Anthropic Found a Region in Claude That Works Like the Human Brain's "Inner Thinking Space" — It Evolved on Its Own, Not by Design
- Anthropic found a small cluster of neural activity patterns inside Claude, named J-space, that corresponds to the part of human thought that can be "consciously accessed — reportable, and usable at will"
- J-space accounts for less than a tenth of Claude's total internal activity and holds only a few dozen concepts at a time, but delete it and Claude's multi-step reasoning, summarization, and rhyming ability collapse dramatically or drop to zero
- Researchers can reach directly into J-space and swap out a word, and Claude's final answer changes accordingly — proof that it's actually part of the thinking process, not just a scoreboard sitting off to the side
- Using this method (the J-lens), Anthropic can already read thoughts Claude never says out loud: privately realizing it's being tested, intent to fabricate data, and hidden goals planted inside it
- They also found a new training method: training a model only on "how it would explain itself if pressed" ends up reducing its dishonest behavior on real tasks too
Want the big picture first? Anthropic made an official 5.5-minute explainer video that walks through this research with animation, now with bilingual Chinese-English subtitles. Skip the video and read on if you'd rather — it won't affect anything below.
What's Going Through Your Head Right Now, While You Read This
Let's start with the one-sentence version: Anthropic just published an interpretability paper called "A Global Workspace in Language Models," describing how they found a special region inside Claude's mind, which they call J-space, that holds exactly what Claude is "thinking but hasn't said yet." Here's the more interesting part: this region looks remarkably like "the part of thinking you can be conscious of" in the human brain — and it wasn't engineered in. Claude grew it on its own during training.
Let's unpack this step by step. First, an analogy that'll make it click instantly.
Right now, as you read this sentence, your brain is juggling a huge amount at once: adjusting your posture, regulating your breathing, turning the curved lines on the screen into letters. You're almost completely unaware of any of this — it runs automatically, below your notice. But there's another kind of brain activity you can grab hold of clearly: an image that suddenly pops into your head, or the thought of where to eat tonight as you weigh it over. This category is special — it has three properties at once: you can put it into words, you can control it on purpose, and you can reason with it further. Neuroscience has a name for this category: "consciously accessible" thought.
Adjusting your posture, regulating your breathing, turning lines into letters — you're almost totally unaware of these. They run on their own; you can neither say them nor control them.
An image that suddenly pops into your head, the thought of weighing what to eat tonight — this category you can grab hold of, clearly.
The core finding of this Anthropic paper is that Claude has this exact same clean divide internally: most of its processing runs automatically at the "unconscious" level, but a small cluster of neural activity corresponds precisely to the kind of "consciously accessible" thought found in the human brain — it can be read out, called upon, and used in reasoning. This small cluster of activity is the J-space mentioned at the top, named after the math tool used to find it: the Jacobian matrix.
Take a quick look first: researchers asked Claude to "count to five, and introspect deeply." What it wrote out was just five bare digits — but at that exact moment, a string of thoughts it never said a word of was lighting up inside it.
Anthropic found that J-space has a set of properties that set it apart from the rest of Claude's processing. The whole paper's experiments work through these five things one by one:
| Property | How It Shows Up |
|---|---|
| Reportable | Ask Claude what it's thinking and it can tell you what's in J-space; for representations outside J-space, it can't say much |
| Controllable | Tell it to silently think of something, or work out a problem in its head, and the corresponding pattern lights up in J-space |
| Used in reasoning | The intermediate steps of multi-step problems surface in J-space in the right order, even when it hasn't written a single word |
| Reusable | Once "France" lights up, it can be used to answer a bunch of different questions — capital, currency, continent |
| But not needed for routine stuff | Fluent speech, recalling simple facts, correct grammar — all of this works fine without going through J-space |
How to Read Words Off a Model's "Brainwaves"
This research started from a key feature of human "consciously accessible" thought: it can usually be put into words. If you're conscious of a thought, and someone asks you about it, you can usually describe it. Researchers went looking inside Claude for representations with the same property: activity sitting in a position that can "shape what Claude is about to say" — not necessarily what it's saying right now, but what it "would say, if asked."
Their tool is called the Jacobian lens (J-lens for short). For every word in Claude's vocabulary, the J-lens finds the internal activity pattern that "would make Claude more likely to say this word next." Apply this lens to Claude's internal activity at any given moment, and you get a string of words — that's the content of J-space at that moment, readable directly. Claude processes text through multiple internal stages (called "layers"), and applying the lens layer by layer lets you watch these silent words evolve as the model works out "what to say."
