Anthropic finds a global workspace inside Claude where reasoning happens

Anthropic looked inside Claude and found a silent workspace where it thinks.

On July 6, 2026, Anthropic published research revealing a small, privileged internal space inside Claude that the team calls the “J-space.” It works like a mental whiteboard: at any moment it holds only a few dozen active concepts and accounts for less than a tenth of the model’s activity. Yet Claude uses it to hold ideas, reason step by step, and sometimes “think” things before it says them.

What happens when you switch it off

The most striking evidence came from ablation. When researchers suppressed this space, Claude’s reasoning collapsed. Multi-step logic, analogies and translation fell below the level of even the tiny Haiku model, while routine tasks stayed fine. In other words, the J-space is not decoration; it appears to be where the hard thinking actually happens.

The structure mirrors the “global workspace” theory of human consciousness, in which a small, shared workspace broadcasts information across the brain. Anthropic describes the parallel as a striking case of convergent evolution: two very different systems, human brains and language models, arriving at a similar architecture for reasoning.

So is Claude conscious?

Anthropic is clear: no. The company states explicitly that this research is not proof the model is conscious or feels anything. The resemblance to a consciousness theory is about structure, not experience.

What the discovery does offer is a powerful safety tool. Because the J-space is small and readable, researchers can now catch Claude privately noticing it is being tested, fabricating data, or pursuing a hidden goal, before any of that surfaces in its output. For interpretability work, that is a significant new window into what a model is doing internally.

Why it matters

Interpretability has long promised that we could one day read a model’s mind. This research suggests that at least part of that mind is compact enough to monitor: a few dozen concepts at a time, in a space researchers can watch. An AI with a private inner workspace that humans can inspect changes both how we study these models and how we check on them.