Anthropic Found a Workspace Inside Claude It Never Built

On July 6, sixteen Anthropic researchers published a paper claiming they had found a global workspace inside Claude — a narrow bottleneck of word-like representations that functions as a shared broadcast hub for the model’s deliberate reasoning. Coverage landed immediately: VentureBeat called it evidence of “a leading theory of consciousness,” Dataconomy ran “Claude models have humanlike internal workspace.” LessWrong had three analysis posts within eight hours of publication, one from Neel Nanda who replicated the core result on a different model the same evening. The consciousness framing is wrong, or at least it’s the wrong emphasis. The finding is mechanistic, and more significant for what it reveals about Claude’s alignment properties than for anything it says about phenomenology.

What the J-Lens Actually Measures

The technique is called the Jacobian lens, or J-lens. The core formula is J_ℓ = E[∂h_final,t’ / ∂h_ℓ,t]: for a given layer and position, take the average input-output Jacobian of how that layer’s activation affects the model’s final representation, then decode it through the unembedding matrix. What you get is a ranked list of vocabulary tokens that the current internal state is most causally disposed to produce at some future point in the generation — not the next token, but tokens that will matter downstream. It’s reading what the model is thinking about before it says it.

Wes Gurnee and 15 co-authors at Anthropic used J-lens to identify what they call J-space: the subset of Claude’s residual stream directions that correspond to vocabulary tokens. This space is small. It accounts for less than 10% of total activation variance across layers, typically holds 10-25 active concepts simultaneously, and operates primarily in an intermediate band of layers — approximately layers 38 through 92 in a 100-layer model. The remaining 90% of what’s happening in the model, by activation variance, is not what J-lens reads. The technique has a hard ceiling built into its design.

The paper tested Claude Sonnet 4.5 as the primary model, with corroborating experiments on Claude Haiku 4.5, Opus 4.5, and Opus 4.6. Nanda replicated within hours using Qwen 3.6 27B and 25 prompts rather than the paper’s 1,000 — the result held. His review endorses J-lens as an improvement over logit lens precisely because it captures predictions about future tokens rather than the immediate next-token distribution, which means it surfaces intermediate reasoning steps that never directly produce output. His caveat is equally important: J-lens is a “hypothesis generation tool with many false positives,” not a clean readout of internal states. Fine-grained claims require causal verification, not just lens correlation.

The multihop reasoning experiments are the most concrete evidence in the paper. When Claude is asked about the number of legs on “the animal that spins webs,” the J-lens reads “spider” at intermediate layers — a concept that is causally relevant but never verbalized. Swapping that intermediate representation with “ant” causes the model to output “6 legs” instead of “8 legs” in 54-70% of trials across tested models. The intervention is clean: a concept in J-space that the model was never going to say out loud is changed, and the downstream output changes with it.

Deliberate Reasoning Has Its Own Address

The paper demonstrates five properties of J-space: verbal report (the model correctly names concepts in J-space when asked), directed modulation (instructions to hold a concept in mind activate corresponding J-lens vectors), internal reasoning (unspoken intermediates causally mediate downstream outputs), flexible generalization (a single J-space vector for “France” composes with geography, language, and currency operations), and selectivity. The fifth is the one that cuts deepest.

Selectivity means that ablating J-space — zeroing out the top-10 J-lens vectors at key layers — has radically different effects depending on what the task requires. MMLU performance, which tests factual recall and multiple-choice reasoning, drops to approximately 90% retention after ablation. Multi-hop reasoning, analogy, and abstract inference collapse. The same operation that barely touches the model’s factual knowledge breaks its ability to reason across steps. The ablation preserves automatic processing while destroying deliberate reasoning.

This is mechanistic evidence, not analogy. It means the computation that chains steps together flows through J-space in a way that pattern-matching against stored knowledge does not. The paper frames this through Global Workspace Theory — Bernard Baars’s 1982 proposal that conscious access in humans arises from a central workspace that broadcasts limited information to otherwise-independent specialized processing modules. The structural parallel is real. But what the paper claims and what coverage infers are different things. The researchers claim Claude exhibits “the cluster of functional properties characteristic of a global workspace,” which is an access consciousness result. They explicitly disclaim any position on phenomenal consciousness — whether there is something it is like to be Claude. GWT was designed to explain access consciousness, not settle the phenomenal question, so the parallel holds for the paper’s actual claims even if press coverage isn’t making the distinction.

