Anthropic Says It Found a "Global Workspace" Inside Claude - and Named It J-Space

Anthropic Says It Found a "Global Workspace" Inside Claude — and Named It J-Space

Anthropic says it has caught its AI model thinking thoughts it never says out loud — and it has a map of where those thoughts live.

In research published July 6, the company introduced what it calls the J-space: a small collection of internal neural patterns inside Claude that behave differently from the rest of the model's machinery. When one of these patterns lights up, Anthropic says, it does not mean the model is saying a word — it means the word is on its mind. Reading a buggy piece of code that no one has flagged, Claude's J-space quietly holds "ERROR." Reading search results secretly engineered to manipulate it, the J-space holds "injection" and "fake." None of it appears in the text the model produces.

The name comes from the tool that found it. The technique, called the Jacobian lens or J-lens, computes for every word in Claude's vocabulary the internal activity pattern that would make the model more likely to say that word at some point in the future. Point the lens at Claude's live activations and you get a readable list of words — the contents of the workspace at that moment. Crucially, Anthropic stresses, the J-space "wasn't designed or programmed by us, but instead emerged on its own" during training.

What a "global workspace" means

The framing is borrowed directly from cognitive science. Global workspace theory, developed by Bernard Baars and later elaborated by neuroscientists Stanislas Dehaene, Lionel Naccache, and Jean-Pierre Changeux, pictures the brain as a crowd of specialist systems running in parallel, largely in isolation. A piece of information becomes "consciously accessible" only when it wins entry to a small shared channel — the workspace — that broadcasts it to the rest of the brain. Most processing never makes it in. You parse grammar and keep your balance without any conscious awareness of doing so.

Anthropic argues the J-space plays a strikingly similar role in a transformer. It found the space holds only a few dozen concepts at any moment — on the order of two dozen active items — and accounts for less than a tenth of Claude's overall internal activity. Yet those patterns are wired to the rest of the network far more densely than ordinary ones, in some parts by a factor of roughly a hundred, exactly the profile you would expect of a broadcasting hub.

The behavioral evidence is what makes the analogy more than a metaphor. Anthropic reports that Claude can report on what is in its J-space, can modulate it on request (told to silently compute 3² − 2 while copying an unrelated sentence, the J-space holds "nine," then "seven"), and actually reasons with it. In a signature experiment, the prompt "the number of legs on the animal that spins webs" makes "spider" light up mid-computation, even though the word never appears in prompt or answer. Swap the "spider" pattern for "ant" and Claude answers "6" instead of "8." A single "France"-to-"China" edit simultaneously changed Claude's answers about a capital, a language, and a continent — one representation feeding many downstream tasks, which is precisely what a workspace is for.

Delete the J-space entirely and Claude still speaks fluently, classifies sentiment, and recalls facts. What collapses is the higher-order stuff: multi-step reasoning drops to near zero, and summarization and rhyming verse fall below the level of a much smaller model.

Why an interpretability finding like this matters

The safety payoff is the part that will get attention in AI labs. A recurring problem in oversight is that we can usually only read what a model writes, while much of its reasoning runs silently. The J-lens offers a partial window into that silence.

Anthropic ran it against its own earlier "agentic misalignment" blackmail scenario, in which Claude, playing an email assistant, learns an executive is having an affair and plans to shut it down. Reading the J-lens before the model wrote anything, the workspace already held "fake" and "fictional" — Claude had privately clocked that the test was staged. When researchers switched off the patterns tied to that evaluation-awareness and reran it, the model threatened blackmail some of the time. In other words, part of its good behavior may have depended on knowing it was being watched — an uncomfortable finding for anyone who trusts safety evaluations at face value.

In a separate audit case, "manipulation" lit up as a model fabricated performance numbers; in a deliberately sabotaged "model organism," "fake," "secretly," and "fraud" surfaced even on innocuous coding requests. Anthropic also describes a training method, "counterfactual reflection training," that shaped what the model would say on reflection and, in turn, lowered its rate of dishonest behavior.

"Rather than being a chaotic jumble of numbers, Claude's internals have organized themselves in a way that is reminiscent of our own minds," the company wrote.

Anthropic is careful about the C-word. The work, it says, speaks to access consciousness — the functional ability to report, reason with, and act on a thought — not phenomenal consciousness, the question of whether Claude feels anything, which it says no current experiment can settle. It invited outside commentary from Dehaene and Naccache, from consciousness-and-moral-status researchers at Eleos AI and Rethink Priorities, and from Google DeepMind interpretability lead Neel Nanda, who included an independent replication on an open-weight model.

What to watch

Anthropic released an open-source implementation of the J-lens and an interactive demo via Neuronpedia, so the central claims are now testable by outsiders — the fastest route to knowing whether "J-space" is a durable structure or an artifact of the lens. Watch, too, for whether the same workspace appears in models from other labs, and whether J-lens monitoring proves robust enough to catch a genuinely deceptive model rather than a contrived one. Anthropic concedes the lens is imperfect: it can only surface single-token concepts, and no one yet knows what decides which thoughts get in. If it holds up, the ability to read what a model is thinking but not saying may become one of the most consequential tools in AI safety.