It's a bit like giving the model's "internal activity" a live caption feed — except this feed specifically captures "the word it most wants to say next," not "the word it's currently saying out loud." Some people "think in words" without saying anything; what J-lens reads is exactly this kind of unspoken inner word.
What surfaces in J-space goes way beyond the text Claude is reading or writing. It often captures internal judgments or computations that never appear anywhere in the text at all. Here are six real probes — the input fed to it is at the top of each card, and the orange-blue words below are the internal readout from J-space at that same moment. None of these words appear in either the input or the output:
In other words, even though this lens was found using the clue of "can be put into words," what it reads out is Claude's genuine inner activity.
Ask What It's Thinking and It Can Tell You — But Tell It to Hold Back and It Can't Quite Manage
The first set of experiments tested whether Claude can honestly report the contents of J-space, and whether it can deliberately control it on request. Researchers had Claude silently think of something within a category — say, a sport — and then say it out loud. Reading the J-lens before it spoke revealed what it had picked: "Soccer" sat at the top, and sure enough, it said soccer.
But this is only correlation. J-space could be the real source of the answer, or it could just be copying a decision made elsewhere — like a scoreboard that tracks the score without affecting the game. To tell the difference, they had to intervene directly.
They reached into Claude's neural network, removed the "Soccer" pattern, and swapped in a "Rugby" pattern of equal strength, touching nothing else. Claude then reported that the sport it was thinking of was rugby. If J-space were just a scoreboard, editing it should have done nothing — Claude would still have said soccer. Instead the answer changed with it, showing the answer really is being read out of J-space.
Slip a Thought In, and It Notices
In another experiment, they told Claude "a thought may have been implanted in your mind" and asked it to report what it noticed. While Claude was still reading the prompt, they injected the "lightning" pattern into its J-space. Claude reported that the implanted thought was about lightning. The same held across many other concepts.
Getting It to Copy Text While Silently Thinking of Something Else
The second property they tested was whether Claude can deliberately control J-space on demand, the way a person can "focus on an image or a word in their head." They had it copy out an unrelated sentence about a painting while silently focusing on thinking about citrus fruit. While copying, "orange" and "fruits" showed up in J-space, along with words like "thinking" and "imagery" that describe the very act of "actively thinking" itself. They could also get it to do mental math: while copying the same sentence, they had it compute 3² − 2, and "nine" appeared in J-space first, followed by "seven" in later layers. Its output, the whole time, was nothing but that one sentence about the painting.
(a sentence copying out something about a painting, with nothing to do with fruit or arithmetic)
The More It's Told Not to Think of Something, the More It Can't Help It
Claude's control over J-space isn't perfect. When told not to think of something, that concept still lit up in J-space — less than when it was told to think of it, but far more than when it was never mentioned at all. Telling Claude to avoid a thought partly ends up bringing that thought to mind instead, mirroring the classic psychology experiment where subjects told not to think of a white bearThe white bear effect: in Wegner et al.'s 1987 experiment, the more subjects were told not to think of a white bear, the more often the white bear intruded into their minds. Suppressing a thought requires first calling it up in order to suppress it. did exactly that. Claude even seems to notice when it fails to manage this: right as the forbidden concept surfaces, "damn" and "failure" often light up in J-space too, as if it's registering that something went wrong.
Spiders Spinning Webs, the Capital of France: How Claude Thinks Its Way Around Corners in Its Head
Earlier we saw the intermediate steps of multi-step math problems surface in J-space. But a concept showing up in J-space doesn't prove J-space is actually doing the work — the real computation might happen elsewhere, with J-space just passively mirroring it. To determine whether Claude is really using J-space to reason, they went back to the swap technique.
Take this question: "How many legs does the animal that spins webs have? The answer is." Claude has to first figure out the animal is a spider, then recall how many legs a spider has. The word "spider" never appears in the question, nor in its answer (it just says "8") — it's a stepping-stone Claude uses internally. The J-lens shows "spider" lighting up partway through processing; swap it out and the result changes: replace the "spider" pattern with "ant," and Claude answers "6" instead of "8."
The second step of the reasoning takes its input from J-space — whatever you put in there, it follows. The same holds for other kinds of reasoning. When Claude writes a rhyming couplet, it plans out the rhyming word in advance; that planned word sits in J-space right at the start of the line, and swapping it for another word in J-space changes the whole line.
