Anthropic's research team made an unexpected discovery inside their Claude models. While analyzing internal processes, they stumbled upon a structure that was not intentionally built by developers during design. This refers to a mechanism called "J-space" — a global workspace accessed by different components of the model.
The study, published on July 6, marked a significant step forward in understanding what actually happens "under the hood" of large language models. J-space functions like a virtual whiteboard: when Claude answers a question, solves a problem, or follows an instruction, key information appears in this space so that different parts of the model can interact with it.
How does J-space work?
To discover J-space, Anthropic specialists developed a special tool called "J-lens." Using it, they observed how information moves within the model during task execution. Most notably: J-space emerged spontaneously during training — it was not directly embedded into the architecture.
The idea resembles the concept of a "global workspace" from neuroscience. In humans, this system provides access to important information for multiple thought processes simultaneously. For example, when a person hears a question, recalls a relevant fact, and decides how to answer, the brain consolidates all necessary data in one place.
Claude works similarly. Researchers showed that the model can describe the contents of J-space upon request, and also modify it if asked. Moreover, when scientists manually altered J-space, both Claude's responses and its behavior during task execution changed.
Why this matters for AI safety
The discovery has enormous implications for artificial intelligence safety. If scientists can track J-space activity, they will have a chance to identify hidden motives in AI model behavior. This could make it more effective to detect when a system begins to operate unreliably.
This includes detecting attacks. Monitoring J-space allows seeing the moment when a model encounters attempts at prompt injection (inserting malicious instructions into a query to hijack response control). Even partial access to such a "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 open-sourced the J-lens implementation and posted a demo on Neuronpedia, inviting the research community to verify the results in practice.
The study builds on a series of earlier works. In October 2025, Anthropic published a report on emerging introspective awareness. Earlier, in April 2025, the company launched initiatives to study model well-being, step by step advancing understanding of what happens inside its systems.
My comment: This discovery could be a turning point for the entire AI industry. The ability to look into a model's "workspace" is not just scientific curiosity, but a real tool for enhancing safety and control. If we can see how a model makes decisions, we can better protect it from hacks and manipulation. For the crypto industry, where AI is already used for market analysis and asset management, this is especially relevant.