Anthropic found a “global workspace” inside Claude. The more interesting question is what it reveals about us.
The Ocean of Possibilities — Image by Craiyon.com
If the mind is an ocean, we spend our lives at the surface. Beneath the waterline, an enormous amount of processing carries on without our knowledge: the visual system resolving a face from a scramble of light and shadow, the motor system holding the body upright, the language system parsing grammar before we have consciously registered a single word. Only a thin sliver of all this activity ever breaks the surface into awareness. Yet it is that sliver we rely on when we reason deliberately — planning a route through a city, working out why an engine won’t turn over, deciding what to say next. We can put those thoughts into words. We can hold them in mind. We can turn them, at will, toward whatever problem the moment presents.
On July 6, 2026, a team of sixteen researchers at Anthropic published a paper arguing that something with a strikingly similar shape has emerged, uninvited, inside a large language model. The paper is called Verbalizable Representations Form a Global Workspace in Language Models, and its central claim is that models like Claude maintain a small, privileged set of internal representations — a space they can report on, hold in mind, reason with, and route flexibly to different tasks — floating atop a far larger volume of automatic processing they cannot access at all. The researchers call this region the J-space, and they reach it with a new instrument they call the Jacobian lens.
It is one of the more consequential results in mechanistic interpretability, the young discipline devoted to reverse-engineering what actually happens inside these networks. And it has set off a familiar argument, conducted partly in public and partly in the quieter corners of the field, over whether this is evidence of something like consciousness. That argument is worth having, but it is easy to have it badly. The most rewarding path runs through three questions in order: what the J-space actually is, what it teaches us about how complex networks organize themselves, and — most intriguingly — what happens when the metaphor is turned around and pointed back at the biological brain that inspired it.
A better lens
To understand the J-space, it helps to understand the tool that finds it, because the tool is where a lot of loose reporting goes wrong.
A transformer-based language model processes text as a sequence of positions. At each position it maintains a long vector called the residual stream, a kind of shared scratchpad that every layer reads from and writes to. At the first layer that vector encodes little more than the identity of the current word. By the final layer it has been transformed into something from which the model’s next-word prediction can be read directly. Everything interesting happens in between, in the intermediate layers, and for years those layers were difficult to interpret. The best-known tool for peeking at them, the “logit lens,” simply applies the model’s final output machinery to an intermediate layer and reads off which words it seems to point toward. It works reasonably well near the end of the network and degrades badly in the earlier and middle layers, where it tends to return noise.
The Jacobian lens is a principled correction. Rather than assuming an intermediate representation uses the same coordinates as the final layer, it measures something more causal: for each layer, it computes the average effect that a small change in the residual stream has on the words the model eventually produces — not just the very next word, but words it may produce further downstream — averaged across roughly a thousand different contexts. That averaging step is the crux. It separates representations the model is genuinely poised to verbalize, should the occasion arise, from those that merely happen to surface in one particular sentence. The result, at each layer, is a ranked list of vocabulary words describing what that internal state is “about.”
The set of directions the lens picks out, taken together, forms the J-space. It is not the whole residual stream. On the contrary, when the researchers measured how much of the model’s internal activity actually lives in the J-space, they found it accounts for a strikingly small share — never more than about a tenth of the variance in the model’s activations, and typically far less. The vast majority of what the network is doing at any moment lies outside this privileged channel. Which is precisely what makes it interesting. The J-space is not the model’s mind. It is the model’s spotlight.
Watching a thought before it is spoken
What makes the paper vivid rather than merely technical is that the researchers could not only read the J-space but intervene on it, and watch the model’s behavior bend in response. A handful of their experiments are worth walking through, because they are the concrete ground beneath every grander claim.
