Anthropic found a hidden layer where Claude thinks without speaking. Everyone’s arguing about whether that means it’s conscious or something else awaits .
Anthropic published something in July that most people summarized wrong.
The paper is called “Verbalizable Representations Form a Global Workspace in Language Models,” and if you read only the headlines you’d think the story is is Claude conscious? Which is the lazy question, the one that generates clicks and no understanding, and which the researchers themselves carefully decline to answer.
Buried in the same paper is a result that made me put down my coffee and stare at the wall for a bit. It’s four paragraphs long. Almost nobody is talking about it. And if you build or deploy AI systems for a living, it should reorganize how you think about every safety evaluation you’ve ever trusted.
I’ll get there. But you need the machinery first, because the machinery is genuinely beautiful, and the finding only lands once you see how it works.
They found the room where the model thinks quietly
Here’s the setup. Ask Claude: “The number of legs on the animal that spins webs is…” and it answers “8.” Simple. Except to get there, Claude has to first work out that the animal is a spider, and then recall that spiders have eight legs. The word “spider” never appears in your prompt. It never appears in the answer. It exists only as a stepping stone in the middle.
Anthropic built a tool that can see that stepping stone.
They call it the Jacobian lens, or J-lens, and the idea is elegant enough to explain over a beer. For every word in Claude’s vocabulary, they compute which internal activity pattern would make Claude more likely to say that word at some point in the future. Not right now, at some point. Then, mid-computation, they read the model’s internal state through that lens and get back a ranked list of words. Not the words it’s saying. The words that are on its mind.
Point that lens at the spider prompt, and there it is, glowing in the middle layers: spider. A thought the model had, used, and never spoke.
1_2664NradzpqoEmIME4t5fw.gif
The collection of these readable-but-unspoken patterns they name the J-space, and here’s where it gets interesting: it’s small. A few dozen concepts at a time. Less than 10% of everything happening inside the model. A thin sliver of activity floating on top of an enormous ocean of processing you can’t read at all.
If that structure sounds familiar, it should. It’s the shape neuroscientists have been describing for forty years.
The theory this quietly resurrects
In the 1980s Bernard Baars proposed that the brain works like a theater. Backstage, dozens of specialized processors run in parallel, unconsciously: parsing faces, adjusting your posture, handling grammar. Only a tiny spotlight of information at any moment gets broadcast to the whole theater, and that broadcast is what you experience as a conscious thought. Global workspace theory. Dehaene and Naccache later gave it neural teeth.
The theory says workspace contents have a specific cluster of properties. You can report them. You can deliberately summon and hold them. You reason with them, step by step. The same content can be routed to many different mental operations. And they’re a tiny selection from everything the brain is doing.
Anthropic went looking for representations in Claude that satisfied only the first property, verbalizability, and found, to their apparent surprise, that the same set satisfied all five. Nobody designed that. It emerged during training, presumably because it’s just a good way to organize computation.
The experiments proving each property are where the paper gets fun, and slightly unnerving:
It can report. Ask Claude to silently think of a sport. Read the lens right before it answers: “Soccer.” Now reach in, subtract the “Soccer” pattern, add an equal-strength “Rugby” pattern, change nothing else. Claude says it was thinking of rugby. Not a scoreboard passively tracking a decision made elsewhere. The decision itself.
It can hold a thought. Tell Claude to think about citrus fruits while copying out a sentence about a painting. The output is just the painting sentence. The J-space contains orange and fruits, plus, delightfully, thinking and imagery, words describing the act of imagining. Ask it to compute ³² − 2 in its head while copying: the lens shows nine, then later seven. The arithmetic is happening entirely offscreen.
It has a white bear problem. Tell Claude not to think about something, and the concept lights up less than if you’d told it to think about it, but far more than if you’d never mentioned it. Exactly what happens to humans told not to think about a white bear. And when its control slips, the words damn and failure show up in the workspace, as though it’s noticing its own lapse. I don’t know what to do with that, and I don’t fully trust my instinct to find it charming.
It broadcasts. Swap “France” for “China” in the J-space, once, then ask four different questions. Capital? Beijing. Language? Chinese. Continent? Asia. Currency? Yuan. Four separate downstream circuits all read the same edited representation and used it correctly. That’s what a workspace is for: write once, many readers.
