Anthropic researchers made an unexpected discovery: a structure formed inside their flagship language model Claude that developers did not design during the engineering phase. This internal mechanism, dubbed "J-space," functions as a common workspace that various components of the model access to exchange data.
The research results, published on July 6, represent a significant step forward in understanding what exactly happens "under the hood" of large language models (LLMs). J-space is a virtual "board" where Claude collects and transmits key information across the network. When the model answers a question or executes an instruction, critically important data appears in J-space, making it accessible to all involved neural circuits.
How J-space works and why it matters
Anthropic identified J-space using a specialized visualization tool called J-lens. By observing the movement of information within the model during task execution, scientists discovered that this structure emerged spontaneously during the training process. It was not directly programmed by engineers.
Conceptually, J-space is strikingly similar to what cognitive neuroscience calls the "global workspace." In humans, this system enables multiple thought processes to simultaneously access important information—for example, when you hear a question, recall a fact, and formulate an answer, the brain brings all data together. Claude operates similarly: the model can not only describe the contents of J-space upon request but also modify it if asked to do so. Moreover, manual intervention in J-space directly changed the model's responses and behavior.
Implications for AI safety and interpretability
This discovery has enormous significance for artificial intelligence safety. The ability to track J-space activity opens the way to identifying hidden motives in model behavior—for example, detecting attempts at prompt-injection attacks, where malicious instructions try to hijack response control. Even partial access to this "conscious" layer of information processing gives researchers a powerful tool for monitoring system reliability.
It is important to emphasize: Anthropic does not claim that Claude possesses consciousness or subjective experience. The term "consciously accessible information" is borrowed from cognitive science and does not imply the presence of true consciousness. Nevertheless, the company has already published the source code for J-lens and posted a demo version on Neuronpedia, inviting the scientific community to verify the results in practice.
Expert opinion: The discovery of J-space is not just an academic curiosity. It is a practical step toward creating "X-ray vision" for AI. In the coming years, monitoring such internal structures will become a safety standard for all serious LLM projects. If we can "see" how a model makes decisions, we can make them safer, more predictable, and, critically, more protected from manipulation.