Anthropic Reports Finding a 'Global Workspace' Inside Claude

On July 6, 2026, Anthropic published research identifying an emergent internal workspace called "J-space" that lets Claude hold and manipulate reasoning steps it never writes into its output. The findings appear on the company's research page, "A global workspace in language models," and according to Anthropic's research page the region formed on its own during training rather than being designed. The full paper, "Verbalizable Representations Form a Global Workspace in Language Models," is available at transformer-circuits.pub.