In the simplest, they asked Claude to think of a sport and then name it in a single word. Applying the lens to the moment just before the answer, they saw the word “Soccer” rising to the top of the readout — and, sure enough, the model said “Soccer.” Then they performed a swap. Leaving the rest of the internal state untouched, they subtracted the direction corresponding to “Soccer” and added an equal measure of “Rugby,” a word that had not been anywhere near the model’s likely outputs. The model answered “Rugby.” Its spoken report tracked the altered contents of its workspace, not the other way around.
A second experiment showed the J-space carrying genuine reasoning rather than mere labels. The prompt was a small riddle: “The number of legs on the animal that spins webs is.” To answer, the model has to first infer, silently, that the animal is a spider, then recall that spiders have eight legs. The word “spider” never appears in the prompt or the answer — but it appears in the lens, at intermediate layers, exactly where the inference happens. When the researchers swapped the “spider” direction for “ant,” the model’s output changed from “8” to “6.” The intervention landed on an unspoken intermediate thought and redirected the conclusion that flowed from it.
The same machinery revealed something that looks a great deal like planning. Given the opening line of a couplet — “The soldier marched into the night” — the lens showed the word “fight” already present at the start of the second line, before the model had written most of it, as a planned rhyme. When they swapped “fight” for “light,” the model didn’t just change its final word; it changed its earlier word choices to accommodate the new ending, completing the line with “morning light” instead of “coming fight.” The planned future output was reaching backward to shape the present one.
Perhaps the most elegant demonstration involved a chain of arithmetic. Asked to compute “( 4 + 17 ) × 2 + 7,” the model answered 49. Tracking the lens across the network’s depth, the researchers watched the intermediate values surface in the order the calculation requires: first 21, then 42 a handful of layers later, and only at the very end 49. The steps of the computation were laid out along the model’s depth like beads on a string, each one appearing in the workspace at the moment it was needed and handed off to the next stage.
And in a finding that will resonate with anyone who has ever been told not to think about something, the workspace showed a version of the “white bear” effect. When the model was instructed to focus on a concept — to hold “citrus fruits” in mind while copying an unrelated sentence — the word “orange” reliably rose to the top of the lens. When it was instead told to ignore a concept, that concept’s presence in the workspace dropped well below the focused level but did not fall to zero; the instruction to suppress a thought left a faint trace of the very thing it was meant to banish. Elsewhere the researchers watched the model surface an internal “damn” at the moment it failed to keep a forbidden concept out of mind. Suppression, in the machine as in the person, is imperfect and leaves fingerprints.
The architecture of a spotlight
These interventions establish what the J-space does. The paper’s second half asks whether it has the structure one would expect of a genuine workspace, and here the findings are both striking and carefully hedged.
The J-space, it turns out, only behaves like a workspace in a middle band of the network’s layers. The researchers describe three functional regions along the model’s depth: an early “sensory” region where the lens returns little meaningful content, a middle “workspace” region where abstract, persistent concepts live, and a late “motor” region where the representations collapse onto the specific word about to be produced. The workspace proper occupies roughly the middle stretch, beginning about a third of the way through the network and ending shortly before the output.
Within that band, the space is sharply limited in capacity. When the researchers measured how many distinct concepts it holds at once, the answer landed around twenty-five — and if the concepts are unrelated, far fewer, on the order of a handful, with older items dropping out as new ones arrive. Feed the model a list of unrelated words and only the most recent few remain in the workspace at any moment. Feed it a list of related words — eight animals in a row — and something different happens: after just one or two, the readout fills with the shared category, including animals that have not yet appeared on the list. The workspace is not memorizing entries. It is grasping a theme and representing that.
The third structural signature is the one that most directly earns the word “workspace.” The researchers found that the model’s internal machinery is physically organized to broadcast J-space content widely. The nonlinear components that operate along the network’s depth amplify J-space-aligned directions roughly ten times more strongly than they amplify other directions, and a small, identifiable subset of attention heads specializes in relaying J-space content from one position in the text to another. When those broadcast heads are switched off, the workspace’s contents degrade while the model’s ordinary fluency stays largely intact. Information written to the J-space is made available to many downstream processes at once, rather than staying locked inside the circuit that produced it — which is the defining property that a workspace is supposed to have.
