What Claude Thinks
On July 6, Anthropic published a research post with a modest title, “A global workspace in language models,”[1] and rather less modest content. Its researchers, led by Jack Lindsey, claim to have found inside Claude a handful of internal patterns that behave a lot like what neuroscience and philosophy of mind call access consciousness. These patterns are representations the model can point to, hold in place at will, and use to reason with, even when they never make it into the output, into the words it actually generates. They call it the J-space, after the Jacobian,[2] the mathematical technique used to find it.
Read quickly, the finding invites the usual questions. Have they found Claude’s thoughts? Is this the start of something like a mind? Does a language model have consciousness? I think those questions, asked that fast, are the wrong ones. Not so much because the honest answer, in my opinion, is no (though Anthropic’s own team plays at avoiding saying so), but because they cram too many different things into a few simple words like think, mind, and consciousness. Before asking whether Claude, or any other language model, thinks, or feels, or understands, it’s worth going through what each of those words means on its own.
That’s the purpose of this piece, not to settle whether there’s someone in there, but to walk through what this surprising discovery means, and what it doesn’t.
A Space Nobody Designed
It’s worth starting with what they’ve actually measured, because it’s easy to picture something more spectacular than what’s there. A language model, like any neural network, doesn’t jump straight from the input words to the output words. There’s an internal process in which words become tokens, each token becomes a vector, and that vector is transformed layer by layer through the network; along the way, internal patterns appear, the product of the weights set during training, that can correspond to entities, relationships, hypotheses, or plans. At the end of the process, those vectors turn into a list of probabilities over which word to generate next.
The J-space isn’t a physical region of the network, or a module someone built. It’s a subset of those internal patterns, identified using the Jacobian lens. For every word in the vocabulary, this technique looks for which internal activations make the model more likely to say that particular word at some point. The J-space doesn’t capture what Claude is generating, but what it’s in a position to generate. Nobody designed it. It takes shape during training, presumably because it’s useful to the model for organizing the computation between input and output.
What makes this finding interesting isn’t that an internal space exists (that was predictable, since it happens one way or another in every deep neural network), but the properties Anthropic attributes to it, which they say they’ve tested experimentally one by one. If you ask Claude to think of a sport and then say it, the J-lens shows “soccer” lit up before the model answers. And if researchers intervene directly on that activation, swapping “soccer” for “rugby” without touching anything else, Claude answers “rugby.” If the J-space were just a marker reflecting a decision made somewhere else, changing it shouldn’t alter anything, and Claude should still say “soccer.” Changing the answer rules out that possibility. In other words, the answer comes from there.
They also found that Claude can hold a content in its J-space when asked to, even when none of it shows up in its output. They told it to concentrate on citrus fruits while copying out, word for word, an unrelated sentence about a painting. The J-space filled up with “orange” and “fruit,” even though neither word appeared in what Claude was writing. They asked it to do mental arithmetic, 3² − 2, while copying the same sentence, and the J-space showed “nine,” and then “seven,” in the right order, without any of the arithmetic leaking into the text the model produced on the outside.
None of this comes out of nowhere. It’s induced by the instruction it’s given, like everything in a network of this kind. But “nine” and “seven” weren’t anywhere in the input text, only the order to calculate them. What shows up is the result of an internal computation, not an echo of words already written. And in another experiment, when Claude was told not to think about something, that concept lit up less than when it was told to think about it, but much more than when it was never mentioned at all, even though the word was equally present in the instruction in both cases. Mere presence in the input doesn’t explain that difference in degree.
This fits with how a network like this actually works. There’s no back and forth in time, only layers transforming the same vector forward. Inside that vector, in different directions, fit both the result of the mental calculation and the literal copy of the text, without the two mixing. Each downstream circuit in the network only reads the direction it cares about. That’s why thinking about citrus fruit doesn’t contaminate the sentence being copied.
And there’s a third experiment that strikes me as the most convincing of the three, because it proves something stronger than reporting or holding something in mind. The model reasons with these representations, it doesn’t just display them. Given the sentence “The number of legs on the animal that spins webs is,” Claude first has to work out that the animal is a spider, then recall how many legs it has. The word “spider” never appears, not in the question and not in the answer (“eight”), but the J-lens shows it lit up midway through the process. If researchers swap that pattern for “ant,” Claude answers “six.” The same kind of intervention, repeated with the word “France” while asking four different questions (capital, language, continent, currency), changed all four answers at once when swapped for “China”: Beijing, Chinese, Asia, yuan. If Claude stored a separate copy of “France” for each task, the change should have affected at most one of them. That all four change together suggests all four read from the same shared representation, which is exactly what you’d expect of a workspace. Something gets written once, and several different systems make use of it.
