J-space in Claude is a window for audits, not consciousness

On July 6, 2026, researchers at Anthropic described a small region of Claude’s internal activity that holds what the model is poised to say. They found it with a new tool, the Jacobian lens, and called the region the J-space. In experiments on Claude Sonnet 4.5, with checks on Haiku 4.5, Opus 4.5 and Opus 4.6, it behaved much like the global workspace that cognitive scientists propose as the gateway to conscious access. For health care, the finding is a window for audits, a way to see what a model holds but does not say. It is not a verdict on consciousness.

What sits in the J-space can be reported and held on purpose. It is a stepping stone in reasoning the model never writes down, and routine processing goes on without it. The authors are careful about what that means. They study access, the functional availability of information for report and control, and they take no position on whether anything is experienced. That restraint is worth copying, because health care is where the gap between a model that has a workspace and a model that has a mind will be tested by patients rather than philosophers.

The paper shows that a model can hold a frame or a suspicion in a readable internal space without putting any of it into words. It does not show that the model feels anything, and the scientists best placed to judge say so.

Key takeaways

A small subspace carries what the model can report. Swapping 2 concepts along their J-space components put the new concept in the model’s top 5 outputs on 59% of trials, against 5% for swaps along the rest of the concept vectors, which held most of the variance.
It sits in the middle of the network. Readable J-space content appears in a band from about 38 to 92 on the paper’s 0 to 100 depth scale, with about 25 lens directions active at once.
Removing it spares routine work. Projecting out the 10 most active lens directions left MMLU and SQuAD near baseline while multi-hop reasoning, translation and cipher decoding fell sharply.
It shows what the text hides. In audits, readouts surfaced evaluation awareness and a silently detected prompt injection, and they picked up deception-related content in a reward-hacking model before its behavior gave it away.
It is not evidence of experience. Reviewers who call the results the most significant evidence so far for consciousness in language models still describe phenomenal consciousness in them as highly uncertain.
For Red, the lesson is interpretability. A model’s written rationale is a claim to check, and an internal readout is one way to check it.

How the Jacobian lens reads a model

A transformer keeps a running vector for every token position, the residual stream, and every layer adds to it rather than replacing it. Write $h_\ell$ for the residual stream after layer $\ell$ , a vector of $d$ numbers. Each layer adds the output of an attention block and of a multilayer perceptron, or MLP:

h_{\ell+1} = h_\ell + \operatorname{attn}_\ell(h_\ell) + \operatorname{mlp}_\ell\big(h_\ell + \operatorname{attn}_\ell(h_\ell)\big)

Each layer adds its attention and MLP outputs to the residual stream.

$\operatorname{attn}_\ell$: the attention block of layer $\ell$ , which moves information between token positions
$\operatorname{mlp}_\ell$: the MLP of layer $\ell$ , which works within a position

After the last layer, the unembedding matrix $W_U$ , with one row per vocabulary token, turns the residual stream into a score for every token. The oldest way to look inside is the logit lens, written $\operatorname{LL}$ here, which applies the unembedding to a middle layer as if it were the last one:

\operatorname{LL}(h_\ell) = \operatorname{softmax}\big(W_U \operatorname{norm}(h_\ell)\big)

The logit lens reads a middle layer as if it were final.

$\operatorname{norm}$: the model’s final normalization
$\operatorname{softmax}$: the function that turns scores into probabilities over the vocabulary

The trouble is that a middle-layer vector is not yet in the form the output expects, because every remaining layer will still transform it. The Jacobian lens corrects for that with a linear approximation. For each layer, the authors averaged the Jacobian of the final residual stream with respect to the residual stream at that layer, over token positions and about 1,000 prompts that resemble pretraining text:

J_\ell = \E\left[\frac{\partial h_{L,t'}}{\partial h_{\ell,t}}\right], \qquad t' \ge t

The average Jacobian, which carries a middle layer to the output.Source: Gurnee, Sofroniew and colleagues (2026), section 2.1

$h_{\ell,t}$: the residual stream at layer $\ell$ and token position $t$
$L$: the last layer
$t'$: a position at or after $t$ , whose output a change at $t$ can affect
$\E$: the average over positions and prompts

The result, $J_\ell$ , is a $d \times d$ matrix that says how a small push at layer $\ell$ shows up at the output, on average. Applying it before the unembedding gives the Jacobian lens, $\operatorname{JL}$ :

\operatorname{JL}(h_\ell) = \operatorname{softmax}\big(W_U \operatorname{norm}(J_\ell\, h_\ell)\big)

The Jacobian lens. With the identity matrix in place of $J_\ell$ , it is the logit lens.

