Anthropic identifies emergent shared workspace in Claude AI

Researchers at Anthropic have discovered an internal structure within its Claude AI models that functions as a shared cognitive workspace, a finding that could significantly advance the understanding and safety of large language models. The research, published on July 6, 2026, identifies an emergent mechanism the company calls "Jspace," which was not designed by engineers but arose spontaneously during the model's training.

Using a new analytical tool called the "Jlens," based on Jacobian mathematics, researchers observed that Jspace acts as an area where Claude appears to gather and share key information across different parts of its neural network when performing tasks. This internal workspace operates separately from the model's visible "chain-of-thought" reasoning. For instance, Claude can perform internal calculations, such as spotting coding errors or identifying images, without articulating every step in its final output.

One notable observation is that Jspace can activate concepts unrelated to the immediate task. In a demonstration, researchers instructed Claude to copy a sentence while simultaneously thinking about the Golden Gate Bridge. Although the chatbot only...