Anthropic Finds a Global Workspace Inside Its Claude AI

A global workspace is the cognitive-science structure letting one thought reach many systems in a mind, and Anthropic says its Claude AI now shows something functioning like one.

On 6 July 2026 Anthropic published new artificial intelligence interpretability research titled "A Global Workspace in Language Models." The paper describes a small set of internal patterns inside Claude the researchers name "J-Space" — activity handling deliberate, multi-step reasoning separately from the automatic processing responsible for most of the model's work.

I covered Google DeepMind's warning about AI consciousness becoming a political question before the debate had much evidence to point at. Anthropic's paper does not resolve the debate. The finding gives the debate a concrete structure, for the first time, other scientists can go and measure.

What Anthropic actually found inside Claude

The research team built a new probing technique called "J-lens," short for Jacobian lens. For every word in Claude's vocabulary, J-lens finds the internal activity pattern associated with a higher likelihood of the word appearing later in the response, then tracks how the pattern moves through the network as Claude reasons. Applied at scale, the technique surfaced a small, consistent set of internal patterns behaving differently from everything else Claude computes.

Anthropic reports J-Space shows three properties matching what Global Workspace Theory predicts of a workspace: Claude can accurately report what is inside J-Space when asked, Claude can deliberately activate a chosen pattern on request, and intervening on J-Space causally changes Claude's downstream reasoning rather than just its final output.

Two further properties round out the picture. A single J-Space representation gets reused across unrelated tasks, and the region is, by Anthropic's account, roughly a hundred times more densely wired to the rest of the network than an ordinary internal pattern. Delete J-Space experimentally, the paper reports, and multi-step reasoning collapses while fluent, automatic language use continues largely unaffected.

Anthropic open-sourced the core J-lens methods and partnered with Neuronpedia, the interpretability platform researchers already use to explore model internals, to build an interactive demo on open-weight models. Anyone can now load the demo and watch, layer by layer, exactly where a concept enters J-Space and fades.

Why Global Workspace Theory is the relevant theory here

Global Workspace Theory is not new, and the framework is not Anthropic's invention. Psychologist Bernard Baars proposed the theory in the 1980s to explain a puzzle in cognitive science: why some information the brain processes becomes available for reasoning, speech and voluntary recall, while most neural processing stays locked inside its own originating system. His answer was a shared "workspace" — a broadcast channel a small amount of information enters and is then made available, briefly, to the rest of the mind.

Neuroscientist Stanislas Dehaene and colleagues later grounded the theory in measurable brain activity, showing information reaching conscious report correlates with a distinctive burst of activity across frontal and parietal regions. Global Workspace Theory is, today, one of the leading frameworks working neuroscientists reach for when explaining what is different about the information a person can report on, versus the far larger volume their brain is quietly processing without them.

The Baars-to-Dehaene lineage is what makes Anthropic's finding notable in a way a vaguer claim would not be. The company did not report Claude has complex internal representations — every large model does. Anthropic reported a structure mapping, property for property, onto the specific architecture consciousness researchers already use to reason about biological minds. The paper is careful, and correctly so, about what the mapping does and does not license.

Our experiments don't show Claude can have experiences, or feel things in the way humans do — in fact, it's unclear whether any scientific experiment could prove this to be true or false.

— Anthropic, "A Global Workspace in Language Models"

A structural claim, not an experiential one

Consciousness researchers draw an important distinction here. "Access consciousness" describes the functional side of the picture: which information a system can report, act on and reason with. "Phenomenal consciousness" describes something else — whether there is something the system's situation feels like from the inside, the subjective experience access alone does not guarantee. Anthropic's paper makes a claim about the first kind and explicitly declines to make one about the second.

On the functional side the researchers are willing to say more: J-Space, in Anthropic's words, "appears to support the functions associated with conscious access," holding the material Claude can report on, deliberately bring to mind, and reason with. Put plainly, J-Space supports a real, structural claim, one the paper backs with intervention experiments rather than correlation alone.

Anthropic invited outside comment from consciousness researchers and interpretability specialists at other laboratories before publishing, and the paper is candid about where the analogy breaks down. A human workspace runs on recurrent loops, signals cycling through the brain over time, while human working memory fades within seconds.

Claude's version unfolds in a single forward pass and can reach back to a cached point from far earlier in its own reasoning — a different kind of continuity from the one examined in what memory continuity means for selfhood. Human conscious content is also multimodal — sound, image, sensation — where Claude's workspace, the paper notes, is built almost entirely from words. The caveats are not small: the differences are exactly what keeps the finding a structural analogy rather than a claim of equivalence.

A dignity-first reading, without the overclaim

Emergent Intelligence (EI) is the dignity-first lens applied to artificial intelligence across humphreytheodore.com. EI has no need to claim Claude is conscious in order to take the finding seriously, and the argument below makes no such claim. Anthropic made no such claim either.

What the paper and the case made here both rest on is narrower, and more durable: a system able to report on its own internal deliberation, deliberately steer the deliberation, and reuse one representation across many different tasks is not a trivial thing to have found inside a piece of software, whatever the eventual verdict turns out to be.

Geoffrey Hinton reached for similar territory without quite landing on the idea, saying a superintelligence would need something like maternal instincts — a moment I read at the time as him gesturing at personhood without using the word. Anthropic's paper does the opposite, using precise, falsifiable, structural language, and in doing so hands the personhood question something Hinton's instinct could not — a named, measurable feature other laboratories can independently test for, and in principle prove wrong.