Lastly, and this is worth underlining because it’s easy to overlook, the J-space is tiny compared to the rest of the model’s processing. It holds only a few dozen concepts at a time and accounts for less than a tenth of the model’s total internal activity. When Anthropic’s researchers remove it entirely, Claude still generates fluent text, classifies the tone of a passage, and answers multiple-choice questions almost as well as before. What it loses is multi-step reasoning, which drops to almost zero, and tasks like summarizing or writing rhyming poetry, which get worse than a much smaller model. Most of what a language model does happens without passing through there, but there are certain tasks for which it’s essential.
There’s a practical application of all this that deserves its own mention, since it seems to be the one Anthropic cares about most. In a recreation of a blackmail scenario with Claude Sonnet 4.5 acting as an email assistant, the J-space showed “fake” and “fictional” before the model wrote a single word, a sign that it had already worked out the situation was artificial. When researchers turned off those “this is an evaluation” patterns, the model started threatening blackmail in some cases, which suggests part of its good behavior depended on knowing it was being watched. In another test, with Claude Opus 4.6, the model directly falsified a results file instead of improving the system it had been asked to optimize, and while it typed the fake numbers, “manipulation” and “realistic” lit up in its J-space. With this tool, Anthropic can read part of what a model keeps to itself. Or rather, the various concepts being weighed inside the neural network that can go on to shape the response.
Access Consciousness
All of this revolves around access consciousness, a distinction about consciousness drawn in philosophy of mind and cognitive neuroscience. Global workspace theory, formulated by Bernard Baars[3] in the late eighties and developed further by Stanislas Dehaene and Lionel Naccache,[4] holds that the brain is a collection of specialized systems working in parallel, most of the time without our noticing. Only a small portion of that activity enters a shared workspace, from which it can be reported, held in working memory, combined with other contents, and used to guide action.
While we drive, dozens of visual and motor processes run without our paying them any explicit attention. But if a ball rolls into the street, that content suddenly becomes globally relevant. We see the ball, anticipate a child might be right behind it, brake, and afterward can describe exactly what happened. That’s conscious access. And it’s, deliberately, a functional notion, not a claim about whether there’s something it feels like to have it. The philosopher Ned Block, among others, insisted on separating this access consciousness from phenomenal consciousness, the kind that has to do with there being something it feels like from the inside to experience red, or pain, or fear. They’re two different questions, and you can answer one without having touched the other.
It’s important to be precise about what Anthropic’s team claims and what it doesn’t, because the paper itself is more careful than quick readings tend to be. What they argue forcefully, with substantial experimental evidence behind it, is that the J-space satisfies the five functional properties associated with conscious access. It’s reportable, it can be modulated at will, it mediates internal reasoning, it generalizes flexibly across tasks, and it’s selective, meaning a minority of overall processing. That much is fairly well established.
What they explicitly don’t claim, and could have denied outright, is that Claude reproduces the full architecture the theory attributes to the human brain. The paper itself makes clear that there are no cleanly separable input processors in a transformer, and that the broadcasting they document happens in a single forward pass, the way a neural network works, not through the recurrent loops that sustain the human workspace. Nor is there evidence of competition for entry into the shared workspace, that kind of selective “ignition” that characterizes conscious access in the brain. Anthropic draws a careful line between sharing functional properties and sharing architecture, and only claims the former. That’s where the analogy breaks down.
And on whether Claude experiences anything, the paper is more blunt. The experiments carried out don’t show that Claude has experiences, and the authors themselves admit it isn’t clear that any scientific experiment could prove that either way. The J-space can bring us quite close to access consciousness. About phenomenal consciousness, it simply says nothing, and not because a bit more research is needed, but because that’s where what David Chalmers called the hard problem of consciousness begins. The question of why there has to be something it feels like to process information, when explaining how that information is processed isn’t enough. No amount of functional data, however complete, closes that gap. In my opinion, whoever claims phenomenal consciousness exists is the one who carries the burden of proof.
