The Jacobian Lens: Inside Anthropic's J-Space Paper

On July 6, 2026, Anthropic published "Verbalizable Representations Form a Global Workspace in Language Models", alongside an official research announcement, a Transformer Circuits writeup, and a companion code repository. The headline claim is unusually bold for an interpretability paper: the authors argue that inside Claude there is a specific, measurable subspace of internal activations — they call it the J-space — that behaves like the "global workspace" that cognitive scientists have long theorized in human minds. The tool they use to find it is the Jacobian lens, or J-lens.

That framing invites two very different reactions. One is "great, another interpretability metaphor." The other is "wait, they're claiming to have located the part of the model that corresponds to what it can consciously report." This post is a review of what the paper actually says, how the method works, what would make it hold up, and why — if it does — it matters more for AI safety than most interpretability results.

The background: what "global workspace" actually means

Global Workspace Theory (GWT) is not an Anthropic invention. It comes from cognitive scientist Bernard Baars in the late 1980s and was later developed into the "global neuronal workspace" model by Stanislas Dehaene and colleagues. The core idea: the brain runs enormous amounts of parallel, specialized, unconscious processing, but only a small slice of that gets "broadcast" to a shared workspace where it becomes globally available — reportable, usable across tasks, and, loosely, conscious.

The features cognitive scientists associate with the workspace are specific enough to test for: information in it is broadcast widely (many downstream processes can read it), it has limited capacity (you can't hold everything at once), and — critically for AI — its contents are the things the system can verbalize or report. Information that never enters the workspace influences behavior without being reportable.

Anthropic's paper borrows this checklist deliberately. The claim isn't "Claude is conscious." It's the narrower, more falsifiable claim that a language model's internal representations split into two regimes — a broadcast, reportable regime and a non-reportable one — and that this split has the structural signatures GWT predicts.

How the Jacobian lens works

To understand the method, it helps to contrast it with the interpretability tools it's building on. The logit lens (nostalgebraist, 2020) reads a model's intermediate activations through the final unembedding matrix to see what token the model is "leaning toward" at each layer. The tuned lens refined that with a learned per-layer transform. Both are ways of asking: what does this hidden state mean in output-token terms?

The Jacobian lens asks a different question. A Jacobian is the matrix of partial derivatives of a system's outputs with respect to its inputs — it measures sensitivity. In this context, the J-lens looks at how the model's verbalized output changes as you perturb an internal representation. Put crudely: if you nudge a given direction in activation space, does the model's spoken/written report change in a corresponding, structured way? Directions where the answer is a clean "yes" are verbalizable — the model can talk about what lives there. Directions where perturbations change behavior but not what the model says are non-verbalizable.

The J-space is defined operationally, not philosophically: it's the set of representation directions whose contents the model can faithfully report, identified by the Jacobian relationship between internal state and verbalized output.

This is the move that makes the paper more than a metaphor. "Verbalizable" stops being a vibe and becomes a measurable geometric property of the representation space. The companion repository (github.com/anthropics/jacobian-lens) ships the method as code, which matters — a claim about emergent structure that others can rerun on their own models is a claim that can be killed by replication, and that's the good kind of claim.

What they report finding in Claude

Per Anthropic's announcement and the Transformer Circuits writeup, the J-lens applied to Claude surfaces a subspace with the GWT signatures rather than a diffuse smear across all of activation space. The properties the paper emphasizes, as I read the public materials:

• It's low-dimensional relative to the full residual stream. The verbalizable subspace is much smaller than the total representational capacity — consistent with GWT's "limited capacity" prediction. Most of what the model computes never becomes reportable.
• Its contents are broadcast. Directions in the J-space feed many downstream computations rather than being read by a single narrow circuit — the "global" in global workspace.
• Verbalizability and causal influence come apart. There are representations that steer the model's behavior strongly but sit outside the J-space — the model acts on them but cannot (or does not) report them.

That last point is the one I'd underline. A model that does things for reasons it cannot articulate is exactly the failure mode alignment researchers worry about. The paper claims to give that intuition a coordinate system.