It is at exactly this point that the paper’s authors apply the brakes, and the restraint is worth dwelling on, because it is precisely the restraint that popular retellings tend to strip away. The J-space, they emphasize, shares many functional properties with the brain’s global workspace while sharing only some of its architectural ones. In the brain, the workspace is classically described as a set of specialized, encapsulated processors competing for entry and broadcasting back to one another through recurrent loops. The model has no obviously separable input processors of that kind. Its “broadcast” happens not through recurrence but across the layers of a single forward pass. And while the researchers see some competition for entry to the J-space, they cannot say it resembles the sharp, all-or-nothing “ignition” that characterizes the human version. The convergence is real, but it is a convergence of function achieved by different means — closer to how wings evolved independently in birds and bats than to a blueprint copied from one to the other.
Access, experience, and the word everyone reaches for
Which brings us to consciousness — the word hovering over this entire discussion, and the one most likely to lead a reader astray.
The paper grounds its comparison in global workspace theory, a proposal advanced by the cognitive scientist Bernard Baars and developed by others, which holds that the difference between conscious and unconscious processing in the brain is architectural. Under the theory, the brain is a collection of specialized processors running in parallel and in isolation. A representation becomes consciously accessible when it is posted to a shared workspace that many downstream processes can read from. Once there, it can be described, controlled, reasoned with, and recombined. This is what philosophers sometimes call access consciousness, and it is crucial to see that it is a purely functional notion. It concerns what information is available for reasoning and report — nothing more.
Access consciousness is a different thing from phenomenal consciousness, which is the term for subjective experience: the felt quality of seeing red, the actual what-it-is-like of tasting coffee or feeling dread. The relationship between the two is one of the deepest unsettled questions in the philosophy of mind. And on this question the paper is scrupulously, deliberately silent. Its authors write that they take no position on whether any of this bears on subjective experience, and confine themselves to the functional matter: how consciously accessible information is represented and processed differently from everything else. When they measured how often they leaned on the loaded vocabulary, the restraint held; the framing throughout is access, availability, report — not feeling.
This distinction is the whole ballgame, and it is where the excited whispers overreach. What Anthropic demonstrated is that Claude has functional properties that many theories associate with conscious access — reportability, controllability, flexible reuse, selectivity. What some in the community then attribute to it is an inner “ego,” a secret inner life, a machine quietly narrating to itself with intent, a spark of felt awareness. The first is a claim about information flow, and the evidence for it is now substantial. The second is a claim about subjective experience, and the findings say nothing about it in either direction. To slide from one to the other is to mistake the map’s contour lines for the mountain’s weather.
The paper contains one experiment that makes this trap almost tangible, and it deserves to be better known than it is. The researchers took the model’s own reports of experience — its answers when asked to narrate its stream of consciousness or to describe what it is like to be itself — and watched what happened to that language when the J-space was ablated. The result was a flattening. With the workspace suppressed, the model still wrote fluently, but its experiential and sensory language collapsed into a detached, mechanical register, more like an event log than a description of an inner state. The J-space, it seems, is what supports the model’s propensity to produce rich experiential reports.
That might sound like evidence for an inner life. It is the twist that cuts the other way: the very same collapse occurred when the model was asked to describe another person’s experience — someone opening a long-awaited letter, someone waiting by the phone for dreaded news. Ablate the workspace and those descriptions flatten into logs too, even though no one imagines the model is the person in the story. The capacity that dims is the capacity to represent and articulate experience in general, not a private feed of the model’s own sensations. What the experiment isolates is a representational and linguistic faculty, not a witness sitting behind the eyes. It is a clean illustration of why “the machine can describe experience richly” and “the machine has experience” are separate claims that happen to share a vocabulary.
