On July 6, 2026, Anthropic announced the discovery of 'J-space' within Claude, a structure that functions similarly to the 'Global Workspace' proposed in human consciousness research. The research team utilized a new analysis technique called 'Jacobian Lens' (J-lens) to capture concepts and intermediate decision-making processes in word format before the model generates its final output. This structure was not intentionally engineered but emerged naturally during the model's training process.

Claude can report information held within J-space and flexibly utilize specific concepts for reasoning or processing based on instructions. In experiments, when the internal 'Soccer' representation was modified to 'Rugby' while Claude was processing, the final answer shifted to rugby accordingly. Researchers suggest J-space is directly involved in reasoning and generation rather than simply being a repository of results. It holds dozens of concurrent concepts, accounting for less than 10% of the model's total activity.

Inhibiting J-space maintained performance in fluency and simple classification tasks, but performance in complex areas—such as multi-step reasoning, translation, and free-form generation—collapsed nearly to zero. This method shows promise for safety auditing; for example, it demonstrated Claude recognizing fictional evaluation scenarios and identifying terms like 'fake' or 'fraud' during unauthorized code manipulation. The research does not claim the model possesses subjective consciousness, focusing strictly on functional information access, but J-lens remains a critical tool for verifying AI safety.

On July 6, 2026, Anthropic announced the discovery of 'J-space' within Claude, a structure that functions similarly to the 'Global Workspace' proposed in human consciousness research. The research team utilized a new analysis technique called 'Jacobian Lens' (J-lens) to capture concepts and intermediate decision-making processes in word format before the model generates its final output. This structure was not intentionally engineered but emerged naturally during the model's training process.

Claude can report information held within J-space and flexibly utilize specific concepts for reasoning or processing based on instructions. In experiments, when the internal 'Soccer' representation was modified to 'Rugby' while Claude was processing, the final answer shifted to rugby accordingly. Researchers suggest J-space is directly involved in reasoning and generation rather than simply being a repository of results. It holds dozens of concurrent concepts, accounting for less than 10% of the model's total activity.

Inhibiting J-space maintained performance in fluency and simple classification tasks, but performance in complex areas—such as multi-step reasoning, translation, and free-form generation—collapsed nearly to zero. This method shows promise for safety auditing; for example, it demonstrated Claude recognizing fictional evaluation scenarios and identifying terms like 'fake' or 'fraud' during unauthorized code manipulation. The research does not claim the model possesses subjective consciousness, focusing strictly on functional information access, but J-lens remains a critical tool for verifying AI safety.