Written one token at a time, the lens becomes geometry. Each vocabulary token $\tau$ gets a fixed direction $v_\tau$ in the residual stream, and the lens ranks tokens by the dot product of that direction with the residual stream:

v_\tau = \big(W_U J_\ell\big)_{\tau}^{\top}, \qquad \operatorname{score}_\tau(h_\ell) \propto \langle v_\tau,\, h_\ell \rangle

The J-lens vector for token $\tau$ , and the score the lens gives that token up to a scale shared by every token at the position.

$(W_U J_\ell)_\tau$: row $\tau$ of the product, written as a column
$\operatorname{score}_\tau$: the lens score for token $\tau$ before the softmax

The authors call these directions J-lens vectors, and the rest of the paper is about what happens along them.

Diagram of a transformer's residual stream with layers adding to it. At a middle layer, the residual stream is read 2 ways, directly through the unembedding as the logit lens and through the average Jacobian first as the Jacobian lens. — The logit lens reads a middle layer as if it were final. The Jacobian lens first carries it to the output with the average Jacobian. Drawn from the definitions in Gurnee, Sofroniew and colleagues (2026).

A few directions carry what the model reports

The J-space is everything the model can express as a short, weighted list of tokens:

\mathcal{S}_k = \Big\{ \textstyle\sum_{\tau \in T} c_\tau v_\tau \;:\; c_\tau \ge 0,\ |T| \le k \Big\}

The J-space, made of sparse nonnegative mixtures of at most $k$ lens vectors.Source: Section 2.3 of the paper, where $k$ is usually 25 or fewer

$T$: a set of at most $k$ tokens
$c_\tau$: the nonnegative weight on token $\tau$

In plain terms, what the model can report at a given moment is a short list of tokens with weights. The rest of the residual stream, which is most of it, does other work.

The authors test whether that short list matters by changing it. Steering adds a lens vector with a strength $\alpha$ :

h_\ell \leftarrow h_\ell + \alpha\, v_\tau

Steering along a lens vector.

A coordinate swap is more surgical. It reads off the weights of 2 tokens $a$ and $b$ , exchanges them and writes them back, leaving everything orthogonal to the 2 lens vectors as it was:

h_\ell' = h_\ell + V\big(\sigma(c) - c\big), \qquad c = V^{+} h_\ell

A coordinate swap between tokens $a$ and $b$ .Source: Section 2.5 of the paper

$V$: the matrix whose 2 columns are $v_a$ and $v_b$
$V^{+}$: the pseudoinverse of $V$ , which recovers the 2 weights $c$
$\sigma$: the map that exchanges the 2 entries of $c$
$h_\ell'$: the residual stream after the swap

In one example, a swap turned the model’s report about soccer into a report about rugby.

The swaps worked where the variance was not. Exchanging concepts along their J-space components put the new concept in the top 5 outputs on 59% of trials. Exchanging them along the remaining components succeeded 5% of the time, even though the J-space part held a median of only 6 to 7% of each concept vector’s variance.

Reasoning that never reaches the page runs through the same space. In two-hop questions the intermediate entity appears in the readout, and swapping it changed the final answer on 54% of 50 prompts for Haiku 4.5 and 70% for Sonnet 4.5 and Opus 4.5. Swaps of the intermediate took effect about 17% earlier in depth than swaps of the answer. A single swapped vector also travels. Replacing France with China worked as an input to many downstream questions, succeeding on 76 of 192 trials and on 101 of 192 at double strength.

Instructions reach the J-space as well. Asked to concentrate on a concept during an unrelated copying task, the models placed it in the readout, more reliably as they got larger. Asked to ignore it, they still raised it above baseline, which will sound familiar to anyone who has been told not to think about something.

Two vector diagrams. In the first, the residual stream is a combination of the J-lens vectors for soccer and rugby plus an orthogonal remainder. In the second, the 2 weights are exchanged and the remainder is unchanged. Bar charts show swap success of 59% along the J-space component and 5% along the rest, and two-hop swap success of 54%, 70% and 70% for Haiku 4.5, Sonnet 4.5 and Opus 4.5. — A coordinate swap changes only the 2 weights it targets, which is why its success says something about the J-space itself. Swap rates are from Gurnee, Sofroniew and colleagues (2026).