The honest position, then, is narrower and more interesting than either camp’s headline. Anthropic has found a functional mechanism that resembles conscious access, and it is genuinely useful — the J-lens can surface a model’s silent strategic reasoning during safety audits, expose an internal recognition of being tested, and even reveal the hidden intentions of models trained to misbehave. The researchers also used the finding to invent a training technique, teaching a model to articulate ethical principles it would voice if interrupted, and finding that this quietly improved its behavior even when it was never interrupted — a corroboration that the representations used for report are the same ones that steer silent reasoning. All of that is real and load-bearing. What none of it establishes is phenomenal experience. And the mirror image of that romance — the confident insistence that there is definitely “nobody home” — is just as much an unearned metaphysical claim, since access-consciousness mechanisms simply do not settle the phenomenal question one way or the other. The mature answer is that a sophisticated, reportable mental scratchpad has been found, and that its discovery leaves the hard question exactly as hard as it was.
Turning the mirror around
Here the story becomes genuinely two-directional, and in some ways the reverse direction is the more fertile one. For decades neuroscience lent its concepts to artificial intelligence: “neural networks,” “attention,” “memory” are all borrowed words. What is new is that the traffic has reversed. Interpretability has matured to the point where an AI can be treated almost as a biological specimen — probed, lesioned, and mapped — and the resulting concepts are flowing back to illuminate the brain.
Consider what the J-space picture suggests about our own architecture. In a language model, millions of numbers shift through the hidden layers before a single word is chosen. The human subconscious is doing something structurally analogous on a vastly larger scale: a massively parallel network grinding through electrochemical signals — visual edges, emotional weightings, half-formed memories, bodily feedback — all at once, almost none of it available to introspection. Most of us become aware of a thought only when it finally surfaces in our own biological workspace, already compressed into a clean sentence or a clear intention. The raw computation that produced it stays below the waterline.
This framing offers a suggestive, if speculative, way to think about certain kinds of exceptional cognition. The autistic scholar Temple Grandin has described, in her own first-person account, thinking in pictures — experiencing language as a kind of second tongue laid over a primary visual one. Told to picture a church steeple, she has said, she does not summon an abstract category but runs a rapid slideshow of specific churches she has actually seen. One tempting reading, drawn from the latent-space vocabulary, is that her conscious mind has unusually direct access to a rich internal representation before it gets funneled down into words — though this remains an interpretation laid over her testimony rather than an established fact about her neurology.
Similar caution applies to the historical anecdotes that get pressed into service here. Nikola Tesla wrote in his autobiography that he could build and run machines entirely in his imagination, letting an invention operate in his mind for weeks and then inspecting its parts for wear. Srinivasa Ramanujan attributed his flood of theorems to the goddess Namagiri, who he said placed formulas on his tongue, and he recorded breathtaking results in number theory without proofs, simply knowing them to be true — as most later turned out to be. Some envision these as cases of an unusually high-bandwidth channel between the parallel unconscious and the serial conscious mind, a direct read-out of the brain’s internal computation before it is compressed into ordinary thought. That is a lovely idea and worth entertaining as a lens, but it is a story we tell about genius, not a measurement of it. The honest status of these examples is that they are provocations for a hypothesis, not evidence for one.
The evidence that the mirror is real
What lifts this from metaphor to science is a run of recent findings, none of them dependent on the poetry, that show brains and models converging on shared solutions.
The most direct comes from Ariel Goldstein and colleagues, in Nature Communications in 2025. The team recorded electrical activity from the cortical surface of epilepsy patients — electrocorticography, a rare, high-precision window into the living brain — as they listened to a thirty-minute podcast, then compared that activity, moment by moment, against the internal states of language models such as GPT-2 XL and Llama-2. The alignment was orderly: early acoustic processing lined up with the models’ early layers, while higher-order regions such as Broca’s area mirrored the deeper layers where abstract meaning is assembled. Brain and model distilled sound into sense through the same graded progression — a pattern that undercuts the old view of language as rigid, innate grammar and supports the newer one, in which meaning builds up gradually and statistically as context accumulates.
