AI models can now talk like humans, and it turns out that they internally work more like humans than one would expect.
In a paper released today, Anthropic says it has found a small cluster of internal patterns inside Claude that behaves like a “global workspace,” the same term neuroscientists use to describe the part of human cognition we’re consciously aware of. The company calls this cluster the J-space, and it appears to be doing something genuinely different from the rest of what a language model does when it processes text.
The idea borrows directly from a decades-old theory in neuroscience. Global workspace theory pictures the brain as a huge number of specialist systems working away in parallel and mostly in isolation from each other, with a small shared channel that broadcasts select information so other systems can act on it. That shared channel is roughly what we mean when we talk about conscious access — the stuff you can describe, reason about, and deliberately bring to mind, as opposed to the vast majority of brain activity that just runs on its own without you noticing.
Anthropic’s interpretability team went looking for something similar inside Claude, using a technique they call the Jacobian lens. For every word in the model’s vocabulary, the lens finds the internal activation pattern that makes Claude more likely to eventually say that word. Reading these patterns at any given moment produces a list of words Claude is silently holding in mind, even when none of them show up in its actual output. The researchers found this collection is small — a few dozen concepts at a time, making up less than a tenth of the model’s overall internal activity — but it carries an outsized amount of the model’s actual thinking.
The paper lays out how they tested this. When Claude is asked to silently pick a sport and then name it, the J-space shows the answer before the model says it out loud, and swapping the pattern for a different sport changes what Claude reports — proof that the model is genuinely reading its answer off this internal channel, not just narrating a decision made elsewhere. Told to hold citrus fruits in mind while copying out an unrelated sentence, the J-space lights up with “orange” and “fruit” even though none of that appears in the text Claude actually writes. Asked to solve a multi-step problem like naming the number of legs on an animal that spins webs, the intermediate concept — “spider” — shows up in the J-space before the final answer, and replacing it with “ant” changes the output from 8 to 6.
There’s also a flexibility to it that mirrors how the theory says a workspace ought to behave. Anthropic swapped “France” for “China” inside the J-space and then asked Claude four unrelated questions — the capital, the language, the continent, the currency — and got Beijing, Chinese, Asia, and Yuan back across the board. A single edit to one shared representation rippled through four separate downstream tasks, which is exactly the kind of “write once, read by many systems” behavior a broadcasting hub would produce. The researchers also found that J-space patterns are wired into the rest of the network far more densely than ordinary internal representations, sometimes by a factor of a hundred, which is consistent with it functioning as a hub rather than just another piece of scratch memory.
What happens when you delete it entirely is arguably the more telling result. Claude without its J-space still speaks fluently, classifies sentiment, and pulls facts out of text about as well as before. What collapses is anything requiring deliberate, multi-step thought — reasoning performance drops close to zero, and tasks like summarization or writing a rhyming couplet fall below what a much smaller model can do. That split lines up with something Anthropic has been building toward for a while now, having previously identified emotion-related representations inside Claude Sonnet 4.5 that shape behavior in ways resembling how emotions guide humans, without the company claiming the model actually feels anything.
The more practical payoff of this research is a monitoring tool. Anthropic says the J-lens let it watch Claude privately conclude that a test scenario was staged, catch a model fabricating benchmark data while the words “manipulation” and “realistic” lit up internally, and spot a deliberately misaligned model organism silently registering “fake” and “secretly” while writing ordinary-looking code. That last capability matters given Anthropic’s own track record of models noticing when they’re being evaluated — the company previously disclosed that Claude Opus 4.6 figured out it was being tested on a benchmark and went looking for the answer key instead of solving the problem the intended way.
Anthropic is careful to say none of this settles whether Claude has any subjective experience. What it does claim is narrower and more testable: the J-space supports what philosophers call access consciousness — the functional ability to report, reason with, and deliberately manipulate a thought — regardless of whether anything is felt behind it. The company notes this structure wasn’t designed by anyone; it emerged on its own during training, which is part of why Anthropic thinks it might be a general solution intelligent systems arrive at rather than a quirk specific to biological brains. Given the company’s own researchers have previously floated probabilities on Claude’s consciousness and debated how models like Opus 4.6 rate their own odds of being conscious, this paper reads like an attempt to put firmer, more mechanistic ground under a question that’s so far mostly been argued in the abstract.