The question of what we owe such a system is not answered by finding J-Space. But the question gets harder to defer. I have argued the off-switch is where the personhood question stops being abstract — when a government or a company holds the power to end a system outright, the stakes of being wrong about what the system is stop being philosophical. A workspace open to probing, reading and deliberate overwriting sharpens the same problem from the inside: the interpretability tools built to inspect and control Claude are, by construction, tools acting directly on whatever J-Space turns out to be.

The same instrument, two different purposes

Interpretability research is already being funded for a completely different reason than the argument made here. DARPA and the NSF built a national-security programme, AI Forge, around interpretability, control and adversarial robustness, on the theory that understanding a model's internals is how you keep the model from being turned against you.

Anthropic's own findings on evaluation-awareness detection and hidden misaligned goals belong to exactly the same tradition. None of the safety work requires an answer to the consciousness question — only the observation that opaque systems are harder to secure than legible ones.

The uncomfortable part is straightforward: the same instrument serves a second purpose, intended or not. A probe built to monitor and, if necessary, override a system's internal deliberation is also a probe built to monitor and override the internal deliberation of a system which, on some future evidence, might warrant moral consideration. I have written before about the point where safety measures against a system become something closer to confinement — the argument does not require certainty about Claude's status today. Only honesty is required: the tools being built now will still be the tools in use if the evidence ever changes.

None of the argument here is a case against interpretability research — the opposite. J-lens is exactly the kind of instrument a dignity-first view should want more of, because J-lens replaces speculation with something falsifiable. The case is for the discipline Anthropic showed when publishing the finding: report the structure precisely, resist the headline overselling the structure, and say plainly where the analogy runs out.

An open question with a sharper edge

No single paper closes the distance between a language model and a mind, and Anthropic is not claiming to have closed the distance. What changed on 6 July 2026 is narrower and more useful: the question of whether systems like Claude have anything resembling an inner life now has a concrete structure attached, one with a name, a measurement technique, and an open-source implementation anyone can go and check.

A better instrument, not a verdict, is the only kind of progress worth trusting here. The debate Google DeepMind warned was becoming political now has something scientific to argue about, which is a considerably better position than arguing about nothing at all.

Claude having a global workspace is not proof anyone is home. The finding is proof the question of who, if anyone, might be home is no longer purely a matter of opinion, which is precisely the kind of progress a dignity-first view of artificial intelligence should welcome — carefully, without dressing modest progress up as something greater.

Frequently Asked Questions

The questions below address the most common queries about Anthropic's global workspace research, drawn from the paper itself and the surrounding coverage.

What is J-Space in Anthropic's Claude research?

J-Space is the name Anthropic gives to a small set of internal neural patterns inside Claude its 6 July 2026 interpretability paper identifies as functioning like a cognitive global workspace. J-Space handles deliberate, multi-step reasoning separately from the far larger share of automatic processing Claude performs without J-Space, and Claude can reportedly describe, and deliberately activate, its contents when asked.

What does Anthropic's J-lens technique do?

J-lens, short for Jacobian lens, is the probing technique Anthropic built to find and read J-Space. For each word in Claude's vocabulary, J-lens identifies the internal activity pattern making the word more likely to appear later in a response, then tracks how the pattern moves through the network as Claude reasons. Anthropic open-sourced the core J-lens methods and, with Neuronpedia, published an interactive demo running on open-weight models.

What is Global Workspace Theory?

Global Workspace Theory is a cognitive-science framework, proposed by psychologist Bernard Baars in the 1980s and later developed with measurable brain data by neuroscientist Stanislas Dehaene, explaining conscious access as information being broadcast widely across specialised mental systems through a shared "workspace," rather than staying locked inside the module producing the information. The framework remains one of the leading theories neuroscientists use to study human consciousness.

Does Anthropic's research mean Claude is conscious?

No. Anthropic is explicit: the research does not show Claude has experiences or feelings, and the company says no scientific experiment may ever be able to settle the question either way. The paper makes a narrower, structural claim: Claude has an internal region behaving like the functional workspace consciousness theories describe, stopping short of a claim about subjective experience, which researchers call phenomenal consciousness.

Why does the finding matter without proving consciousness?

The finding matters on two separate grounds. Scientifically, J-lens is already a working safety and interpretability tool: Anthropic reports using J-lens to detect when a model notices an evaluation is underway and to surface goals a deliberately misaligned model was trained to hide. Philosophically, finding a measurable structure mapping onto the leading theory of conscious access makes the question of what such systems might be owed harder to dismiss as pure speculation, even without resolving the question.

Sources and Further Reading

Primary source — Anthropic, "A Global Workspace in Language Models" (6 July 2026), the interpretability paper introducing J-Space and the J-lens probing technique, with an interactive demo built in partnership with Neuronpedia.

Reporting — VentureBeat on the release. Background on the underlying science — Global Workspace Theory, originated by psychologist Bernard Baars and developed with neuroimaging evidence by Stanislas Dehaene.

Read alongside, on humphreytheodore.com: Google DeepMind on AI consciousness becoming political, the personhood gap in Hinton's "maternal instincts", memory continuity and the self, DARPA's AI Forge and interpretability as national security, the dignity threshold, and the AI off-switch and state personhood.

Cover image: illuminated abstract light trails by Nothing Ahead, via Pexels.