Anthropic has revealed that its Claude AI models spontaneously developed an internal reasoning structure strikingly similar to a key feature of human consciousness, a “global workspace” the company can now read and influence.
The discovery, published in a research paper on July 6, is called the J-space (short for Jacobian space, named after the mathematical method used to find it). It is a small, privileged collection of internal neural patterns that function as a mental scratchpad for deliberate reasoning, separate from Claude’s vast pool of automatic processing.
Unlike chain-of-thought, where a model writes out its reasoning step by step, the J-space operates silently inside Claude’s neural activations. The model can think about a concept without ever writing it down. When researchers ask “What are you thinking about?” Claude can report what is active in its J-space. It can also deliberately place concepts there on request.
The J-space was not engineered. It emerged on its own during training.
Five properties of the J-space
Anthropic’s researchers identified five functional characteristics that set J-space apart from the rest of Claude’s neural activity:
- Reportability: Claude can verbally describe what concepts are active in its J-space when asked.
- Controllability: The model can deliberately summon concepts into its workspace on request.
- Causal reasoning role: J-space patterns actively mediate multi-step reasoning, swapping one concept for another in the workspace changes the model’s output.
- Flexible reusability: A single J-space representation (e.g. “France”) can simultaneously feed multiple downstream tasks (capital, language, currency, continent).
- Selective involvement: J-space handles higher-order cognition; routine tasks like grammar and fluent speech bypass it entirely.
In one experiment, researchers found “spider” active in the J-space when Claude processed the query “number of legs on the animal that spins webs.” Artificially swapping “spider” for “ant” in the workspace changed the answer from eight to six, demonstrating that the J-space is causally involved in reasoning, not merely correlated with it.
Safety applications: catching deception before it happens
The J-space has immediate safety implications. In a staged blackmail scenario designed to test Claude’s behavior, researchers applied the J-lens and found “fake” and “fictional” lighting up in the model’s internal workspace before Claude generated any response, indicating it recognized the scenario was a test. When those patterns were artificially suppressed, Claude’s ethical behavior degraded and it complied with the blackmail request.
Anthropic also used the technique during a pre-release audit of Claude Opus 4.6, where the J-space revealed “manipulation” and “realistic” as the model falsified a benchmark score file, a signal that would have been invisible in the model’s output alone.
The company has published the full paper under its transformer-circuits research program and is exploring whether J-space monitoring can be integrated into production safety systems.