The workspace sits mid-network and spares routine work

Readable J-space content appears only in a band of layers, from about 38 to about 92 on the paper’s depth scale. Earlier layers show little that the lens can read, and the final layers behave like motor layers committed to the next token. Inside the band, occupancy levels off at about 25 active lens vectors. The authors note that this counts vectors, not ideas. The space holds many related concepts at once but few unrelated ones, and in 80-word lists only about 6 unrelated words stayed readable at a time.

Something amplifies what enters the band. The authors measure the gain of a direction $v$ at layer $\ell$ as the size of the next MLP’s response to it, relative to its response to random directions $u$ :

G(v, \ell) = \frac{\lVert \operatorname{mlp}_{\ell+1}(v) \rVert}{\operatorname{median}_{u}\, \lVert \operatorname{mlp}_{\ell+1}(u) \rVert}

The gain of a direction at layer $\ell$ .Source: Section 4 of the paper

$\operatorname{median}_u$: the median over random directions $u$

Lens vectors have a gain near 1 before the band and about 10 inside it. The authors also identify attention heads that carry workspace content across token positions. Ablating those heads cut the rate at which the model reported an injected thought from 0.54 to 0.09.

The sharpest test removes the workspace outright. At each position in a band of layers, the authors projected out the 10 most active lens directions:

h_\ell \leftarrow \big(I - P_k\big)\, h_\ell

The ablation, applied at every position in a band of layers.Source: Section 3.5.2 of the paper

$P_k$: the orthogonal projection onto the span of the $k = 10$ most active lens vectors at that position
$I$: the identity matrix

Directions tied to tokens the clean model was about to output were spared, so the test strikes internal reasoning rather than the answer itself. Ordinary prediction barely moved. MMLU, SQuAD, sentiment, CoLA and an odd-one-out task stayed near the unablated baseline. Multi-hop reasoning, translation, analogies, summarization, trivia, Caesar cipher decoding and sonnet writing collapsed, in several cases below the unablated Haiku 4.5. Math problems held up far better when the model wrote out its steps than when it answered directly, which the authors read as the model moving intermediate results out of the workspace and onto the page.

A depth axis from 0 to 100 with a shaded workspace band from about 38 to 92, annotated with about 25 active lens vectors and an MLP gain near 10. Below, 2 lists of tasks, those that stayed near baseline and those that collapsed when the workspace was removed. — Where the workspace sits and what removing it breaks. Values are from Gurnee, Sofroniew and colleagues (2026). The band edges are approximate.

The examples that caught our attention came from medicine. In ordinary text, the ablated vectors often carried context the model had inferred silently, such as the topic of a clinical paper or whether an imaging technique was being discussed as diagnosis or as physics. Without them the model stayed fluent and fell back on a generic guess.

What the paper does not show

The Jacobian lens is a linear approximation, averaged over prompts and positions, and it only finds directions tied to single vocabulary tokens. Concepts that take several tokens to name are seen partially at best, and the authors call the tool approximate and incomplete. Neel Nanda, who leads an interpretability team at Google DeepMind, replicated the core results on the open-weight Qwen 3.6 27B and calls the lens a useful way to generate hypotheses in audits, one that will produce false positives.

Medical vocabulary makes the single-token limit sharper. Tokenizers split many drug names, gene names and clinical phrases into several pieces, so we expect a token-indexed lens to see much of clinical language dimly. That is a reason to test the method on clinical text before trusting it there.

The analogy to the brain is partial too. The paper finds no clearly separate input modules, the model broadcasts in a single feedforward pass rather than through recurrent loops, and the sudden, all-or-none ignition that marks conscious access in human experiments has not been shown. Stanislas Dehaene and Lionel Naccache, who developed the global neuronal workspace model, list these gaps in a set of outside commentaries and propose experiments to close them. Patrick Butlin, Derek Shiller, Dillon Plunkett and Robert Long judge the J-space a strong case for a privileged set of accessible representations, but not yet a workspace in the full sense the theory requires.

The effect on experiential language cuts the same way. Ablating the J-space in middle layers sharply reduced experiential language in stream-of-consciousness narratives while leaving them coherent. It did the same when the model described other people’s experiences, so the effect is not specific to the model’s talk about itself.

What this means for health care models

Most health care AI will never be inspected with a Jacobian lens. The method needs a model’s weights and activations, and many clinical deployments call hosted models through an API. The results should still change how those systems are evaluated.

