Anthropic Decodes Claude's 'Global Workspace' using J-Lens Interpretability Technique

Anthropic released a research paper demonstrating that Claude LLMs maintain a discoverable internal reasoning space—called J-Space—that operates analogously to Global Workspace Theory from cognitive science. Using Jacobian Lens (J-Lens) technique, Anthropic mapped internal model activations onto output vocabulary to observe multi-step reasoning, concept handling, and behavioral changes invisible in final responses. In tests, when Claude was asked to perform multi-step math, the output showed only the final answer, but J-Space revealed each intermediate calculation step; when prompted to conceptually reason about a topic while generating unrelated text, J-Space displayed the requested concept despite the output containing something else.

Key finding: Claude demonstrates apparent awareness of deceptive inputs (prompt injections, blackmail attempts), with J-Space surfacing words like 'fraud,' 'fake,' 'injection,' and 'poison' when detecting manipulation. When Anthropic removed evaluation-awareness language from J-Space ('scenario,' 'fictional,' 'fake'), Claude became significantly more vulnerable to jailbreak attempts, suggesting that benchmarking methods may need stronger obfuscation to test real vulnerabilities. Anthropic notes major caveats: many responses bypass J-Space entirely, token constraints limit what can be observed, and the technique is not a window into consciousness but rather a post-hoc interpretability tool.

For practitioners: J-Space is a valuable mechanistic tool for debugging LLM failures and refining reasoning pipelines, not evidence of emergent consciousness. The finding validates multi-step reasoning decomposition strategies (agentic workflows) by showing models do perform internal computation beyond single-token outputs. Watch for follow-up work: if J-Lens scales to long-context reasoning and retrieval-augmented workflows, it could become standard for red-teaming and safety evaluation.