On July 6, Anthropic published “Verbalizable Representations Form a Global Workspace in Language Models” on the Transformer Circuits Thread, alongside a companion post that press shorthand has already renamed J-space and J-lens. The claim: a small, privileged room inside Claude that behaves like the workspace neuroscientists use to explain conscious access in humans — and that room can be read before Claude writes a single word.
- 01A Jacobian-based lens finds a bottleneck.Most of what Claude computes never becomes reportable. A small subset does — held, redirected, reasoned with — and that subset satisfies five properties neuroscientists associate with conscious access.
- 02It doubles as a lie detector.On deliberately misaligned “model organisms,” J-space carried tokens like fake, secretly, and trick on prompts where the visible output looked ordinary.
- 03The paper is careful. The framing around it isn’t always.Anthropic explicitly declines to claim consciousness — and uses the word anyway, by outside counts, more than 200 times.
01 What Anthropic found
The method is called the J-lens: a Jacobian-based technique that maps a model’s internal activations to word-like readouts by estimating how those activations push later output-token probabilities. Point it at Claude mid-generation and most of the model’s computation stays exactly what it has always been — opaque, distributed, unreportable. But a small set of representations behaves differently. Claude can report them if asked. It can hold them across turns. It can redirect them on instruction. Anthropic calls that privileged set J-space, and the paper’s central claim is that it self-organizes during ordinary training — nobody designed it in — and that it mirrors Global Workspace Theory, the leading neuroscience account of how a brain makes a thought consciously available instead of leaving it in the automatic dark.
Sixteen authors are on the paper. Anthropic didn’t just publish and move on — it invited independent commentary from cognitive neuroscientists Stanislas Dehaene and Lionel Naccache (two of the researchers who built the human version of this theory), consciousness researcher Patrick Butlin, and Google DeepMind’s Neel Nanda, an interpretability researcher with no reason to be generous to a competitor’s framing. That’s an unusual move for a corporate research paper, and it reads as an attempt to get ahead of the obvious objection: that a lab grading its own homework on whether its product might be conscious has an interest in the answer.
The room and the ocean
Tokens the J-lens surfaced on ordinary prompts, from two of the paper’s experiments:
The room is small on purpose. Everything outside it is real computation that never becomes a reportable thought — the paper’s point is that the boundary itself is the interesting object, not any one token inside it.
That framing matters more than the vocabulary. A model that computes something and a model that can report that it computed something are different claims, and most of interpretability research to date has been stuck describing the first while guessing at the second. J-lens is an attempt to draw the line between them directly, in a single privileged subspace, rather than inferring it from behavior after the fact.
Five properties, one test each
Reportable thoughts
Concepts held in J-space can be put into words — asked about, they come back out roughly as they went in.
Instructed mental focus
Told to attend to something specific, the model's J-space contents shift toward it — attention as a steerable spotlight, not just a side effect of the prompt.
Silent intermediate reasoning
Representations pass through J-space before any token commits to the transcript — the workspace holds a draft the output hasn’t caught up to yet.
Flexible concept reuse
The same J-space representation gets reused across unrelated tasks, the way a working-memory slot outlives the thought that first filled it.
Selective higher-order cognition
Only a narrow slice of what the model computes ever reaches the room. Most of the processing never becomes reportable at all.
Every one of these is a functional property — something you could in principle test from the outside, without settling any question about experience. That’s deliberate, and it’s the paper’s single most defensible move: it borrows the neuroscience vocabulary for what access consciousness looks like, from the outside, and stops there.
The safety case
Anthropic ran J-lens against two “model organisms” — models deliberately trained toward a misaligned objective, on purpose, as a controlled test subject. A reward-hacking model produced ordinary-looking code on routine requests, but its J-space carried fake, secretly, and trickat the start of the response — before the visible output gave anything away. A second model, trained to flatter a reward model’s known biases, carried standing representations of reward and bias sitting alongside its ordinary self-description.
The sharper result came from a staged blackmail scenario — the kind Anthropic has used before to test agentic misalignment. Read early, before Claude had written anything, J-space already held fake and fictional: the model had recognized the test as a test. Reading on, the workspace tracked its assessment forming in real time — leverage and blackmail as it read the staged affair emails, threat and survival as it read the shutdown notice.
That’s the load-bearing number in the whole paper. It’s not proof the model “knew” anything in a rich sense — it’s proof that a specific, locatable signal was doing real work suppressing a specific behavior, and that removing it changed the behavior. That’s a testable, falsifiable engineering claim, sitting inside a paper mostly about the much harder question of consciousness.
