A hidden "workboard" discovered inside Claude: what is J-space and why it's a revolution in AI
Anthropic's research team has made an unexpected discovery that could fundamentally change our understanding of how large language models work. A structure was found deep within Claude that developers did not originally build into the architecture. This is the so-called "J-space" — an internal workspace where the model collects and processes key information necessary for performing tasks.
Imagine a virtual whiteboard that different components of the neural network can access simultaneously. When Claude receives a request, solves a problem, or follows an instruction, critically important data first enters J-space, and only then is distributed to other modules. This was not designed by engineers — the mechanism emerged spontaneously during the model's training process.
How J-space works and why it resembles the human brain
To detect this structure, Anthropic specialists developed a special tool called "J-lens." Using it, researchers observed how information moves within Claude while performing a task. It turned out that J-space functions as a global workspace — a concept well known to neuroscientists. In humans, a similar system provides simultaneous access to important information for multiple thought processes: for example, when we hear a question, recall a relevant fact, and simultaneously decide how to answer it.
Furthermore, the Anthropic team demonstrated that the content of J-space can not only be read on demand but also modified. When researchers manually altered the data in this space, Claude's responses and behavior changed accordingly. This opens up a direct possibility to influence the model's decision-making process.
Why this is critically important for AI safety and interpretability
The ability to monitor J-space activity is a breakthrough for artificial intelligence safety. Monitoring this "conscious" processing layer allows for the detection of hidden motives in the model's behavior and timely identification of when the system begins to operate unreliably. In particular, this opens new horizons for detecting prompt-injection attacks, where malicious actors attempt to insert harmful instructions into a query to hijack the response.
Anthropic has already published the source code for the J-lens implementation and released a demo on the Neuronpedia platform, inviting the research community to verify the results in practice. This continues a series of the company's works: in October 2025, a report on emerging introspective awareness in models was published, and in April 2025, initiatives to study the well-being of AI systems were launched.
My expert perspective: This discovery marks a transition from a "black box" to a "transparent mechanism" in AI. The ability to look into the model's "workspace" and even influence it is not just a scientific sensation, but a practical tool for creating truly safe and controllable systems. However, it is worth remembering that Anthropic emphasizes: the term "consciously accessible" information is borrowed from cognitive science and does not imply that Claude possesses true consciousness or subjective experience. We are still far from that, but a step in the right direction has been taken.