Researchers at Anthropic have discovered what they call a "silent space" or J-space within Claude—a region of the model where complex reasoning appears to be processed. When this space is suppressed, Claude loses significant capability in inference, creativity, and multi-step logic.
The research team claims that Claude's architecture exhibits five properties associated with human conscious access: verbal report (the ability to articulate reasoning), directed modulation (control over processing), internal reasoning (silent deliberation), flexible generalization (applying knowledge across contexts), and selectivity (focusing attention on relevant information). These characteristics suggest structural parallels between how the model processes information and how human consciousness operates, though the findings raise ongoing questions about what consciousness in AI systems actually means.