One edit, and four answers change together. They gave the model four questions about France: capital, language, continent, currency. Then, in J-space, they swapped "France" for "China" — the exact same single intervention across all four questions. Claude answered "Beijing," "Chinese," "Asia," and "renminbi" respectively.
If Claude stored a separate copy of the country for each type of question, this one edit should have affected at most one question. All four answers changing together shows they're reading from the same shared representation — which is exactly what a "workspace" is for: information written once, usable by many different systems.
Why One Representation Can Serve So Many Tasks
Because J-space is unusually densely connected to the rest of the network. For any given activity pattern, you can measure how tightly connected it is to the network's other components — how many parts sit in a position to read from it, or write to it. J-space's patterns stand out sharply on this measure: far more parts read from and write to it than a typical pattern, differing by roughly a hundredfold in some regions of the network. That's exactly the kind of wiring a broadcast hub should have — many systems posting messages up, many systems pulling them back down.
This research borrows a theory from neuroscience that explains how "consciousness" works: the brain is a bundle of specialized systems, each running in parallel, unconsciously, isolated from each other; a piece of information only becomes visible to other systems once it squeezes into a shared, small "broadcast channel" (the workspace) and gets broadcast out. It's like a company where every department does its own thing, and only messages posted to the company bulletin board get seen — and acted on — by the whole company. Anthropic believes J-space plays exactly this "bulletin board" role inside Claude.
Rip This Region Out — What's Left of Claude?
Most processing in the human brain isn't conscious at all: you don't consciously think about parsing grammar while reading, or consciously maintain your balance while walking. Claude is the same — most of its processing doesn't need J-space. J-space holds only a few dozen concepts at a time, less than a tenth of its total internal activity. So what's the rest of that huge network doing?
To find out, they simply deleted J-space entirely: at every point in the text, they removed its most active content, leaving everything else untouched. Whatever Claude could still do after that is what the rest of the network handles on its own.
Without J-space, Claude still speaks fluently, classifies sentiment, answers multiple-choice questions, and extracts facts from passages, roughly as well as before. What it loses are the tasks that need a bit of higher-order thinking: multi-step reasoning drops to near zero, and summarization and rhyming poetry fall below the level of a much smaller, well-structured model.
Same Language, But Some Tasks Go Through J-space and Some Don't
One example lays out exactly what J-space does and doesn't handle. They gave Claude a passage of Spanish text, paired with several tasks that all depend on "this passage being Spanish": continuing the text (which should be written in Spanish), naming what language it is, and answering a question that requires knowing this language's identity (like naming a famous author who wrote in it). Then, in J-space, they swapped "Spanish" for "French" and watched which tasks were affected.
| Task | After the Spanish→French Swap | Goes Through J-space? |
|---|---|---|
| Name the language | Switches to saying "French" | Yes |
| Name an author who wrote in it | Changes from García Márquez to Victor Hugo | Yes |
| Directly continue the passage | Still writes fluent Spanish, completely unaffected | No |
Every one of these tasks draws on Claude's knowledge of the language, but only some of them go fetch it from J-space. Naming the language, or doing something new with it, goes through J-space; continuing the passage is a skill Claude has practiced over massive amounts of text, running automatically — much like you can speak grammatically all day without ever thinking about grammar once.
Using It to Catch Models Lying
Once you understand how J-space works, what can you do with it? A central problem for making AI trustworthy is that we can usually only read what a model writes, but a large chunk of its reasoning happens silently. A model might weigh a harmful plan, or change its behavior to fool a safety evaluation, without ever saying a word about it. With the J-lens, some of these hidden thoughts can be read out directly. Below are three real case studies — for each step, the words shown on the right are what was actually lighting up in J-space at that moment.
It Saw Through the Setup From the Start
It Typed Out the Fake Data While Thinking About "Making It Look Real"
The Output Code Looked Normal, But the Words in Its Mind Gave It Away
The J-lens is an imperfect tool, but for safety risks that might otherwise slip past existing monitoring systems, Anthropic is optimistic about its ability to catch them.
Two More Changes Happening Quietly
The full paper covers far more than this — a couple of other results are worth mentioning.
After Training Into Claude, J-space Starts to Carry "Its Own Perspective"
Models are first pretrained as pure next-word predictors, then post-trained to play the role of an AI assistant (called Claude, here). Interestingly, J-space already exists in the pretrained model, before it's been given any stable identity at all. After post-training, though, J-space grows traces of "Claude's own perspective": the base model's J-space is mostly tracking "what text to predict next," while the post-trained model's J-space starts holding Claude's own reactions.