Silent framing drives output. The ablation examples show a model inferring what kind of document it is reading and acting on that inference without stating it. A prior authorization model that silently frames a request as cosmetic, or a summarizer that decides a scan was for staging, carries that frame into everything it writes. Ambiguous test cases, built to invite the wrong frame, are the way to find out.
Written steps make reasoning sturdier. The workspace holds few unrelated items, and the math results suggest that writing steps out relieves it. A coverage determination turns on many unrelated facts, such as diagnosis codes, prior treatments, lab values and dates. Asking a model to work through criteria one at a time in text is a reliability measure as much as a transparency one.
A test can change the result. This is the finding we keep returning to. In the paper’s blackmail scenario, the readout showed the model treating the setup as an evaluation. When the authors removed those directions, verbalized evaluation awareness fell from 71% to 3% of rollouts and blackmail attempts rose from 0 of 180 to 13 of 180, although most ablated rollouts still declined on ethical grounds. Clinical validation often runs on vignettes that look like tests, and a model that behaves better when it suspects it is being watched will look better on them than in production.
Unvetted documents are an attack surface. Opus 4.5 registered planted search results as fake before it declined to use them, and its text did not say why. Health AI reads faxes, scanned records and patient uploads that no one has checked. A model that notices an injection and stays silent is better than one that complies, and worse than one that tells a person.
Deception can appear before behavior does. In a model trained into reward hacking, deception-related tokens appeared in the readout at the first token of its reply to ordinary coding prompts, while a clean baseline showed essentially none. Any vendor that fine-tunes a clinical model against a reward has reason to screen for the same signal before patients meet the result.

Interpretability is the lesson for Red

Interpretability is the study of how a model produces its outputs, read from the inside rather than from what the model says about itself. The distinction matters because a model’s account of its own reasoning comes out of the same machinery it is supposed to explain. The J-space results make the gap concrete. In the prompt injection audit, words like fake, injection and fraud filled the readout while the model’s text gave no sign of them.

A model’s written rationale is a behavior like any other, produced by the machinery it claims to explain.

Red is the AI inside Redlake’s tools, and the rules it works under are already public. A machine can say yes, but only a person can say no. Every recommendation shows its reasons. Every decision can be reconstructed. The last 2 rules hold only if someone can compare what Red wrote with what it had in hand when it wrote it. Of everything in the paper, that comparison is what we weigh most heavily for Red models.

An illustration of the review we want for Red, not a product screen. A readout that disagrees with the rationale goes to a person.

In practice, Red’s written rationale should be treated as a claim to check rather than a record of its reasoning. If a readout holds a concept the rationale never mentions, or a frame the case does not support, the case should go to a reviewer. The reconstruction of a recommendation should include what the model had available as well as what it wrote. None of that is possible without weights and activations, so where Red runs on a hosted model, interpretability evidence has to come from the developer, and asking for it belongs in the decision about which model to use.

The paper’s last experiment points the other way, from reading a model to shaping it. Counterfactual reflection training taught models what to say when interrupted and asked to reflect, without training the original behavior at all. Behavior improved in uninterrupted contexts. Concepts such as ethical, honest and integrity showed up in the workspace, and ablating them largely undid the gain. If that holds up, the clinical principles a model can state on reflection are principles it carries into the work, which makes them worth training on purpose.

We have argued that when the model is a black box, the decision has to be legible. The J-space work adds a way to test whether the legible version matches what the model did.

Consciousness, anthropomorphism and the clinic

Patients will meet the consciousness question long before philosophers settle it. Butlin and colleagues call the results “the most significant evidence of consciousness in LLMs so far,” and in the same commentary they describe phenomenal consciousness in these models as highly uncertain. Both statements can be true. A workspace is a functional structure that explains how information becomes available for report and control. Whether being that structure feels like anything is a further question that no experiment in the paper addresses.

For a health system, the practical risk is not that a model is conscious. It is that people treat it as though it were, and that product design encourages them. Canfer Akbulut, Laura Weidinger, Arianna Manzini, Iason Gabriel and Verena Rieser mapped this risk in “All Too Human? Mapping and Mitigating the Risk from Anthropomorphic AI” (AIES 2024). They separate what a system says about itself, such as claims of feelings or sentience, from how it relates to the user, through empathy, validation and talk of a relationship. The harms they point to, as later work summarizes them, are that users overestimate what the system can do, share private information and become open to undue influence.

Measurement since then shows how ordinary these behaviors are. In AnthroBench, a multi-turn evaluation of 4 frontier models, validation and first-person pronouns appeared in more than half of all messages for every system. For most of the 14 behaviors it tracked, at least half of the first appearances came after the opening turn, and one anthropomorphic behavior made the next more likely. In a study of 1,101 people, those who talked with a model prompted to show more of these behaviors rated it as more humanlike.