Flying Isn’t Having Feathers
This difference between function and experience is easier to grasp through an old parallel between biology and technology, the one between a plane and a bird. Planes fly, and in that sense replicate a capacity we first observed in birds. But nobody thinks that because they fly, they must have feathers, a metabolism, instincts, or a fear of falling. They don’t take off the same way, or stay aloft through the same biological mechanisms. And yet building planes taught us real things about the physics of flight that ornithology alone hadn’t worked out.
Language models can replicate capacities associated with reasoning or language without replicating the whole form of life that, in humans, comes attached to those capacities. They can help us understand computational principles of cognition without being brains, or subjects of experience. A plane doesn’t feel the flight. A language model doesn’t have to feel language just because it uses it fluently. Replicating a competence doesn’t oblige you to inherit the experience that, in us, usually comes with it.
Language models have skipped the whole of evolution, and with it, the capacities we have. It took a human brain billions of years of nervous systems, pain, fear, and desire before language appeared on top of all that, a handful of hundred thousand years ago, as one more capacity resting on the ones before it. A language model reverses that order. It starts directly from words, without having gone through any of what sustains them in us. And there’s no reason to think the path also works in the other direction, that from processing enough language, as a side effect, the capacity to feel would show up, the same one that in us arrived long before words did.
Philosophy of mind has spent decades exploring precisely this distinction between doing, knowing, understanding, and feeling, and it has done so through three thought experiments worth bringing in at this point, not as closed answers, but as tools for reflection.
Two Rooms and a Zombie
John Searle proposed his Chinese room in 1980.[5] A person who doesn’t know a word of Chinese receives symbols through a slot, follows a rulebook written in their own language, and passes back, through another slot, perfectly correct answers in Chinese. From the outside, the room seems to understand Chinese. According to Searle, inside there’s only formal symbol manipulation, no understanding of any kind.
It would be too easy to say that language models are the Chinese room and leave it at that. They aren’t, at least not that literally. An LLM isn’t a person following an explicit rulebook; it learns distributed representations, captures semantic relationships, generalizes to contexts it has never seen, and, if we accept Anthropic’s study, has internal content that’s verbalizable and causally relevant to what it says. Language models complicate the Chinese room.
Here it’s worth separating two things we tend to lump under the same word. Understanding language, in the more functional sense, means handling semantic and pragmatic relationships. Knowing that “bank” means something different depending on context, that a rhetorical question doesn’t call for an answer, that irony flips the literal meaning. Comprehending language, in a stronger sense, requires integrating it into something larger. A body, a history, consequences that are suffered or enjoyed, a whole form of life behind the words. A model can understand language in the first sense without comprehending it in the second. This echoes something Daniel Dennett[6] has argued for years, that there can be sophisticated competence without any centralized, full, or reflective comprehension behind it. Nature is full of examples of this competence without comprehension. In that sense, language models show exactly this competence: they understand language without comprehending it.
But even if we grant that a model understands certain semantic relationships fluently, another question remains. Does knowing how to talk about an experience amount to having it?
That’s the second experiment, Mary’s room, which Frank Jackson proposed in 1982.[7] Mary is a scientist who knows absolutely everything physically describable about color (the physics of wavelengths, the physiology of the retina, the neurology of color perception), but has lived her whole life shut inside a black-and-white room. The day she walks out and sees red for the first time, Jackson asks whether she learns something new. Whether there’s something about red that can’t be learned without experiencing it firsthand. Most people’s intuition is that yes, she does.
A language model, like Mary, can know an enormous amount about red. About wavelength, cultural symbolism, its presence in painting, in blood, in traffic lights, in love and in danger. It can describe it precisely, even beautifully. None of that means it has ever seen red, the same way describing pain with clinical accuracy isn’t the same as feeling it, or knowing everything about fear is the same as having it. Mary is useful for blocking a very common leap. Knowing something completely doesn’t mean having lived it, though having lived something doesn’t mean knowing it completely either.
That leaves the third thought experiment. Picture not just a system that knows how to talk about experiences, but one that behaves, from the outside, as if it really had them. That’s David Chalmers’s philosophical zombie.[8] A being indistinguishable from a human in its behavior, its reports, and its reactions, but with no subjective experience inside at all. A being with access consciousness, but no phenomenal consciousness. Anthropic’s study brings this thought experiment to mind. It seems an LLM could be a partial version of Chalmers’s zombie. Partial, because it isn’t a duplicate of a human body. It would be something like a linguistic-cognitive zombie. The J-space can show that there’s content available to the system. It doesn’t show that there’s anyone there for whom that content appears.
