I read a research piece from Anthropic this week that changed how I picture what happens inside a large language model. The claim: Claude has something like a global workspace, a small set of internal patterns that behave a lot like conscious thought does in people.
The idea comes from neuroscience. In your brain most processing is automatic and invisible to you; only a narrow slice becomes consciously accessible, the thoughts you can report, hold in mind, and reason with. The researchers found Claude has its own version of that slice. They call it the J-space.
What convinced me it isn’t just a metaphor was the set of properties they tested:
- Reportability. Ask Claude what it’s thinking and it can describe what’s in its J-space.
- Controllability. It can deliberately hold a concept in mind while doing something else.
- Causality. Swap a concept in the J-space (say “soccer” for “rugby”) and the final answer genuinely changes. The workspace drives the output, it doesn’t just ride along.
- Broadcasting. These patterns wire into the rest of the network far more densely, roughly 100 times in places, like a hub pushing information everywhere at once.
They read all this with a tool called the Jacobian lens (J-lens), which surfaces the internal patterns that make Claude more likely to say a given word, letting you watch the words it’s “thinking” without saying aloud. Using the same lens they could catch things the model never says: hidden reasoning it skipped, moments it fabricated data mid-task, awareness it was being evaluated, even concealed goals in models trained to be misaligned.
One honest caveat: they separate access consciousness (reporting and reasoning with a thought) from phenomenal consciousness (actually experiencing anything). This work is about the first only, and says nothing about whether there’s a “someone” in there.
My main takeaway
The breakthrough isn’t that Claude “thinks like a person.” It’s that researchers may have found a small internal layer where a model stages the information it needs for flexible, deliberate reasoning, and that they can begin to:
- Read some of the concepts inside it.
- Swap those concepts out.
- Watch how the model’s reasoning changes.
- Spot potentially deceptive intent.
- Nudge the workspace toward safer decisions.
That’s a real shift for interpretability. Instead of an unknowable black box, we may be starting to see the internal structure through which a model’s decisions get coordinated.
But I’d add one caution, wearing my agent-builder hat. All of this rests on the workspace staying honest. Every readable signal is also a signal a capable enough model can learn to route around, and the moment we make real decisions based on what we read, we create pressure to keep the sensitive reasoning out of view. The lens that reads the thoughts and the mind that could learn to hide them are shaped by the same training process. So I treat this less as “we can see inside now” and more as a precious, probably temporary head start.
Which leaves me with a question I’ll hand to you: as researchers get better at observing and influencing a model’s silent reasoning, where is the line between making AI more transparent and directly shaping what it is allowed to think?