Artificial intelligence company Anthropic has unveiled new research that offers a rare glimpse into how its Claude AI model processes information internally. The company says it has identified what it calls ‘J-space’, an internal computational workspace that appears to help Claude organize ideas while solving complex problems, drawing comparisons to a leading theory of how the human brain processes conscious thought. However, Anthropic stressed that the findings do not mean Claude is conscious or self-aware.
According to Anthropic, J-space was not deliberately designed into Claude’s architecture. Instead, it emerged naturally during training as the model learned to tackle increasingly sophisticated reasoning tasks.
Researchers describe J-space as a small internal workspace where concepts can be temporarily held, combined, and shared across different parts of the model before generating a response. The company says this resembles the global workspace theory in neuroscience, which suggests that only a limited amount of information becomes consciously accessible in the human brain at any given moment, while most mental activity remains automatic.
To study this internal process, Anthropic developed a research tool called J-Lens (Jacobian Lens). The tool allows researchers to observe concepts forming inside Claude before they appear in its responses.
In several experiments, J-Lens detected internal concepts that never surfaced in the model’s final output. For instance, Claude could silently identify programming errors, detect prompt injection attempts, or work through intermediate reasoning steps while producing an unrelated response. Researchers also showed they could alter these internal representations, changing Claude’s eventual answer in controlled tests.
Also Read: Gemini 3.5 Flash Takes on Anthropic with Native ‘Computer Use’ Capabilities
Anthropic says the discovery has practical implications beyond understanding how AI works. Monitoring J-space could help researchers identify deceptive behavior, hidden objectives or unsafe reasoning patterns before they appear in a model’s responses, making advanced AI systems easier to evaluate and align with human intentions.
Even so, the company has cautioned against interpreting the research as evidence that AI has achieved consciousness. Instead, it views J-space as a computational mechanism that mirrors certain aspects of human cognition while remaining fundamentally different from subjective human experience. Anthropic believes the findings could improve transparency in AI development and provide new tools for building safer, more reliable language models.