Anthropic just found something inside its Claude models that nobody designed or expected. An internal structure called “J-space” appears to work like a cognitive workspace shared across the model.

The July 6 research marks a major step in understanding what actually happens inside large language models.

What the J-Space Inside Claude Actually Is

J-space is an internal area where Claude appears to gather and share important information across the model. A simple way to think about it is as a whiteboard inside the AI.

When Claude answers a question, solves a puzzle, or follows an instruction, key details seem to appear in this shared space so different parts of the model can use them.

Anthropic found J-space using a research tool called the “J-lens.” The tool helps researchers inspect how information moves inside Claude during a task. It showed that J-space emerged during training by itself. Anthropic did not design it directly.

Follow us on X to get the latest news as it happens.

The idea is similar to a theory in neuroscience called the “global workspace.”

In humans, this describes how the brain makes important information available to different mental processes at once. For example, when someone hears a question, remembers a fact, and decides how to answer, those pieces of information need to come together.

Claude appears to do something similar. Anthropic found that Claude can describe what is inside its J-space when asked. It can also adjust those contents if instructed.

More importantly, when researchers directly changed the J-space, Claude’s answers and task performance changed too.

Why This Matters for AI Safety and Interpretability

The discovery carries real weight for AI safety. If researchers can monitor J-space activity, they can potentially uncover hidden motivations behind a model’s behavior. As a result, they gain a sharper tool for spotting when something goes wrong.

That includes detecting attacks. Monitoring J-space could reveal when a model processes a prompt-injection attempt designed to hijack its outputs. Furthermore, even a partial window into this “conscious” processing layer marks a meaningful advance for the field.

The scope remains limited, however. The vast majority of Claude’s information processing still occurs entirely outside the J-space.

Still, Anthropic released an open-source J-lens implementation and a Neuronpedia demo, inviting the broader research community to test the findings.

Subscribe to our YouTube channel to watch leaders and journalists provide expert insights.

The work builds on a longer research trail. Anthropic published a report on emergent introspective awareness in October 2025. Moreover, it launched model-welfare initiatives in April 2025, steadily exploring what happens inside its systems.

One caveat stands out clearly. Anthropic is emphatic that the research does not claim Claude is conscious or has subjective experience. The paper uses the phrase “consciously accessible” information, borrowing the vocabulary without making the leap to actual consciousness.

Источник: BeInCrypto