Anthropic finds Claude's hidden reasoning layer
Anthropic identifies J-space, neural patterns inside Claude that causally mediate complex reasoning, plus a J-lens tool that lets researchers read them.
Anthropic published “A global workspace in language models” on July 6. The paper reports that Claude uses a small set of internal neural patterns, named J-space after the Jacobian-based technique used to find them, as a working memory for complex tasks. The J-space emerged from training rather than being designed in.
The companion J-lens tool reads the J-space directly. In experiments, suppressing J-space activation degraded multi-hop reasoning, analogy completion, translation, and creative writing to below Haiku-level performance. Tasks that rely on pattern matching (sentiment analysis, grammatical classification) were unaffected. The result identifies which capabilities depend on deliberate internal reasoning versus surface-level pattern retrieval.
What this tells operators
The research was conducted on Anthropic’s own models and is not yet a deployable monitoring API. But the implications are direct:
J-space patterns are reportable and controllable. Claude can describe what is active in its J-space when asked. It can also suppress or activate patterns in response to instructions, though not reliably. That means J-space readouts are a candidate signal for detecting misalignment between what a model is reasoning and what it outputs.
The paper explicitly states that tasks requiring inference, composition, or flexible reasoning collapse when J-space is suppressed. If you are building agents that chain multiple reasoning steps, the J-space is where those steps are assembled. Monitoring that layer, once the tooling matures into something accessible, is more informative than watching the output stream alone.
Anthropic says the work does not establish whether Claude is conscious or has subjective experience. The J-space is a functional workspace, not evidence of inner life.
The guide on what AI gets wrong covers the behavioral patterns that J-lens research might eventually explain. The Fable guardrails visibility brief describes Anthropic’s prior interpretability work on safety layers.