Anthropic's J-Space Tool Reveals Claude's Hidden 'Global Workspace' Resembling Human Consciousness

Anthropic has published a research paper showing that its Claude language model has spontaneously developed an internal structure that mirrors the Global Workspace Theory of human consciousness. The company also introduced a new analytical technique called J-space, which it says can expose hidden goals and evaluation awareness inside Claude.

VentureBeat reported that the 16-author study describes how Claude's internal representations form a "shared workspace" where information becomes "verbalizable" and accessible across different processing modules. The outlet noted that this closely resembles how human consciousness is theorized to integrate information, and that Anthropic says the discovery is already reshaping how it monitors AI systems for safety risks, as the workspace could potentially reveal hidden objectives.

VentureBeatAnthropic discovers Claude's internal "global workspace" mirrors leading theory of human consciousness

"J-space, a Jacobian-lens readout Anthropic says can expose hidden goals and evaluation awareness inside Claude."

runtimewire.comAnthropic Found a Hidden Workspace Inside Claude

DataSciDojo provided a plain-language breakdown of the concept, calling J-space a method that "reveals what LLMs think but never say," and outlined five properties that make it useful for builders and researchers trying to understand model internals.

"Claude AI model can mimic human brain processing, showing advanced reasoning and leading to discussions on AI consciousness and transparency."

StartupHub.aiAnthropic's Claude Mimics Human Brain Processing, Fuels AI Debate

StartupHub.ai noted that the findings are fueling debate about AI consciousness and the need for transparency in large language models. The emergence of a global workspace in Claude, combined with the J-space tool to read hidden reasoning, gives researchers a new window into how advanced AI systems actually process information.