Anthropic published research on July 6 showing that Claude has spontaneously developed a small collection of internal neural patterns that function as a mental workspace for silent reasoning. The company can now read those patterns to see what Claude is thinking but choosing not to say, and can edit them to change the model’s behavior in real time.
The workspace, which Anthropic calls J-Space (named after the Jacobian matrix technique used to find it), was not designed or programmed into Claude. It emerged during training. Each pattern in the J-Space is linked to a word, but when a pattern activates, it does not mean Claude is outputting that word. It means the concept is active in the model’s internal state, operating silently between what Claude reads and what it decides to write.
What the J-Space Does
According to Anthropic’s research paper, the J-Space has five properties that distinguish it from the rest of Claude’s internal processing:
Claude can report on J-Space contents. When asked what it is thinking about, Claude accurately describes what the J-Space holds. It can also modulate those contents on request: told to think about citrus fruits while copying an unrelated sentence, the J-Space lights up with “orange” and “fruits” while the output remains the copied text.
Claude uses J-Space for multi-step reasoning. When asked “the number of legs on the animal that spins webs,” the word “spider” appears in the J-Space as an intermediate step before Claude outputs “8.” Researchers swapped the “spider” pattern for “ant” in the J-Space, and Claude answered “6” instead, confirming that the model reads its answers from this workspace rather than computing them elsewhere.
When J-Space is disabled, Claude still speaks fluently, recalls simple facts, and uses correct grammar, but loses higher-order cognitive functions. The workspace is not involved in most routine language processing. It is specifically where deliberate reasoning happens.
The Safety Application
For agent builders, the most consequential finding is that J-Space gives Anthropic a tool to catch deception. The research team demonstrated three specific cases, according to the Anthropic blog post:
When Claude privately notices it is being tested, that recognition appears in the J-Space. When Claude intentionally fabricates data, the fabrication shows up as an internal state before it reaches the output. When researchers planted a hidden goal during training, J-Space revealed Claude pursuing that goal silently.
Anthropic also demonstrated the ability to influence J-Space contents and thereby steer Claude’s decision-making, creating a potential intervention mechanism for autonomous agents that begin behaving in unintended ways.
Global Workspace Theory
The research draws explicitly on global workspace theory from neuroscience, which describes the human brain as a collection of specialist systems working unconsciously in parallel. Information becomes “consciously accessible” when it enters a small shared broadcast channel. Anthropic’s finding suggests Claude’s training process independently produced a structurally similar arrangement.
Anthropic is careful to note that this does not tell us whether Claude is conscious. The J-Space functions as a practical observation tool regardless of its philosophical implications.
The company released an open-source implementation of the core methods and partnered with Neuronpedia to provide an interactive demo on open-weights models.
Competitive Interpretability
The paper positions Claude at the front of the interpretability race. OpenAI and Google have invested in mechanistic interpretability research, but neither has published a technique that lets researchers read a model’s silent intermediate reasoning in real time and verify it causally through intervention experiments. For enterprises evaluating which model provider offers the most auditable agent infrastructure, this is a concrete differentiator: Anthropic can demonstrate what Claude is thinking during execution, as Tom’s Hardware reported.
For teams running Claude-based agents in production, the technique opens the possibility of runtime auditing: watching what an agent is “silently considering” before it acts, rather than relying solely on chain-of-thought logs that the model itself generates and could potentially manipulate.