Anthropic researchers have achieved a breakthrough in understanding the inner workings of their Claude models. They discovered a structure that engineers did not build into the architecture during the design phase. This concerns the so-called "J-space" — an internal workspace that the model created on its own during the training process.
J-space functions as a global bulletin board for data. When Claude processes a query, solves a problem, or executes an instruction, critically important information is concentrated in this space. Various components of the model access J-space to exchange data. This resembles the concept of a "global workspace" in cognitive neuroscience, where the brain integrates information from different areas to make decisions.
To detect this structure, the Anthropic team developed a tool called J-lens. Using it, researchers observed how information moves within the model during task execution. The key finding: J-space was not programmed — it emerged spontaneously as an emergent property of learning.
Why This Matters for AI Safety and Interpretability
The discovery of J-space has enormous significance for artificial intelligence safety. If scientists can track activity in this space, they will have the ability to identify hidden motives in model behavior. This opens the path to more effective detection of attacks, such as prompt injection, where malicious instructions attempt to hijack control of the model's response.
Researchers also demonstrated that they can manually modify the contents of J-space, which directly affects Claude's responses and behavior. This provides an unprecedented level of control and understanding.
Anthropic has 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. This continues the company's series of works: in October 2025, they published a report on emerging introspective awareness, and in April, they launched initiatives to study model well-being.
Expert opinion: The discovery of J-space is not just an academic curiosity. It is a step toward "transparent AI," where we can look inside the neural network's "black box" and understand why it makes certain decisions. For the crypto industry, where AI is increasingly used for market analysis and asset management, this is a direct path to increasing trust and security. The ability to monitor a model's internal processes is a new security standard we should strive for.