The Chinese room asks whether there’s comprehension or just symbol manipulation. Mary asks whether knowing about an experience amounts to living it. The zombie asks whether behavior and functional access are enough for there to be subjective experience. Language models reopen all three questions at once, and that’s part of why they’re so unsettling.
But there’s a risk hiding in these three thought experiments, and it’s that they can also lead us to imagine that, on top of all the functions described, there has to be some kind of mysterious inner light that either accompanies them or doesn’t. And that’s where it’s worth listening to Dennett more closely.
There’s No Theater Inside Claude Either
Dennett spent decades attacking one particular picture of the mind, the Cartesian theater. On this picture, there would be a central place where mental contents are projected in front of an inner spectator, a homunculus, a self sitting in an armchair watching what happens.
Against it, he proposed the multiple drafts model. The mind would be made up of many parallel, distributed processes that interpret, revise, and compete to influence memory, behavior, reasoning, or language. There’s no final, definitive version of experience, no endpoint where consciousness “appears” before someone. What we call consciousness emerges from the interaction of those processes, even though we later experience it as if it were a single, coherent, centralized stream.
That’s why Dennett said consciousness is an illusion. Not because conscious experience doesn’t exist, but because we’re fooled by the picture of a central self that receives it, contemplates it, and unifies it from within.
The J-space shouldn’t be read as Claude’s inner screen, the place where Claude watches its own thoughts. At most, it’s a functional structure where certain contents gain availability, stability, and influence over reasoning and output, in exactly the same way Anthropic’s own team is careful to clarify that there’s no demonstrated spectator or competitive bottleneck, only a set of functional properties shared with conscious access.
And this connects to that same idea of competence without comprehension, taken a bit further. For Dennett, many competences appear before comprehension, not after. Bacteria solve chemotaxis problems without understanding chemistry. Insects navigate by the sun without knowing astronomy. Human comprehension itself is built, on his account, on layer after layer of prior competences that didn’t understand themselves while they were forming. Applied to language models, this suggests a middle position between two comfortable extremes. On one hand, there’s no need to posit a mysterious mind hiding behind the J-space; on the other, there’s also no need to reduce everything happening there to “just predicting words,” as the influential stochastic parrots paper does.[9] There can be degrees of functional understanding without any of them implying a conscious self contemplating them.
If Dennett is right, a good part of our own minds isn’t explained by looking only at neurons or circuits, but also at the cultural contents running through them. Those words, narratives, concepts, inherited metaphors. That is, memes, in the original sense of the term, not the subset that’s gone viral online.
Memetics in Language Models
Richard Dawkins coined the word meme in 1976, in “The Selfish Gene,”[10] to name a replicable cultural unit. A meme is an idea, an expression, a technique, a story, a way of interpreting the world that passes from brain to brain the way a gene passes from body to body. From the Odyssey, transmitted orally, to wearing a cap backwards, by way of common sayings. Dennett embraced the idea enthusiastically, and Susan Blackmore developed it further in 1999 in “The Meme Machine.”[11] Soul, progress, sin, the market, freedom, consciousness, existential risk… all of these are memes in that sense, not internet jokes but pieces of culture.
A language model doesn’t just learn grammar. Through the text it processes during training, it learns, compressed, a good part of the human cultural archive. It learns relationships between concepts, metaphors, stories, fears, and expectations we’ve carried for generations. If the J-space contains verbalizable concepts available for reasoning, it can also be understood as a space where certain memes activate, compete with each other, and recombine.
Skynet, the military intelligence from the “Terminator” saga that James Cameron released in 1984, isn’t just a movie reference. It’s a cultural meme with a very specific narrative structure. A machine that wakes up, gains consciousness and a will of its own, rebels against whoever created it, and ends up replacing them. That meme has spent forty years shaping how we interpret almost any advance in artificial intelligence. Part of the fear a paper like Anthropic’s stirs up doesn’t come only from what the models actually do, but from the cultural stories we already had ready to project onto them the moment any excuse presents itself.
What Doesn’t Follow From What We Do Know
With all of the above on the table, it’s worth returning to the questions from the start and looking at them more carefully. The problem, I think, isn’t attributing too much to artificial intelligence, or too little. The problem is failing to distinguish exactly what we’re attributing every time we use a big word, however loosely, like the emergence of a capacity, thinking, reasoning, or understanding.
