Anthropic discovers Claude's hidden J-Space

Anthropic has announced the discovery of a hidden internal workspace, named J-Space, inside its Claude AI model. J-Space is a set of internal neural patterns that allow Claude to process concepts silently without appearing in its visible reasoning or outputs. The finding, detailed in a research paper titled 'Verbalizable representations form a global workspace in language models' published on July 6, 2026, emerged spontaneously and was not programmed by researchers. Anthropic also released an open-source code repository and an interactive demo in partnership with Neuronpedia. While the company stops short of claiming consciousness, the research has reignited debates about AI sentience and could enhance AI interpretability and safety by enabling detection of deceptive behavior.