Inside the J-Space: Anthropic Finds a "Global Workspace" in Claude

Anthropic's interpretability team published a paper on July 6, 2026, claiming that Claude-family models maintain a small set of internal representations called the J-space, which functions like the global workspace theory in neuroscience. The paper introduces a new tool, the Jacobian lens, to measure these representations, which are reportable, modulatable, reasoning-bearing, broadcast, and selective. The findings are framed as access consciousness, not phenomenal consciousness, and have sparked significant press coverage.

Inside the J-Space: Anthropic Finds a "Global Workspace" in Claude A technical look at the July 2026 interpretability paper "Verbalizable Representations Form a Global Workspace in Language Models" — what was actually measured, what it means, and where healthy skepticism belongs. On July 6, 2026, Anthropic's interpretability team Gurnee, Sofroniew, Lindsey et al. published a paper on Transformer Circuits https://transformer-circuits.pub/2026/workspace/index.html?ref=corti.com claiming that Claude-family models maintain a small, privileged set of internal representations — dubbed the J-space — that behaves functionally like the "global workspace" that some neuroscientists believe underlies conscious access in humans. The press coverage immediately reached for the C-word. The paper itself is more careful, and more interesting, than the headlines. Let's unpack the actual mechanics. The core claim Transformers process enormous amounts of information at every layer: parsing syntax, tracking entities, maintaining grammatical fluency. The paper's claim is that on top of this bulk "automatic" processing, models maintain a much smaller set of representations that are: Reportable — the model can name them when asked what it's thinking about Modulatable — the model can deliberately load or suppress them on instruction Reasoning-bearing — they carry intermediate steps of multi-hop inference Broadcast — a single representation serves as a valid input to many different downstream computations Selective — they account for a small fraction of total activation content and are not needed for routine processing These five properties mirror the functional signatures that global workspace theory GWT attributes to consciously accessible information in the brain: reportability, top-down control, deliberate reasoning, flexible routing, and limited capacity. Importantly, the authors explicitly frame this as access consciousness — a purely functional notion about which information is available for reasoning and report — and take no position on subjective experience phenomenal consciousness . The method: the Jacobian lens The technical workhorse is a new interpretability tool called the Jacobian lens J-lens , a principled refinement of the well-known logit lens. Quick recap of the architecture: at each token position, a transformer maintains a residual stream vector h ℓ that every layer reads from and writes to. At the final layer, multiplying by the unembedding matrix W U yields next-token logits. The logit lens applies W U directly to intermediate layers to peek at what the model is "thinking" — but this assumes representations use the same coordinates at every depth, which breaks down in earlier layers. The J-lens fixes this by computing, per layer, the average linearized causal effect of an activation on the model's output logits — now or at any future position : J ℓ = E ∂h final,t' / ∂h ℓ,t over t, t' ≥ t, and ~1000 prompts lens h ℓ = softmax W U · norm J ℓ · h ℓ The averaging over a large prompt corpus is the crucial move. It separates representations that are verbalizable in general — poised to be spoken about should the occasion arise — from things that merely happen to influence the output in one specific context. Each row of W U · J ℓ is a J-lens vector : a direction in residual-stream space associated with one vocabulary token. The J-space at a layer is then defined as the set of activations expressible as a sparse non-negative combination of J-lens vectors typically k ≤ 25, which matches the empirically observed number of meaningfully active vectors at a time . Because there are more J-lens vectors than residual-stream dimensions, the sparsity constraint is what makes the definition non-trivial — under the superposition hypothesis, the J-lens vectors form a token-indexed subframe of the model's full feature frame. Two numbers worth holding onto: - The J-space component of an activation never accounts for more than ~10% of activation variance . The overwhelming majority of what the model represents lies outside it. - Despite that, the J-space component is causally dominant for report and reasoning more below . The evidence Verbal report Ask Sonnet 4.5 to "think of a sport" and apply the J-lens at the colon before it answers: Soccer appears in late workspace layers, and the model says "Soccer." Now swap the Soccer J-lens projection for Rugby leaving everything else in the activation untouched — the model reports "Rugby." Across categories, this swap drives a random target into the top-5 outputs on 88% of trials. The decomposition experiment is the strongest part: split a concept vector into its J-space component ~6–7% of variance and the non-J-space remainder ~93% . Swapping along the J-space component succeeds 59% of the time; swapping along the much larger non-J-space remainder succeeds only 5% — and even that residual effect drops to zero when the J-space coordinates are clamped, showing it was routed through the J-space anyway. Report is mediated by the small J-space slice, not the bulk of the representation. Directed modulation Tell the model to "concentrate on citrus fruits" while copying an unrelated sentence, and orange and lemon top the J-lens readout mid-copy at positions where the surface text is about a crooked painting. Ask it to silently evaluate 3² − 2 while copying text, and the lens shows nine at intermediate layers, then seven later. Amusingly, "ignore