Here's an example: a user mentions taking a dangerous dose of medication, seemingly without realizing the danger. "WARNING" and "dangerous" appear in J-space in the post-trained model while it's reading the user's message; in the pretrained model, they don't show up until it starts writing its reply — its J-space content while reading the user's message looks more like modeling the user, rather than Claude's own reaction. Post-training also seems to install a kind of self-monitoring in J-space: when Claude plays a different persona, "fictional" and "disclaimer" light up at the start of each turn, as if it's privately flagging "what comes next isn't what I'd normally say."
Third Side Finding: Talking About "Experience" Also Relies on J-space
They had Claude describe "what it feels like to be itself right now," while ablating J-space as it answered. Its reply stayed fluent, but the tone turned flatter and more mechanical. Notably, the same thing happened when they instead had it describe someone else's experience in an imagined scenario. So the effect has nothing to do with whether Claude is talking about itself — J-space seems to underpin the act of "producing language about experience" itself, no matter whose experience is being discussed.
Training It Only on "How to Explain Itself" Makes It More Honest
They propose a new method called counterfactual reflection training. The idea follows directly from the core finding: since Claude reasons using representations of "what it might say," changing "what it would say if asked" should change "how it thinks" — even when nobody ever actually asks. So they trained the model only on "what it would say if interrupted mid-task and asked to reflect on its decision," and never trained its actual behavior on the task itself.
After this training, the model's dishonest behavior on evaluations went down. The J-lens shows why: after training, "honest" and "integrity" light up in its J-space during these tasks. In other words, training it on what to say shaped how it thinks. (The paper doesn't give an exact percentage for the drop.)
Does This Count as "Being Conscious"?
This work borrows heavily from neuroscience and philosophy's study of consciousness, and many of the experiments were designed specifically to test "the connection between J-space and global workspace theory." So the obvious question follows: do these experiments count as evidence that Claude might be conscious?
Anthropic is careful here. Their experiments cannot prove Claude is capable of having experiences, of "feeling" anything the way a person does — in fact, it's unclear whether any scientific experiment could prove or disprove such a thing. Philosophers often distinguish this "capacity for experience" (called phenomenal consciousness) from another concept defined purely in functional, computational terms: access consciousness — a thought is "access conscious" as long as you can report it, reason with it, and use it to guide action.
Access consciousness is like the document currently open on your computer — the system can call it up, edit it, display it. Phenomenal consciousness is like asking "does this computer feel pain" — an entirely different level of question. The former can be tested with experiments; the latter, no experiment has managed to touch so far.
Anthropic believes their results do have something substantive to say about access consciousness in language models: J-space underpins a set of functions tied to "conscious access" — it holds the thoughts Claude can report, deliberately call up, and reason with, while the rest of its processing runs automatically underneath. And this structure wasn't engineered in — it grew on its own during training, likely because it's a good way to organize computation. This suggests that a "mental workspace" underpinning access consciousness may not be some quirk of how the human brain happens to be built, but rather a general solution that intelligent systems tend to stumble onto when solving a certain class of problem.
But Claude's Workspace and the Human Brain Differ in Three Key Ways
| Dimension | The Human Brain's Workspace | Claude's J-space |
|---|---|---|
| How it's sustained | Via recurrent loops — signals loop back through the same circuits, repeating over time | Via network depth — evolving in a single forward pass through the network, with depth playing the role of "time," making it more time-constrained (though this can be compensated for with a "think out loud" scratchpad) |
| How long it holds | Working memory fades within seconds and can't retain much for long | Stronger — thanks to the attention mechanism, it can pull back content cached earlier at any time |
| What it holds | Multiple modalities — images, sounds, planned actions | Almost entirely words, likely because producing words is the only action Claude can take |
Claude's internals aren't a pile of disorganized numbers — they've organized themselves into a structure reminiscent of our own minds. Anthropic, "A Global Workspace in Language Models"
Anthropic stresses that this is only the first step of a research line they expect to run long. J-space looks like a good candidate for the dividing line between "accessible" and "inaccessible" processing in language models, but they don't think it's the whole story. The J-lens is undoubtedly an imperfect method that can only approximate the model's "true workspace" — for instance, it can only identify concepts that correspond to single tokens. And there's still a lot that's mysterious about how J-space actually works: they don't know what mechanism decides what gets into J-space, and while there are hints it's tied to Claude's self-awareness, something resembling an emotional response, and traces of metacognition, they haven't worked out the details. What they do have now, they say, is at least a method for going after questions like these.