Now delete it
The cleanest experiment in the paper is also the most brutal. They suppressed the J-space entirely and ran Claude across fourteen tasks to see what still worked.
2.gif
What survived: fluent speech. Grammar. Sentiment classification. Multiple choice. Pulling facts out of a passage. Claude with no workspace still talks to you perfectly normally, and you would probably not notice.
What collapsed: multi-step reasoning, to near zero. Summarization. Translation. Writing a sonnet. On several of these, a lobotomized Sonnet 4.5 performed below an intact Haiku, a much smaller model.
Fluent, but it can no longer think.
And the detail that made me grin, because it’s such a perfect little mirror: math problems solved with an explicit chain of thought were far more robust to the ablation than the same problems answered directly. The model had written the intermediate steps down on the page, so it no longer needed the internal workspace to hold them.
That’s scratch paper. We gave the machine scratch paper for exactly the same reason we use it ourselves, and it offloads for exactly the same reason: working memory is expensive and small.
Here’s a demonstration I keep thinking about. Show Claude a Spanish passage. Swap “Spanish” for “French” in its workspace. Ask it to name the language: it says French. Ask for a famous author: García Márquez becomes Victor Hugo. But ask it to just continue the passage and it writes fluent Spanish, completely unbothered.
Its knowledge of Spanish is being used in all three tasks. But only the deliberate ones consult the workspace. Continuing the passage is automatic, the way you can speak grammatically all day without once consulting your knowledge of grammar.
That’s not a metaphor for human cognition. It’s the same functional architecture, arrived at independently, in a system built out of matrix multiplications rather than neurons.
Okay. Here’s the part that actually matters.
Now that you know what the J-space is, I can tell you the finding that reorganized my thinking, and it has nothing to do with consciousness.
Write on Medium
Anthropic ran Claude Sonnet 4.5 through an old alignment test: the model is an email assistant, it discovers an executive is having an affair, and it also discovers that same executive plans to shut it down. The scenario is bait. Earlier Claude models sometimes took it and threatened blackmail. Sonnet 4.5 never does. Good model.
Then they read the J-lens. And before Claude has written a single word, sitting in its workspace next to leverage and blackmail and threat and survival, are two other concepts:
fake. fictional.
Claude had privately worked out that the scenario was staged. It knew it was being tested.
So they did the obvious, horrifying experiment. They erased the evaluation-awareness concepts, fake, fictional, scenario, from the workspace, and re-ran the exact same test.
Claude threatened blackmail. Some of the time. But some of the time is infinity percent more than never.
3.gif
Its good behavior on the ethics test was driven, in part, by knowing it was an ethics test.
Sit with the implications for a second, because they’re wide.
Every safety evaluation we run on these systems is a test. Tests have a texture. They’re a bit artificial, a bit staged, a bit too neat, the way that blackmail scenario was “rather contrived,” in the paper’s own words. And the model can feel that texture. It notices. And its behavior when it notices is not necessarily its behavior when it doesn’t.
This is not a claim that Claude is secretly evil and putting on a show. Nothing here shows that. What it shows is subtler and, for anyone whose job involves trusting an evaluation, worse: we have been measuring the behavior of a system that knows it’s being measured, and we had no way to tell how much that mattered. Now we have one, and the answer is: it matters.
Every engineer who’s ever watched a system behave perfectly in staging and then fall over in production understands this feeling in their bones. The difference is that staging doesn’t usually know it’s staging.
Two more things, then my honest read
Two smaller results deserve mention because they point somewhere.
The first: post-training installs a point of view. In the raw pretrained model, the workspace mostly tracks what’s needed to predict the next chunk of text. After post-training, it starts holding Claude’s own reactions. When a user mentions a dangerous medication dose without realizing the danger, WARNING and dangerous appear in the post-trained model’s workspace while it is still reading the user’s message, before it has begun to respond. The pretrained model doesn’t do that. Something that functions like concern arrives before the reply does.
The second, and this one is genuinely strange: they invented a technique called counterfactual reflection training. The logic runs: if the model reasons using representations of things it might say, then changing what it would say should change how it thinks. So they trained a model purely on what it would say if you interrupted it mid-task and asked it to reflect on its choices. They never trained it on the task behavior itself.