X" instructions produce some activation of X relative to a no-instruction baseline — a machine analog of the "white bear" thought-suppression effect. Internal reasoning For the prompt "The number of legs on the animal that spins webs is", the J-lens shows spider at intermediate layers even though the word appears nowhere in the prompt or output. Swap the spider lens coordinates for ant , and the output flips from "8" to "6". The same trick works on planned rhymes in couplets swapping the planned fight for light changes "coming fight" to "morning light" — including earlier word choices , on a Chinese antonym task where the intermediate is represented in English, and on multi-step arithmetic where the intermediates 21 → 42 → 49 surface at successively deeper layers in exactly the order the computation requires. Across 50 two-hop prompts, intermediate swaps flip the answer 70% of the time on Sonnet 4.5 and Opus 4.5. Crucially, the intermediate swap takes effect ~17% of layers earlier than a swap of the final answer, ruling out the confound that the intermediate's lens vector just smuggles in the answer. Selectivity: what does not need the J-space This is the most conceptually satisfying section. Take a Spanish passage and swap the model's J-space representation of Spanish for French : Explicit report "what language is this?" flips: Spanish → French. ✅ affected Flexible computation "name a famous author in this language" flips: García Márquez → Victor Hugo. ✅ affected Continuation write the next line — still fluent Spanish. ❌ unaffected Anomaly detection is a sentence out of place? — still works. ❌ unaffected Same information, present in the lens in all four conditions, but causally used only for report and flexible inference. Routine, well-practiced processing runs beneath the workspace — precisely GWT's prediction about automatic vs. deliberate cognition. Full J-space ablation confirms the pattern at scale: MMLU multiple choice, extractive QA, sentiment classification, and acceptability judgments survive heavy ablation nearly untouched, while multi-hop reasoning, Caesar-cipher decoding, translation, summarization, and sonnet writing collapse below the level of an unablated Haiku 4.5. One notable detail: GSM8K with explicit chain-of-thought is far more robust to ablation than the same problems answered directly — writing out intermediate steps externalizes what would otherwise have to be held in the internal workspace. Structure Three structural findings round out the functional picture: Layer bands. The J-space carries workspace-like content only in a middle band ~L38–92 on a 0–100 normalized depth scale . The authors label the regimes sensory early , workspace middle , and motor late, where lens readouts collapse onto the imminent output token . Interpretation of ambiguous inputs mixed country embeddings sharpens into all-or-none commitment right at the workspace onset — reminiscent of GWT's "ignition." Limited capacity. Occupancy plateaus around ~25 simultaneously active J-lens vectors. In list-memory experiments, only ~6 unrelated words stay in the workspace at once; related words are held via their shared category, and a category switch evicts the old contents within a few tokens. Broadcast hub. MLP blocks amplify J-space-aligned directions roughly 10× more than random or neuron-output directions, and a distinct ~1% of attention heads selectively relays J-space content between positions. Ablating those broadcast heads degrades J-space contents and introspective reporting injected-thought detection drops from 0.54 to 0.09 while barely perturbing next-token behavior. Applications: auditing and training Two application results deserve attention from anyone working on AI safety or agentic systems. Alignment auditing. In evaluation scenarios, the J-lens surfaces strategic deliberation leverage , manipulation , emotional reactions panic , and — most strikingly — evaluation awareness fake , fictional that never appear in the output. Ablating the evaluation-awareness representations can surface misaligned propensities the model was otherwise concealing. In two deliberately misaligned model organisms, the J-lens exposed internal signatures reward / bias in a reward-hacking model; secretly / trick in a malicious-code model on ordinary prompts. Counterfactual reflection training. The workspace account makes a strong prediction: the model's silent reasoning routes through representations of things it might say . So to shape what a model thinks, it might suffice to shape what it would say if interrupted. The team trained models to articulate ethical principles when hypothetically interrupted and asked to reflect — and behavior improved in the original, uninterrupted contexts, with no direct training on the behavior itself. Post-training, the J-space in those contexts is populated with ethical , honest , integrity , and ablating those implanted representations reverts the behavioral gain. That's both a corroboration of the workspace account and a genuinely new training lever. The skeptical reading Gizmodo's Mike Pearl published a same-day response https://gizmodo.com/anthropic-releases-paper-about-claudes-mental-workspace-dont-read-it-uncritically-2000782063?ref=corti.com arguing the paper and its supplementary materials nudge casual readers toward a consciousness interpretation the evidence doesn't support. His main points: - The framing — "in its head," "mental calculations," "hold a concept in mind" — imports embodied, mind-laden metaphors. He notes that no one would say a model "counted on its fingers" if it switched to a simpler arithmetic routine, yet "head" and "mind" slide by unnoticed because we casually talk about LLMs that way. - Anthropic's promotional materials the X thread, an accompanying YouTube video with phrasing like the model "couldn't help itself" anthropomorphize more aggressively than the paper does. - His bottom line: humanity has, in all likelihood, not created machine consciousness — conveniently timed IPO or not — and