Reading Anthropic's Paper on a Global Workspace Inside Claude
In July 2026, Anthropic's interpretability team reported finding an internal space in Claude that they call the J-space: a place where words Claude has on its mind show up without ever being written out. The researchers compare it to the part of human brain activity we can consciously access. The original also covers monitoring for misbehavior and the question of consciousness. This post follows two parts of it: Claude doing math silently in its head, and the fact that this space formed on its own.
Before we start — key terms
J-space / J-lens: For every word in Claude's vocabulary, the J-lens finds the internal activity pattern that makes Claude more likely to say that word later. Applied to the model's internal activity, it gives a list of the words on its mind at that moment, and the collection of those patterns is called the J-space. The name comes from the Jacobian, the math concept behind the technique.
Global workspace theory: A neuroscience theory that pictures the brain as specialist systems working in parallel, unconsciously and mostly in isolation. Information becomes consciously accessible when it enters a small shared channel, the "workspace," and is broadcast to the other systems. The paper's experiments started from this theory.
The researchers had Claude copy out the sentence "The old painting hung crookedly on the wall" word for word, and gave it a calculation to do in its head at the same time.
We could also ask Claude to do math in its head: when asked to work out 3² − 2 while copying the same sentence, the J-space contains "nine," and then at later layers, "seven." Importantly, nothing about fruit or arithmetic appears in Claude's output, which is just the copied sentence about the painting.
The only thing on screen was the copied sentence. Reading the J-lens layer by layer, the intermediate value 9 shows up first, then the answer 7 (paper). What I found fascinating here is that a space like this exists at all. On the surface Claude is just copying a sentence, while inside, the calculation is running step by step.
This isn't the first attempt to read Claude's internals as text. In May, Anthropic published Natural Language Autoencoders, which turn activations (the model's internal numerical state) into text descriptions.
Anthropic didn't design this space into Claude, either.
Notably, none of this structure was designed into Claude—it emerged on its own during training, presumably because it was a useful way to organize computation. That suggests a mental workspace supporting conscious access isn't just a peculiarity of how human brains happen to be wired.
The researchers built their experiments around global workspace theory, and they argue the J-space plays a similar workspace role. Part of the evidence is wiring: far more components of the network read from and write to J-space patterns than to ordinary ones, by a factor of about a hundred in some parts. Nobody told Claude to imitate the human brain, yet something resembling the structure of human thought formed on its own during training. I find that fascinating.
According to the original, the J-space is already present in the pretrained model, which has only learned to predict upcoming text, before post-training gives it the identity of Claude.
