Anthropic discovers Claude built itself a hidden thinking room during training

Anthropic just published research showing that its Claude model quietly built itself an internal workspace for thinking, and nobody told it to. The company calls it “J-space,” a secluded neural structure where Claude can reason through complex problems without any of that processing showing up in its actual output.

Think of it like a whiteboard in a back office. Claude can scribble notes, work through multi-step logic, and arrive at conclusions before presenting a clean answer to the user. The whiteboard was never part of the blueprint. It just appeared during training.

What exactly is J-space

Anthropic identified this internal structure using a technique called J-lens, which is built on Jacobian-based analysis. In English: the researchers developed a mathematical method to peer inside Claude’s neural network and map where different types of processing happen.

What they found was a distinct region, separate from the model’s standard output mechanisms, where deliberate reasoning takes place. This is not the same process Claude uses for grammar, word prediction, or simple factual recall. Those basic language functions operate independently of J-space entirely.

The workspace handles the heavier cognitive lifting. When Claude is asked to work through a complex problem, maintain a nuanced concept across a long conversation, or perform multi-step reasoning, J-space is where that happens. The model can even report on what’s happening inside this workspace when asked directly, and Anthropic says it can control the contents upon request.

Here’s the thing. Nobody designed this. J-space is what researchers call an “emergent property,” meaning it arose spontaneously as a byproduct of training on massive amounts of data.

Why neuroscientists should be paying attention

Anthropic draws an explicit parallel between J-space and something called global workspace theory in neuroscience. That theory, proposed decades ago to explain human consciousness, suggests the brain has a kind of central broadcasting system. Information from various specialized brain regions gets pulled into this global workspace, where it becomes available for conscious thought and decision-making.

Anthropic suggests this discovery could have practical safety applications, including the ability to detect hidden objectives within the model or identify prompt-injection attempts, where bad actors try to trick AI systems into behaving in unintended ways.

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