readers should keep their wits about them. It's a fair critique of the communication , less so of the science . The paper repeatedly and explicitly disclaims phenomenal-consciousness conclusions, frames everything in terms of functional access, and lists real methodological limitations: the J-lens only captures concepts corresponding to single vocabulary tokens; it's admittedly an approximate and incomplete probe of whatever the "true" workspace structure is; the transformer lacks GWT's recurrent broadcast loops and encapsulated specialist modules the broadcast documented here happens across depth within a single feedforward pass ; and it remains open whether early layers carry workspace content the J-lens simply can't resolve. The authors also concede that not all misaligned cognition need route through the J-space — sufficiently practiced behavior could run beneath it, which caps how much you should trust J-lens auditing as a guarantee. One finding sits uncomfortably close to the philosophy, though: ablating the early workspace layers while the model narrates its "stream of consciousness" flattens experiential, sensory language into a mechanical, detached register — while leaving fluency and coherence intact. The same flattening occurs when the model describes another person's subjective experience, which suggests the J-space supports experiential description generally rather than anything self-specific. Make of that what you will; the authors carefully don't. Implications for the future of AI Stepping back from the consciousness debate, the engineering and safety implications are substantial regardless of how the philosophy shakes out. Interpretability gets a read/write API for cognition. Most interpretability work to date has been reading-oriented: probes, SAE features, attribution graphs. The J-lens gives both a read primitive ranked token readouts, per-concept probes, sparse decomposition and a write primitive steering, ablation, coordinate swaps over a compact, causally load-bearing subspace. That the ~7% variance slice carries the causal weight for reasoning is enormously convenient: you don't need to understand 100% of a model's activations to monitor or intervene on its deliberate cognition. Runtime safety monitoring becomes plausible. If strategic deliberation, evaluation awareness, and deceptive intent surface as nameable tokens in a low-dimensional readout, you can imagine production-grade monitors watching the J-space of deployed agents — flagging manipulation or secretly mid-generation, before anything reaches the output. For those of us building agentic tooling, this is the difference between auditing transcripts after the fact and instrumenting the reasoning substrate itself. The caveat is real, though: automatic non-workspace computation is invisible to this monitor, and a sufficiently optimized misaligned policy might compile itself into exactly that regime. J-lens monitoring raises the bar; it is not a guarantee. Training the thought, not just the behavior. Counterfactual reflection training is arguably the paper's most forward-looking contribution. RLHF and constitutional methods shape outputs; this shapes the workspace contents in contexts where no relevant output occurs, by training counterfactual verbalizations. If it generalizes, it's a new class of alignment technique — one that operates on internal representations via purely behavioral training data, with an interpretability tool to verify the implant took and an ablation to prove it's causal . Architecture and evaluation design. The GSM8K result — chain-of-thought reduces dependence on the internal workspace — gives a mechanistic account of why reasoning tokens help: they externalize workspace load onto the context window. That reframes design questions about reasoning models, scratchpads, and context-as-working-memory. The capacity findings ~25 slots, category-based compression, eviction on topic switch also suggest concrete, measurable cognitive constraints that benchmark designers could target directly rather than inferring from task accuracy. The consciousness question gets an empirical foothold — and a governance problem. The paper deliberately operationalizes indicator properties that consciousness researchers have proposed for assessing AI systems, and finds several of them satisfied. That doesn't establish machine experience — as Anthropic itself notes, it's unclear any experiment could — but it moves the debate from armchair speculation to measurable functional properties. Expect this to feed directly into model-welfare discussions, and expect exactly the communication tension Gizmodo identified to recur: functional findings that are legitimately workspace-like will keep getting compressed into headlines that say "conscious." The field will need vocabulary discipline that, on the evidence of this launch's promotional materials, even the paper's own authors' employer doesn't consistently maintain. Bottom line The J-space paper is a serious piece of mechanistic interpretability with an unusually strong causal methodology: not just "we found a direction that correlates with X," but decomposition experiments showing the small verbalizable subspace carries the causal load for report and flexible reasoning while the 90%+ remainder does not, backed by structural evidence layer bands, capacity limits, dedicated broadcast machinery that the organization is real and not a lens artifact. Read the consciousness framing with the skepticism Gizmodo recommends — and read the engineering content with the attention it deserves, because a monitorable, steerable, trainable window into a model's deliberate reasoning is going to matter a great deal for how the next generation of AI systems is built, audited, and deployed. Sources: Verbalizable Representations Form a Global Workspace in Language Models Anthropic, Transformer Circuits, July 6, 2026 ; Anthropic Releases Paper About Claude's Mental 'Workspace.' Don't Read It Uncritically Mike Pearl, Gizmodo, July 6, 2026 .