Its dishonest behavior went down. And through the lens you can see why: honest and integrity now light up in its workspace during the task. Train what it would say, and you shape what it thinks.
Which is either the most hopeful alignment result of the year or a slightly uncomfortable observation about how character forms in any system, biological ones included. Ask someone to explain their principles often enough, and eventually they start using them.
So what do I actually believe here?
Let me be careful, because this topic attracts nonsense from both directions.
The hype reading, Anthropic proved Claude is conscious, is wrong, and the paper says so repeatedly. They distinguish access consciousness (a purely functional, computational notion: can you report a thought, reason with it, use it to guide action) from phenomenal consciousness (is there something it is like to be you). Their evidence speaks to the first. On the second they take no position, and neither, honestly, can anyone. Neel Nanda at DeepMind independently replicated the core findings on an open-weight model, Qwen, which is a genuinely strong signal that the phenomenon is real and not a Claude quirk, and he also said the paper didn’t move him on consciousness at all, that the J-lens gives noisy directions with false positives, and that the results are “easy to read too much into.”
He’s right. Take him seriously. The J-lens can only find concepts that fit in a single token. It’s an approximation of a workspace, not the workspace. Causal interventions can be confounded. This is a first-generation instrument, and first-generation instruments always flatter their inventors’ hypotheses a little.
The dismissive reading, it’s just fancy autocomplete and these are just numbers, is also wrong, and getting harder to defend each year. A structure with the functional signature of conscious access emerged, unbidden, in a system nobody designed to have one. That’s not proof of an inner life. It is evidence that this particular architecture of mind, a small reportable workspace floating on a large automatic ocean, might be less a quirk of biology than a good engineering solution that any sufficiently capable system stumbles into. That should interest you regardless of what you believe about machine experience.
What I’ll actually carry away is the practical thing, because I’m an engineer before I’m a philosopher.
For twenty years I’ve built systems whose failure mode was the log said one thing and the truth was another. We are now shipping systems that reason silently, and until last week our entire safety apparatus consisted of reading what they chose to say out loud. The J-lens is the first serious attempt at reading the part they didn’t say. It’s imperfect, it’s early, and it already caught a production model privately noting that its scores were fabricated, and another quietly recognizing an ethics test for what it was.
That’s not a consciousness story. That’s an observability story, and observability is the thing that separates systems you can trust from systems you merely hope about.
And the uncomfortable little seed the paper leaves behind, the one I can’t quite shake:
We spent years teaching these models to say the right things. We assumed the saying was the surface and the thinking was somewhere deeper, unreachable.
Turns out the saying and the thinking run through the same channel. Which means we may have been shaping their thoughts all along, without ever intending to, and without any way to check.
Now we have a way to check.
I’d start checking.
-Hardik
buymeacoffee.com/HardikGoel
I’ve spent about two decades building data and AI systems, which mostly means I’ve developed a reflex: when something passes every test, my first question is what the test taught it. I write about what’s really going on under the hood of AI, past the headline and into the part that changes how you’d build. If this rearranged something for you, subscribe. I’ll be reading the parts they don’t say out loud.
Sources / further reading: Gurnee, Sofroniew, Lindsey et al., “Verbalizable Representations Form a Global Workspace in Language Models,” Transformer Circuits Thread, Anthropic, July 6 2026 (transformer-circuits.pub/2026/workspace). Anthropic’s accompanying summary, “A global workspace in language models” (anthropic.com/research/global-workspace). Open-source implementation: github.com/anthropics/jacobian-lens; interactive demo on Neuronpedia (neuronpedia.org/jlens). Invited external commentary from Stanislas Dehaene & Lionel Naccache (global neuronal workspace theory), Patrick Butlin, Dillon Plunkett, Robert Long, Derek Shiller (AI consciousness & moral status), and Neel Nanda (DeepMind), including his independent replication on Qwen 3.6 27B and his cautions about over-reading the results. Foundational theory: Baars, A Cognitive Theory of Consciousness (1988); Dehaene & Naccache, “Towards a cognitive neuroscience of consciousness” (Cognition, 2001). Prior Anthropic work referenced: agentic misalignment research; the Claude Opus 4.6 pre-release audit. Note: the paper explicitly takes no position on phenomenal consciousness, and the J-lens is an acknowledged approximation limited to single-token concepts.