When researchers blocked J-space access experimentally, Claude continued normal conversation but lost higher-order cognitive functions. The research draws on global workspace theory from neuroscience, which posits that conscious access arises when information is broadcast from specialized subsystems to a shared workspace. Anthropic describes this as evidence that a functionally analogous distinction between 'conscious' and 'unconscious' processing has spontaneously emerged in large language models.