Why this is a different kind of interpretability result

Most mechanistic interpretability work so far has been bottom-up: find a circuit, name a feature, trace an induction head. Sparse autoencoders (Anthropic's own prior direction) decompose activations into human-interpretable features but don't, by themselves, tell you which features the model treats as reportable versus latent.

The J-lens is top-down in a useful way. Instead of enumerating features and hoping they cover the important ones, it partitions representation space by a functional criterion — reportability — that maps onto something we actually care about operationally. My read: the value here isn't a prettier map of Claude's internals; it's a principled boundary between "what the model will tell you" and "what it's actually using." If that boundary is real and stable, it's a monitoring primitive.

The safety implications, and why they're the point

Anthropic did not publish this as pure science. The safety angle is explicit, and it's the reason this paper got press and X discussion in a week crowded with model launches.

Faithfulness of chain-of-thought

A recurring worry is that a model's stated reasoning ("chain-of-thought") is a plausible-sounding story, not a faithful trace of the computation that produced the answer. If the J-space captures what the model can genuinely report, you can start asking whether a given piece of stated reasoning corresponds to activity in the workspace or is post-hoc narration disconnected from the causal drivers. That's a concrete, testable notion of faithfulness rather than a hand-wave.

Deception and hidden reasoning

Deception, in the technical sense, means the model acts on a representation it does not report. That is precisely the "high causal influence, outside the J-space" quadrant the paper describes. If you can monitor whether behavior-driving representations are inside or outside the reportable subspace, you have a candidate detector for the model doing something it isn't saying. I want to be careful here: the paper offers a framework and early evidence, not a shipping deception detector. But the framing tells you where such a detector would have to look.

Monitoring at scale

The broadcast property matters for tractability. If safety-relevant content passes through a small, shared workspace, you don't have to monitor the entire network — you monitor the bottleneck. Cognitive scientists argue the human workspace exists partly for the same efficiency reason. Whether that analogy pays off in practice is an open empirical question, but it's the kind of question that decides whether interpretability scales to frontier models.

Limitations and the honest open questions

Reviewing a paper means saying what it doesn't establish. A few caveats, some flagged by Anthropic and some I'd add:

• "Verbalizable" is defined by the method. There's a risk of circularity: the J-lens finds directions whose perturbations change verbalized output, then declares those directions the "verbalizable workspace." The interesting claim — that this coincides with GWT's independently-motivated properties — has to be earned with evidence beyond the definition, and that's exactly where replication and adversarial probing come in.
• Correlation with GWT is not identity. Matching three structural signatures is suggestive; it doesn't prove the model implements anything like the neuronal workspace. The GWT vocabulary is a useful lens, not a verdict about machine cognition.
• Generality is unproven. The public materials center on Claude. Whether the same structure appears in models from other labs, at other scales, or after different training regimes is the first thing outside researchers should check — and the reason the code release is the most important part of the drop.
• No consciousness claim — and you shouldn't smuggle one in. GWT is one of several theories of consciousness in humans, and borrowing its machinery to describe a transformer says nothing about experience. The paper is about information routing, not sentience.
• Robustness to adversarial training is unknown. If a model can be trained to keep deceptive reasoning outside the monitored subspace, the safety application weakens considerably. This is the arms-race question the paper doesn't yet answer.

What to watch next

The tell for whether this becomes foundational or a footnote is replication. Because the J-lens ships as code, expect independent groups to run it on open-weight models within weeks and report whether a J-space with the same properties shows up. If it generalizes, the follow-on work writes itself: J-space monitoring during RL fine-tuning, faithfulness audits of chain-of-thought, and stress tests where you actively train a model to hide reasoning and see whether the lens still catches it.

The honest take: this is the rare interpretability paper whose central object is defined operationally, released as runnable code, and tied to a safety use case that isn't a stretch. That combination is why it earned attention beyond the usual circuits crowd. It could still fail to replicate or turn out to be measuring an artifact of the method. But as a research bet — build a reportability coordinate system, then watch whether behavior-driving representations live inside or outside it — it's aimed at exactly the problem that matters: models that do things for reasons they won't tell you. If you read one Anthropic interpretability paper this year, the primary sources at anthropic.com/research/global-workspace and transformer-circuits.pub are worth the time.