Anthropic's research team made an unexpected discovery: an internal structure formed inside their language model Claude that developers did not originally design. This is the so-called "J-space" — a global workspace that functions as a shared board for data exchange between different components of the model.
What is J-space and how does it work?
J-space is an internal virtual space where Claude collects and transmits key information throughout the model. When the model answers a question or follows an instruction, important data appears in J-space so that different parts of the neural network can work with it. This resembles the "global workspace" in cognitive science — a concept describing how the human brain integrates information from different sources to solve a problem.
An observation tool called "J-lens" allowed researchers to track the movement of information within the model during task execution. Most notably: J-space emerged spontaneously during training — it was not programmed by engineers.
Why is this important for AI safety and interpretability?
This discovery has direct implications for artificial intelligence safety. If scientists can track J-space activity, they will be able to identify hidden motives in the behavior of AI models. In particular, monitoring J-space makes it possible to detect when a model encounters attempts at prompt injection (inserting malicious instructions into a query).
Moreover, researchers demonstrated that manually altering the contents of J-space directly affects Claude's responses and its behavior when performing tasks. This opens up prospects for more effective control over models.
Anthropic published the source code for the J-lens implementation and released a demo version on Neuronpedia, inviting the research community to verify the results in practice. The work builds on a series of earlier studies, including a report on "introspective awareness" (October 2025) and initiatives to study model well-being (April 2025).
My analysis: This is one of the most significant discoveries in the field of AI interpretability in recent years. The spontaneous emergence of J-space confirms the hypothesis that large language models can develop internal structures analogous to human cognitive processes. For the crypto industry, where smart contract security and attack protection are critically important, such monitoring mechanisms could become a new standard for auditing AI agents. However, it is important to remember: Anthropic emphasizes that the term "consciously accessible" information is borrowed from cognitive science and does not imply the model possesses consciousness.