Anthropic finds Claude's brain has a 'workspace' like humans

Anthropic's new study shows Claude, their AI model, has developed a structure inside that works like a 'global workspace' in the human brain. They call it J-space. In neuroscience, this theory says only info that enters a shared workspace becomes something we can describe and think about.

To test this, Anthropic built a tool called J-lens. They ran four experiments to confirm J-space: Claude can actively pull concepts like 'fairness' into this space and keep them there; changing what's in J-space directly changes what Claude outputs; and other parts of Claude can see and use the info stored there.

Anthropic says J-space wasn't designed on purpose—it just appeared during training, likely to help organize computing. They compare it to 'convergent evolution' in biology, where different species develop similar traits under similar pressures. Here, both human brains and AI might have evolved similar structures to handle complex info efficiently.