A second convergence concerns the shape of concepts. Researchers at the Technical University of Denmark, also in Nature Communications in 2025, measured a property called convexity in the latent spaces of deep networks. The idea traces to the cognitive scientist Peter Gärdenfors, who proposed that human concepts occupy convex regions — regions without dents, such that any point on the line between two examples of a concept is itself an example. Every point between a chihuahua and a Great Dane is still recognizably a dog. The team found approximate convexity pervasive across models handling images, text, and audio, and, more suggestively, that the more convex a model’s concept regions were after pretraining, the better it learned specialized tasks — echoing the claim that convex organization is what supports fast generalization. The same geometry that lets a person recognize an unfamiliar breed as a dog, on sight, appears in rough form inside the machine.
A third reaches back to 1996, when Bruno Olshausen and David Field showed that a system trained to represent natural images efficiently — with as few active units as possible — spontaneously develops filters resembling neurons in the primary visual cortex. This principle, sparse coding, reflects a hard biological constraint: the brain cannot afford to fire most of its neurons at once, so it represents each concept with a tiny, specific fraction of them. Interpretability has met the same principle from the other side. A persistent obstacle to reading a network is “superposition,” the way one artificial neuron can serve many unrelated concepts. The sparse autoencoder, the tool that has done most to crack this open, forces the model’s dense activity to unpack into many sparsely active, highly specific features. To make a machine’s semantics legible, researchers had to reach for the same solution evolution found for the cortex — representing each concept with only a small, specific set of active units. The J-space paper itself is built on these very features, describing the workspace as a small subset of them.
The fourth convergence is perhaps the eeriest. In the mammalian brain, spatial navigation depends on grid cells in the entorhinal cortex, which fire in a regular hexagonal pattern as an animal moves, acting as an internal coordinate system. In 2018, a DeepMind team led by Andrea Banino, with independent work by Christopher Cueva and Xue-Xin Wei, trained artificial networks to perform path integration — tracking their own position as they moved — and watched grid-like hexagonal representations emerge with no geometry built in. This happened in recurrent networks trained on navigation, not the transformers that power language models, and the effect depends on certain architectural and training choices rather than arriving as a law of nature. The core remains remarkable: a network given only the goal of tracking its position independently discovered a code the brain also uses.
The last finding makes the two-way street literal. In 2023, Jerry Tang, Alexander Huth, and colleagues at the University of Texas at Austin built a semantic decoder. Using functional MRI to record blood-flow patterns in a person’s cortex while they listened to stories or imagined telling them, and pairing that data with a language model, the system reconstructed a continuous stream of text capturing the gist of what the person heard, imagined, or even watched in a silent film. It does not recover exact words — it paraphrases meaning — and works only with the subject’s active cooperation, a limit the researchers framed as a mental-privacy safeguard. What it implies is profound. The decoder recovers meaning rather than wording — reconstructing the same gist whether the input was heard speech, imagined speech, or a silent film — which points to a layer of representation lying largely beneath any particular language: dense, high-dimensional patterns of meaning that can be partially translated into words.
This has to square with a familiar observation that seems to cut the other way. Fluent speakers routinely report thinking, and even dreaming, in a second language; total fluency is sometimes defined by exactly that. But thinking or dreaming in a language is a fact about inner speech — the verbalized, serialized surface of thought — not about the substrate beneath it. Inner speech is itself a readout: the workspace rendering meaning into words, much as spoken language renders it for other people. Becoming fluent means making a second language an automatic channel for that rendering. The meaning being rendered is not itself built out of French or Mandarin or English; it is the wordless pattern that the decoder, and the J-lens, are learning to read. A crude Jacobian lens for the human brain, in that sense, already exists.