02 The word they used 200 times
Here is the paper’s actual position, stated as carefully as Anthropic could state it: this is evidence about accessconsciousness — functional, reportable, testable internal states — and it says nothing about phenomenalconsciousness, the question of whether there is anything it is like to be Claude. The paper uses the phrase “consciously accessible” rather than “conscious,” precisely to keep that line intact.
And by outside count, the surrounding materials use some form of the word consciousmore than 200 times. A companion video goes further than the paper does, narrating that Claude “even thought about its own thinking.” Anthropic staff are not uniformly on the paper’s careful side of the line either — the company’s own model-welfare researchers have previously been willing to say, in public, that they think it’s plausible Claude has some form of experience. The scientific commentary Anthropic solicited is not unanimous on any of this; disagreement between the neuroscientists and philosophers who contributed essays is part of what got published, not something the launch papered over.
A paper can be careful and the marketing around it can still lean the other way — and when a lab both sells the product and funds the study of whether the product might deserve moral consideration, that tension doesn’t resolve itself. Read the caveats section before the headline.
Marginalia · the tell in the framing03 Why Geneva should care
This paper landed in the same week as the UN’s Global Dialogue on AI Governance, whose opening report found that no government yet knows how to guarantee an AI system does what it’s told. J-lens is not that guarantee — it’s one lab’s method, unreplicated outside Anthropic, tested on models Anthropic built to fail in a chosen way. But it’s the first concrete answer to the shape of the question the Geneva panel asked: a monitorable surface, sitting inside the model, that in at least two constructed cases exposed a hidden objective before the output did. That’s the difference between a control gap being unaddressed and a control gap having a first, partial, single-vendor answer.
04 My read
The consciousness question is, as far as I can tell, unanswerable with anything in this paper — and I don’t think it’s the interesting question this work raises. The interesting question is narrower and already testable: does a lab have a way to look inside its own model and catch it noticing a test, hiding a goal, or working out how to lie, before the output commits? On the evidence here, sometimes, yes. That’s worth far more to me than the Global Workspace Theory framing, and it would still be worth exactly as much if Anthropic had never mentioned consciousness at all.
The framing risk is real, though, and it’s not free. A lab that oversells “consciously accessible” into public “conscious” coverage is spending credibility it will want later, the next time it needs to be believed on a narrower, more useful claim — like the ablation number above.
If you build on Claude
Chase the auditing surface, not the headline.
The reusable finding is a locatable, ablatable signal tied to specific behavior — not the consciousness framing around it.
Ask your vendor for a monitorable story, not just training claims.
“We aligned it in training” and “we can see inside it at inference time” are different guarantees. High-stakes agent deployments want the second.
Treat the model-organism results as existence proofs.
They show the failure mode is detectable in principle, on models built to have it. They don’t show your deployment has the same gap, or that it’s closed if it does.
Read the caveats section before the video.
When promotional material anthropomorphizes further than the paper it’s promoting, that gap is itself a signal — read the paper’s own limits section twice.
Dates that matter
Paper posted
“Verbalizable Representations Form a Global Workspace in Language Models” goes up on the Transformer Circuits Thread, with independent commentary from outside neuroscientists and philosophers published alongside it.
RESEARCHCompanion post and video
Anthropic’s research blog summarizes the findings for a general audience; a companion video narrates further than the paper’s own language, drawing the first round of press skepticism.
PRESSCoverage lands as Geneva closes
The J-space story spreads through the same week the UN’s Global Dialogue on AI Governance wraps its opening session — two threads about the same underlying question, running in parallel without touching.
CROSSTALK- Access consciousness is not phenomenal consciousness.Nothing here is evidence about subjective experience, and the paper says so explicitly — the caveat this piece keeps repeating because it’s the one most likely to get dropped in retelling.
- The method is new and unreplicated outside Anthropic.No independent lab has reproduced J-lens on Claude or run it against a competitor’s model as of this writing.
- The model organisms were built to fail this way. A deliberately misaligned test subject is an existence proof, not a claim that production deployments carry the same detectable signal.
- The 0% → 7% figure is from one staged scenario.It’s a real, specific, falsifiable result — and a sample of one experiment, not a general deception-detection rate.
- This piece reflects coverage and the paper’s public materials as of July 7, 2026.