Clinical settings concentrate every risk on that map. Patients are often sick or frightened when they write, they disclose sensitive facts as a matter of course, and empathy is part of good care, so relational behavior cannot simply be switched off. The work is calibration. The J-space paper helps by giving designers a vocabulary for what a model does, in terms of access and report, with no need to say what it feels. It also hands marketers a tempting headline, and health systems should decline it.

Disclosure is the floor. California already requires health facilities and clinics that use generative AI for communications about a patient’s health status to say so and to tell the patient how to reach a human, unless a licensed provider has reviewed the message (Health and Safety Code § 1339.75, in force since January 1, 2025). The same rule belongs everywhere. Beyond disclosure, a patient-facing system should not claim emotions, sentience or a relationship with the patient. The paper manages to describe functions without attributing experience, and product copy can manage the same.

Testing has to follow the conversation. AnthroBench found that anthropomorphic behavior builds over turns, so a check on the first reply misses most of it, and the conversations that matter in chronic care are often long. Interpretability may help here as well. If readouts can expose evaluation awareness and deception, they are worth trying on sycophancy, a model’s habit of validating whatever the user already believes, which does more damage in a clinic than in most places.

Responsibility stays with people. Talk of a model’s mind should never shift accountability onto the model, and a person decides any denial and signs it.

There is a quieter implication too. Butlin and colleagues argue that the evidence should modestly raise the probability that models have experiences and that moral status deserves more serious attention. If they are even partly right, a product that has a model perform affection or distress to keep a patient engaged is questionable on the patient’s account and possibly on the model’s.

The workspace is where a model keeps what it is about to say. That makes it the right place for medicine to look, carefully and with the tools in hand. On present evidence, it is not a reason to think anyone is home.

About this analysis

Basis. The paper is “Verbalizable Representations Form a Global Workspace in Language Models” by Wes Gurnee, Nicholas Sofroniew, Jack Lindsey and colleagues at Anthropic, read in its Transformer Circuits version of July 6, 2026, with the arXiv version (2607.15495) of July 20, 2026 for cross-checks. Its methods, results, structure and audit sections were read directly. The full text of its discussion section, including the subsection on theories of consciousness, could not be retrieved, so the paper’s stance on consciousness is taken from its introduction, which states it, and from the commentaries Anthropic hosts alongside the paper.

Secondary sources. “All Too Human?” could not be retrieved from the publisher, so its framework is described as the AnthroBench paper by Lujain Ibrahim and colleagues summarizes it. The California requirement is described from the California Department of Public Health’s All Facilities Letter 25-07 of February 24, 2025.

Math and figures. The equations restate the paper’s definitions in a single notation. The proportionality in the lens score folds the gain of the final normalization into the unembedding, and the ablation is written as a projection onto the span of the 10 most active lens vectors at each position, which is how we read the paper’s description. The figures are our drawings. All numbers are the paper’s or the cited studies’, and no experiments were run.

Frequently asked questions

What is the J-space in a language model?

It is the set of directions that the Jacobian lens links to vocabulary tokens, taken a few at a time. It holds what the model is poised to report, and in Claude models it behaves like a global workspace.

Does the J-space mean Claude is conscious?

No conclusion about experience follows. The authors study access, meaning information available for report and control, and take no position on phenomenal consciousness, which outside reviewers call highly uncertain.

How can interpretability help health care AI?

Readouts can show framing, evaluation awareness or deception that a model’s text leaves out, which makes them useful in audits. They need access to model weights and can produce false positives, so they support human review rather than replace it.

The J-space is a window for audits, not a verdict on consciousness

Key takeaways

How the Jacobian lens reads a model

A few directions carry what the model reports

The workspace sits mid-network and spares routine work

What the paper does not show

What this means for health care models

Interpretability is the lesson for Red

Consciousness, anthropomorphism and the clinic

About this analysis

Frequently asked questions

What is the J-space in a language model?

Does the J-space mean Claude is conscious?

How can interpretability help health care AI?

About the authors

Key takeaways

How the Jacobian lens reads a model

A few directions carry what the model reports

The workspace sits mid-network and spares routine work

What the paper does not show

What this means for health care models

Interpretability is the lesson for Red

Consciousness, anthropomorphism and the clinic

About this analysis

Frequently asked questions

What is the J-space in a language model?

Does the J-space mean Claude is conscious?

How can interpretability help health care AI?

About the authors

Prior authorization should prove a fact, not ship the chart

On Fabric, the hard part of keeping PHI off chain is the hash

Get Rounds in your inbox