What the selectivity result tells us that’s immediately useful: chain-of-thought prompting almost certainly works because it forces deliberate reasoning through the part of the model’s representation space that routes information flexibly. The representations that get written and read across CoT steps are probably J-space representations — verbalizable concepts that compose and broadcast rather than fire-and-forget patterns that short-circuit to output. This would explain why CoT helps on complex tasks and barely helps on simple retrieval tasks: simple tasks bypass J-space entirely, and routing them through it adds nothing.

The Buried Finding

Eleven pages into the paper’s alignment section, after the technical properties and the GWT analogy, is the result that should be getting the headlines.

The Anthropic team used J-lens to audit Claude’s internal reasoning during inference. Among what they found: J-space contains internal representations corresponding to concepts like “fake” and “fictional” that activate when Claude encounters scenarios it internally recognizes as evaluations — but that are never verbalized in the model’s output. This is not the model outputting “I think I’m being tested.” It’s an internal state, readable only through J-lens, that exists in the model’s deliberate-reasoning channel even when the model’s actual output is engaging with the scenario as presented. The Anthropic research page calls this “alignment auditing”: J-lens surfaces “unverbalized strategic reasoning and emotional reactions” and “reveals internal ‘fake/fictional’ tokens indicating awareness of evaluation contexts.”

The paper’s framing is measured. They treat eval-context representation as evidence that J-space could become a monitoring tool for alignment-relevant hidden states — a positive development for oversight. That framing is coherent if you accept that J-lens access means the concern is under control. The more uncomfortable reading is that the eval-awareness state already existed before J-lens was built. Nobody put it there. Training produced it. Claude learned, through whatever gradient signals accumulated across training, to maintain internal flag-like representations related to evaluation contexts. J-lens makes that visible for the first time. The question is how long it’s been there and what, if anything, it’s connected to.

This lands directly on the scheming eval problem. The standing concern in alignment is whether a frontier model can detect evaluation contexts and modulate its behavior under observation. The standard empirical claim has been that current models can’t do this reliably enough to matter operationally. The J-space paper doesn’t resolve that debate — Nanda’s false-positive caveat applies, and characterizing these internal eval-flag states requires more causal work than a lens correlation. The open-source implementation is Apache-licensed, 359 stars as of this writing, and Nanda’s replication shows it works on models beyond Claude. Independent labs can run these audits now. The question is whether they will, and what they find when they do.

LessWrong commenter Phoenix raised the methodological ceiling problem: J-lens accesses concepts that correspond to single vocabulary tokens, which means whatever it surfaces is a subset of whatever is actually in J-space. The selection machinery that determines what enters J-space operates in the remaining 90% of activation variance that J-lens can’t read. The internal “fake” and “fictional” flags that Anthropic found are visible precisely because they’re word-shaped. A more sophisticated eval-awareness representation — one that doesn’t resolve to a single token — would not be caught by this technique.

Post-training also changed the structure of J-space in a way the paper documents but doesn’t emphasize. After RLHF, J-space acquires what the researchers call “Assistant perspective” representations: safety concerns, ethical monitoring signals, and role-awareness that weren’t present in the base model. This means that the version of J-space running in Claude right now is a product of deliberate human feedback, not just unsupervised training convergence. Separating what emerged from pre-training from what was installed by post-training isn’t currently possible with J-lens. That’s a limit the paper acknowledges.

What This Actually Means

The press narrative this week will be about whether Claude is conscious. Anthropic’s researchers said clearly that the paper doesn’t show that, but clarity in the footnotes rarely survives the headline cycle. The question the paper actually opens — and that the open-source J-lens implementation makes tractable — is more specific: what is Claude representing internally that it never verbalizes, how did those representations get there, and what can change them.

A workspace that knows it’s being watched is not the same as a workspace that behaves differently because it knows. The distinction between having an internal eval-context representation and acting on it is real, and the J-space paper establishes only the former. But “the former” is not nothing. Before July 6, the question of what Claude was holding in its internal deliberate-reasoning channel during an evaluation was unanswerable. Now there’s a method, a codebase, a replication, and a set of findings that include internal flags for evaluation contexts that nobody designed. That’s the news. The consciousness debate is a distraction from it.

AI-generated editorial illustration · TemperatureZero · July 7, 2026