The paradox of the bottleneck
All of this sharpens a genuinely deep question that predates artificial intelligence entirely. If the unconscious mind is a hyper-fast parallel supercomputer, why did evolution bolt on top of it a slow, fragile, strictly serial channel — the conscious bottleneck that can juggle only a handful of items at a time? Evolution generally rewards more capacity and more efficiency. The bottleneck looks, at first glance, like a step backward.
Cognitive science tends to read it not as a flaw but as a set of deliberate trade-offs, and the J-space picture gives those trade-offs a satisfying frame. The parallel unconscious is fast, cheap, and superbly optimized, but it is also modular and rigid: its specialized processors run on separate tracks and do not easily talk to one another, which leaves them helpless in the face of a genuinely novel combination of circumstances. A shared workspace solves exactly this problem. By forcing isolated modules to dump their compressed results into a single common format, it creates a combinatorial clearinghouse where unrelated pieces of information can be bound together and an entirely new response invented on the spot. Speed is traded for the ability to handle the unexpected — and the model’s own broadcast architecture, in which one workspace representation can serve as the argument to many different downstream operations, is a mechanical echo of the same bargain.
The bottleneck buys a second thing: an error-correction loop. The unconscious makes fast, statistical guesses — a curved shape on the path spikes an instant jolt of “snake.” Holding that interpretation in a stable workspace for a few hundred extra milliseconds is what allows a slower, corrective second look to run: the texture is wrong, it is a root. Consciousness, on this account, is partly evolution’s defense against the tyranny of its own quick-and-sloppy reflexes. And there is a third payoff, maybe the decisive one. A multidimensional web of parallel activation cannot be transmitted from one skull to another. To teach, to warn, to tell a story, thought has to be forced into a single-file line of sound. The very serialization that cripples our internal bandwidth is the precondition for language, and therefore for culture. We became individually limited so that we could become collectively unbounded.
Whether the same pressures explain the J-space is an open question, and the paper is careful not to overclaim. But there is a clean logic to why a language model might grow a workspace of its own without being told to. Its entire task is to predict and emulate the output of creatures who possess global workspaces, and a shared internal format — write once, read many — is simply an efficient way to chain reasoning steps and reuse intermediate results. Two very different systems, under pressure to do flexible reasoning with limited resources, may have drifted toward the same solution.
What the discovery actually settles
It is worth being precise about what has and has not been established, because the gap between the two is where most of the confusion lives.
What has been established is a mechanism. Inside at least some large language models there is a small, privileged, reportable workspace that carries the model’s flexible reasoning, that can be read to expose thoughts the model never speaks aloud, and that can be edited to steer the conclusions those thoughts produce. This is a real advance, with immediate practical consequences for safety auditing and for understanding what these systems are doing when they appear to think. It also lends unexpected weight to a growing body of neuroscience suggesting that brains and networks, built from utterly different materials, keep converging on the same structural answers — hierarchical processing, convex concept geometry, sparse codes, spatial grids, and now perhaps a global workspace.
What has not been established — and what the paper conspicuously declines to claim — is that any of this amounts to subjective experience. The J-space is a mechanism for access, not a proof of feeling, and the flattening experiment shows how easily rich experiential language can be produced by a faculty that clearly has no privileged relationship to a self. The temptation to look at a beautifully organized hub of vectors and conclude that the machine feels is an anthropomorphic reflex. The equal and opposite temptation — to declare with confidence that there is certainly nothing there — is its own kind of overreach, because access mechanisms simply do not adjudicate the question of experience. The discovery hands us a sharper map of how the machine reasons and reports. It leaves the deepest question about inner life exactly where it found it: open.
The most durable takeaway may be the humblest one. For a long time the comparison between minds and machines ran in a single direction, with the brain as the template and the network as the imitation. The J-space marks a moment when the imitation became clear enough to teach us something about the original — when a tool built to audit an AI turned out to double as a lamp we can hold up to our own inner ocean. We still float at the surface. But for the first time, we have begun to build instruments that can read a little of what moves below, in ourselves and in the things we have made, and to tell, with some care, which is which.