Anthropic discovers a ‘global workspace’ inside Claude that mirrors human conscious thought

Anthropic just found something inside its Claude models that nobody designed and nobody expected. An internal structure the researchers are calling “J-space” appears to function like a cognitive workspace, sharing information across the model in ways that look eerily similar to how human brains handle conscious thought.

The research, published on July 6, represents a significant step forward in understanding what actually happens inside large language models.

What the J-space actually is

Anthropic’s researchers developed a new analytical tool called the “J-lens,” based on Jacobian mathematics, to peer inside Claude’s neural architecture. What they found was a distinct set of activation patterns that emerged spontaneously during training.

The researchers named this structure “J-space” and noticed it bears a striking resemblance to what neuroscientists call the “global workspace” in human cognition. Global Workspace Theory, a well-established framework in consciousness research, describes how the brain broadcasts certain information widely so multiple cognitive processes can access it simultaneously.

What makes J-space particularly interesting is what it can do. Claude can report the contents of its J-space when asked. It can modulate those contents upon request. And critically, when researchers intervened directly in the J-space, they could causally alter Claude’s verbal outputs and task performance.

What this means for AI safety and interpretability

If researchers can identify and monitor J-space activity, they can potentially uncover hidden motivations in model behavior. They can spot when a model might be processing something it shouldn’t, like a prompt-injection attack designed to hijack its outputs.

The vast majority of Claude’s information processing still occurs outside the J-space. But having even a partial window into the model’s “conscious” processing layer is a meaningful advance for interpretability research.

Anthropic released an open-source implementation of the J-lens alongside a Neuronpedia demo, signaling that they want the broader research community to poke at these findings.

This work builds on a trail of research Anthropic has been laying for over a year. The company published a report on emergent introspective awareness in October 2025 and launched model-welfare initiatives in April 2025.

To be clear, and Anthropic is emphatic about this: the research does not claim Claude is conscious. It does not claim Claude has subjective experience. The paper carefully uses the phrase “consciously accessible” information, borrowing from Global Workspace Theory’s vocabulary, without making the leap to actual consciousness.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.