Researchers at Anthropic have made an unexpected discovery: an internal structure has formed within their flagship language model Claude that developers did not originally build into the architecture. This mechanism, called "J-space," functions as a common workspace that various components of the model access while performing tasks.
What is J-space?
Essentially, J-space is a virtual "board" inside the neural network where key information converges when processing queries. When Claude answers a question or solves a task, critically important data appears in this area so that different parts of the model can interact with it. Anthropic analysts identified this structure using a special tool called "J-lens," which allows real-time observation of information movement within the model.
Notably, J-space emerged spontaneously during training—it was not programmed by engineers. The concept strikingly parallels what neuroscience calls the "global workspace"—a mechanism that in humans provides access to important information for multiple thought processes simultaneously. For example, when you hear a question, recall a fact, and formulate an answer, your brain consolidates all this data into one point.
Experiments showed that Claude can not only describe the contents of J-space on request but also modify it. Moreover, manual intervention in this structure directly affected the model's responses and behavior.
Why this matters for AI safety
The discovery has enormous significance for the safety and interpretability of artificial intelligence. The ability to track J-space activity opens the way to identifying hidden motives in model behavior and detecting anomalies, including prompt-injection attacks, where malicious instructions attempt to hijack response control.
Currently, of course, capabilities are limited: most information processing in Claude still occurs outside J-space. However, the team has already published the source code for the J-lens implementation and released a demo version on Neuronpedia, inviting the research community to verify the results.
This research is part of a broader series of Anthropic studies on the "inner life" of models. The company is consistently moving toward understanding what happens "under the hood" of AI, and in my opinion, this is one of the most important trends in the industry right now.
Expert opinion: The discovery of J-space is not just an academic curiosity. It is a step toward creating truly controllable and transparent AI. If we can "read the thoughts" of models at the level of their internal workspace, we will gain the key to solving fundamental safety challenges that the entire industry faces today.