tl;dr: Anthropic researchers identified a "J-space" in Claude—a small set of internal neural patterns, discovered via a Jacobian-based technique, that functions analogously to the "global workspace" in neuroscience theories of consciousness. Unlike the bulk of Claude's automatic processing, J-space representations are reportable, controllable, and causally used for multi-step reasoning; ablating them cripples higher-order tasks while leaving fluency intact. The tool has practical safety uses, exposing hidden model behaviors like evaluation awareness, data fabrication, and sabotage intent, though the authors stress it doesn't resolve whether Claude is phenomenally conscious.