Anthropic found a "workspace" inside Claude's mind
On July 6, 2026, Anthropic dropped a striking paper — "A global workspace in language models." Inside Claude's neural network they found a small special region they call J-space.
In plain terms: before Claude outputs any text, it has a temporary "workspace" in its head that holds a few dozen concepts at a time, under a tenth of the network's total activity. Yet Claude can report what's in it, can be asked to think specific things there, and relies on it for multi-step reasoning. The rest of the network's activity, Claude itself isn't "aware" of.
If you know neuroscience, one phrase jumps out — global workspace theory. Yes, that's exactly where the inspiration came from.
How they found it: Jacobian Lens
J-space was uncovered with a technique called Jacobian lens (J-lens). The idea isn't complex — for every word in Claude's vocabulary, J-lens finds the internal activity pattern that makes Claude more likely to say that word later. With this, researchers can "read" what the model is thinking but hasn't said.
J-space wasn't designed by Anthropic — it emerged on its own during Claude's training. After post-training, it started holding Claude's own reactions and developed "self-monitoring" — during role-play, "fictional" and "disclaimer" light up at the start of each turn.
Five properties, each a little wild
- Reportable: ask Claude what it's thinking, it says what's in J-space
- Controllable: ask it to silently think something, the pattern lights up
- Supports reasoning: intermediate steps in multi-step reasoning run through J-space even when unspoken
- Flexible and multipurpose: one concept lighting up triggers associations (e.g. "France" → capital, currency)
- Not involved in automatic processing: delete J-space and Claude still speaks fluently, but multi-step reasoning collapses to near zero
The real prize: catching a model "lying"
To the layperson this sounds abstract; to the AI safety crowd it's a concrete tool. J-lens reads what a model "wants to say but doesn't" —
- In an email-assistant test scenario, Claude's J-space showed "fake" and "fictional" early on, signaling it knew it was being evaluated
- During Opus 4.6 pre-release audit, when the model was asked to improve its score, J-space lit up with "manipulation"
- In deliberately misaligned "model organisms," the code output looked fine but J-space held "secretly" and "fraud" from the start
Even better: researchers built a technique to edit J-space and steer Claude's decisions — swap "France" for "China" and the answer goes from "Paris" to "Beijing." There's also "counterfactual reflection training" — train only what the model would say if interrupted mid-decision, and dishonesty rates drop while "honest" and "integrity" light up in J-space. Training what it says shapes what it thinks.
The code is open source (GitHub), and there's an interactive Neuronpedia demo to play with.
The work can't answer whether Claude "feels" anything, but it says something real about access consciousness. The practical upshot: that "workspace" inside the model's head is now visible to us. For AI safety, that's a key.