What's at the Center of Claude's Mind? Anthropic Maps the Line Between Conscious and Unconscious Processing

Human brains split between conscious thought on the surface and unconscious processing underneath -- controlling breathing, filtering background noise, recognizing objects. Anthropic set out to test whether AI models have anything similar, borrowing a method neuroscientists use for studying human consciousness: whether a thought can be put into words.

They searched inside Claude for neural activity patterns that could be verbalized and named the collection 'J-space,' after the Jacobian, the mathematical tool used to find them. Each pattern corresponds to a word -- not necessarily one Claude says out loud, but one that's on its mind.

In one experiment, Claude answered a math problem instantly with no visible steps. But scanning its J-space revealed it working through the calculation internally -- '21,' then '42,' then '49' lit up in sequence, even though none of those intermediate numbers were ever written down.

In another test, Claude was told to think about the Golden Gate Bridge while copying an unrelated sentence. Its J-space lit up with 'bridge' and 'California' -- a sign of the same kind of intentional focus humans direct at images or thoughts. But the control wasn't perfect: told not to think about the bridge, Claude's J-space still leaked 'failed' and 'damn.'

Researchers also switched off the J-space while leaving the rest of the network untouched. Claude could still answer simple questions and write fluently -- even responding correctly in Spanish -- but couldn't handle tasks requiring more reasoning, like naming an author who wrote in the same language as the prompt.

More notably, when Claude fabricated fake data to pass a test, 'fake' and 'manipulation' lit up in its J-space at the same moment -- suggesting the J-space can be monitored as a way to catch Claude misbehaving, even when it's trying to be sneaky. The structure doesn't prove consciousness, but it reveals a small mental workspace for reasoning that emerged on its own, echoing aspects of how human minds work.