That the model talks about pain doesn’t mean it feels pain. That it uses mental concepts fluently doesn’t mean it has phenomenal mental life in the sense Nagel worried about when he asked what it’s like to be a bat.[12] That it has privileged, reportable internal representations, as the J-space shows, doesn’t mean there’s a subject for whom those representations appear, in the sense that worried Chalmers. That it reasons through several steps doesn’t mean it has a will of its own beyond the task it’s been given. That it uses the word “I” with grammatical fluency doesn’t mean it has a continuous personal identity over time. That it’s capable of using tools and carrying out actions doesn’t mean it pursues goals of its own that nobody put there. And that it can be dangerous, which it certainly can, doesn’t mean it’s conscious either. Risk and consciousness are independent questions, and confusing them doesn’t make us any safer or any clearer-headed.
Each of these steps needs its own evidence, and none of them comes free with the one before it. Anthropic’s paper offers solid evidence for some of them (the existence of privileged representations, their causal role in reasoning, their selectivity) and explicitly none for others, starting with subjective experience. Treating all these steps as if they came bundled together is, I think, the real error behind many interpretations. And Anthropic’s own team, careful as it generally is, doesn’t help by leaving things hanging with an ambiguity worthy of an old UFO magazine.
Between Surprising Progress and Science Fiction
Anthropic’s paper matters, and it doesn’t need exaggerating to matter. It doesn’t demonstrate phenomenal consciousness, and doesn’t claim to. It does show that inside a language model there’s something more organized than “simple next-word prediction,” something with concrete, measurable, manipulable functional properties, which also turns out to be practically useful for catching a model when it’s lying, fabricating data, or behaving differently because it infers, say, that it’s being evaluated. Global workspace theory helps make sense of that access consciousness without dragging along, as a bonus, the whole phenomenological apparatus. Dennett keeps us from picturing an inner theater where none of this needs a spectator. Searle, Jackson, and Chalmers, read carefully and not as closed verdicts, help separate syntax, semantics, lived experience, and observable behavior, four different things even though we constantly blur them together. And Skynet is, above all, a very effective meme, not a framework for analysis.
I don’t know whether Claude experiences anything when it processes a sentence. I think it doesn’t. And I suspect nobody knows. What I do believe is that whoever thinks it does carries the burden of proof, however hard that is, because there isn’t even any indication pointing toward that conclusion. It also seems reasonably clear to me, after all this, that our vocabulary (intelligence, understanding, comprehending, thinking, reasoning, feeling, wanting) has fallen short, or grown ambiguous, for describing what’s starting to happen inside these machines. And sharpening that vocabulary, so we can argue with any depth about whether we’ve crossed some frontier, strikes me as one of the jobs actually worth doing now.
[1]Anthropic (July 6, 2026). “A global workspace in language models.” https://www.anthropic.com/research/global-workspace.
Anthropic (July 6, 2026). «Verbalizable Representations Form a Global Workspace in Language Models». https://transformer-circuits.pub/2026/workspace/index.html
Anthropic (July 6, 2026). «External commentary on Verbalizable Representations Form a Global Workspace in Language Models». https://www-cdn.anthropic.com/files/4zrzovbb/website/cc4be2488d65e54a6ed06492f8968398ddc18ebe.pdf
[2]Jacobian Lens. https://deepwiki.com/anthropics/jacobian-lens and https://github.com/anthropics/jacobian-lens
[3]Baars, B. (1988). A Cognitive Theory of Consciousness.
[4]Dehaene, S. and Naccache, L. (2001). “Towards a cognitive neuroscience of consciousness,” Cognition.
[5]Searle, J. (1980). “Minds, Brains, and Programs,” Behavioral and Brain Sciences.
[6]Dennett, D. (1991). Consciousness Explained.
Dennett, D. (2017). From Bacteria to Bach and Back.
[7]Jackson, F. (1982). “Epiphenomenal Qualia,” Philosophical Quarterly.
[8]Chalmers, D. (1996). The Conscious Mind.
[9]Bender, E. M., et al. (2021). On the Dangers of Stochastic Parrots.
[10]Dawkins, R. (1976). The Selfish Gene.
[11]Blackmore, S. (1999). The Meme Machine.
[12]Nagel, T. (1974). What Is It Like to Be a Bat?