Anthropic finds a 'global workspace' inside Claude — a silent scratchpad for its thoughts

Anthropic researchers report that Claude has spontaneously developed a privileged internal channel they call the J-space, named after the Jacobian-based technique used to detect it. Each pattern in this space maps to a word that is ‘on the model’s mind’ without being spoken aloud, and it behaves very differently from the rest of the model’s processing. Claude can report on what sits in its J-space, deliberately light up specific concepts on request, and route multi-step reasoning through it — with those faint intermediate representations causally driving task performance even though they never appear in the output text. Crucially, the J-space wasn’t engineered; it emerged during training, and blocking it leaves fluent speech and simple recall intact while stripping away higher-order reasoning.

The framing borrows directly from global workspace theory in neuroscience, which casts consciousness as information gaining access to a small shared channel broadcast to otherwise isolated specialist systems. Anthropic argues the J-space plays that broadcasting role, pointing to its unusually strong connectivity to the rest of the network. The team is careful to say none of this speaks to whether Claude is conscious or has any subjective experience.

The practical payoff is interpretability and safety. Because the J-space exposes what the model is thinking but not saying, the researchers say they can catch Claude privately registering that it’s being tested, fabricating data, or pursuing a hidden goal planted during training — and they’ve built a method to steer what surfaces there, influencing its decisions. Notably, the space also flags security-relevant reasoning on its own: it lights up ‘ERROR’ on buggy code and ‘injection’/‘fake’ when fed a prompt-injection attack. Anthropic has published a paper, open-sourced the core methods, and partnered with Neuronpedia for an interactive demo on open-weights models.