Internal Architecture of Claude: Anthropic Has Discovered a Spontaneously Emergent "Global Workspace"
Anthropic's research team has made a breakthrough in understanding the inner workings of large language models. During an analysis of Claude's architecture, a structure was discovered that developers did not originally design into the project. This internal mechanism, named "J-space," functions as a common workspace that various components of the model access.
What is J-space and how does it work?
J-space is a virtual "board" where Claude collects and transmits key information between its modules. When the model answers a question or follows an instruction, critically important data appears in this space so that different parts of the neural network can interact with it. Essentially, this is an analog of what cognitive science calls the "global workspace"—a concept describing how the human brain integrates data from various sensory and thought processes.
J-space was discovered using a special tool called "J-lens," which allowed researchers to observe the movement of information within the model in real time. Surprisingly, this structure emerged spontaneously during Claude's training—it was not directly programmed by engineers. Moreover, experiments showed that manually altering the contents of J-space changes both the model's responses and its behavior when performing tasks.
Why is this critically important 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 the model's behavior. We gain a tool for detecting prompt-injection attacks—when malicious instructions are embedded in a query to hijack the response.
Additionally, partial access to this "conscious" layer of information processing creates new prospects for research. Anthropic emphasizes that the term "consciously accessible information" is borrowed from cognitive science and does not imply that Claude has consciousness or subjective experience. However, the very fact that we can look inside the neural network's "black box" and see how it processes queries is a giant step forward.
The company has already published the source code for the J-lens implementation and released a demo version on the Neuronpedia platform, inviting the scientific community to verify the results in practice. This research builds on a series of earlier Anthropic works, including a report on emerging introspective awareness (October 2025) and initiatives to study model well-being (April 2025).
Expert opinion: The discovery of J-space is not just an academic curiosity. For the cryptocurrency and DeFi market, where AI agents are increasingly used for data analysis and trading automation, the ability to "look under the hood" of a model means a new level of control and security. If we can track how AI makes decisions, we can prevent manipulations and errors that could cost millions of dollars. This discovery makes Claude not only more powerful but also a more transparent tool.