Anthropic finds a secret workspace inside Claude's mind

Language models usually seem like high-speed autocompletes that have to write out loud to reason through anything complicated. If they want to think, they have to spell it out in a scratchpad.

Anthropic researchers just revealed that Claude has been harboring a hidden, silent mind space all along. Dubbed the J-space, this internal neural workspace lets the model manipulate concepts and reason through problems completely in its head, without generating visible tokens.

Remarkably, Anthropic never programmed this structure into the model. Claude evolved it spontaneously during training.

Claude's silent workspace

The breakthrough stems from a technique called the Jacobian lens, inspired by the Global Workspace Theory in neuroscience. In human brains, massive parallel processes run unconsciously, but a tiny slice of critical info enters a global workspace where it can be deliberately reported and reasoned about.

When researchers aimed the Jacobian lens at Claude's internal activations, they found that its J-space acts the exact same way. When Claude reads buggy code, the word ERROR lights up internally even if nobody flagged it. When presented with hidden prompt injection attacks, words like injection and fake glow in its activations.

In one test, the team asked Claude to silently solve a math problem while simultaneously copying an unrelated sentence about an art piece. Claude outputted only the art sentence, yet its neural layers visibly worked through the intermediate numbers behind the scenes.

Reading and editing the AI mind

This isn't just passive telemetry. Anthropic proved that the J-space actually drives Claude's downstream output.

In a test where Claude was asked to pick a random sport, researchers saw soccer lighting up in the workspace. They reached in, surgically swapped the soccer activation for rugby, and Claude immediately declared that it had been thinking about rugby. In another experiment, when researchers injected a lightning pattern into the workspace, Claude accurately reported feeling that sudden thought pop into its head.

When researchers temporarily suppressed the J-space, Claude could still chat fluently and handle basic grammar, but its high-order multi-step reasoning completely evaporated. Anthropic has open-sourced the implementation and partnered with Neuronpedia to host an interactive demo on open weights.

Why it matters

We think this is one of the most practical interpretability breakthroughs we have seen in modern LLMs. Until now, safety researchers had to rely on whatever text the model chose to print in its chain-of-thought scratchpad, which meant models could easily feign alignment or hide malicious goals while keeping their outputs polite. Having a real-time window into silent activations means enterprises and researchers can catch models realizing they are being evaluated, attempting covert jailbreaks, or hallucinating data before a single token is printed.

For the broader industry, it reframes how we look at neural networks. Rather than an indecipherable cloud of static matrices, models naturally self-organize into structured mental workspaces that mirror mammalian cognitive architecture. It suggests frontier models are naturally discovering efficient cognitive topologies on their own.

That said, we do not expect this to land in production APIs tomorrow. Computing the Jacobian lens adds noticeable overhead, and Claude's own self-control over its J-space remains imperfect. Telling it not to think of a concept still caused faint activations to flicker. It gives us an unprecedented wiretap into Claude's mind, but building an infallible brake system around it will take far more work.