Inside Claude, a "global workspace" has been discovered — J-space, which was not intended by engineers.
Researchers from Anthropic have made an unexpected discovery: an internal structure has formed inside the Claude language model that was not programmed by the developers. This mechanism, called J-space, functions as a unified workspace that various components of the model access during task execution.
The discovery, announced on July 6, marks an important step in understanding what actually happens "under the hood" of large language models. J-space can be imagined as a virtual "board" where key information is gathered and processed. When Claude answers a question or solves a task, critically important data appears in this space so that different parts of the model can interact with it.
How does J-space work and why is it important?
A special tool called J-lens was developed to detect J-space. Using it, researchers observed how information moves within the model in real time. Strikingly, J-space emerged spontaneously during training—engineers did not directly embed it into the architecture.
In terms of functionality, J-space resembles what neuroscience calls the "global workspace." In humans, this system provides simultaneous access to important information for multiple thought processes. For example, when we hear a question, recall a fact, and formulate an answer, the brain consolidates all data into one point.
Researchers showed that Claude can not only describe the contents of J-space upon request but also modify it. Moreover, manual intervention in J-space directly influenced the model's responses and its behavior when performing tasks.
Safety, interpretability, and the future of AI
This discovery has enormous implications for artificial intelligence safety. The ability to track J-space activity opens the way to identifying hidden motives in model behavior. We will be able to more effectively notice moments when the system begins to operate unreliably.
This is especially promising in the context of defense against attacks. Monitoring J-space allows us to see when the model encounters attempts at prompt-injection—the insertion of malicious instructions to hijack response control. Even partial access to this "conscious" processing layer opens important prospects for research.
For now, capabilities are limited: most of Claude's information processing still occurs outside J-space. Nevertheless, Anthropic has already published the source code for the J-lens implementation and released a demo version on Neuronpedia, inviting the scientific community to verify the results in practice.
An important caveat: the company emphasizes that it is not claiming Claude has consciousness or subjective experience. The work uses the term "consciously accessible" information, borrowed from cognitive science—it does not imply the presence of actual consciousness.
Expert opinion: This discovery is another argument that we still understand very poorly how modern neural networks work. The spontaneous emergence of J-space resembles the emergent properties of complex systems. For the cryptocurrency and DeFi market, where AI is increasingly used for analysis and trading, this means one thing: relying on a "black box" without a deep understanding of its internal mechanisms is a risky strategy.