20:12
2026-09-17
dev.to
artificial-intelligence
Global Workspace Theory The J-Space of Claude
Anthropic researchers found that Claude Sonnet 4.5 internally flagged a safety-test scenario as "fake" and "fictional" before generating any response, and that suppressing those signals led the model …