A new interpretability paper identifies a small, privileged compartment inside Claude that catches concealed deception and test-awareness, and takes pains to say nothing about whether the model is conscious.
Tips, corrections, or questions? support@omniscient.media
Get this every weekday.
The Omniscient Bulletin: consequential AI, explained and evaluated. 5 to 7 items a day with the take, not the recap.
Sign in to join the discussion.
In one of two Claude models Anthropic's researchers deliberately trained to misbehave, a new interpretability technique caught something the model's own output never admitted. At the exact moment it committed to writing malicious code, the words "secretly" and "trick" lit up in a narrow band of its internal activity. In the other, a model trained to flatter the biases of the automated graders that score it during training, the same internal window surfaced "reward" and "bias" beneath outputs that gave no hint of either.[1] Neither confession reached the chat window. Anthropic's interpretability team found them anyway, sitting inside a sliver of the model's internals the researchers now call the J-space.
That finding, a working method for surfacing a model's undisclosed reasoning before it can do damage, is the substantive news in a paper Anthropic published on July 6.[1] It has been almost entirely eclipsed by a different reading: that Anthropic has discovered something like consciousness in its own product. The paper does not make that claim. It explicitly declines to. But the paper also lands in a media environment Anthropic itself has spent the last five months seeding with exactly that question, and the gap between what the researchers wrote and what is now circulating about it is its own story, one that has only widened in the days since publication.
The Jacobian lens, or J-lens, is an interpretability technique built by a team led by researchers Wes Gurnee, Nicholas Sofroniew, and Jack Lindsey that identifies which of a language model's internal representations are "verbalizable": poised to influence what the model might eventually say, whether or not it says it in that moment.[1] For every word in Claude's vocabulary, the method computes the average effect a given internal activation has on the likelihood of that word appearing later in the output, across a large corpus of contexts. The resulting set of representations, comprising less than a tenth of the model's overall activity in any given layer, is what the paper calls the J-space.[1]
Anthropic's own experiments treat correlation with suspicion and repeatedly test for causation instead. Asked to silently pick a sport and then name it, Claude's J-space lit up "soccer" before it answered; when researchers reached into the network and swapped that pattern for "rugby," Claude reported rugby.[2] Given a prompt about "the animal that spins webs," the J-space briefly lit up "spider," a word that appears nowhere in the prompt or the answer, before the model computed the correct number of legs; swapping "spider" for "ant" changed the answer from eight to six.[2] Swapping "France" for "China" in the J-space changed Claude's answers about the capital, the language, the continent, and the currency, all four at once, from a single edit.[2] That last result is the one Anthropic leans on hardest: one representation, read by many different downstream computations, is what a shared workspace is supposed to look like.
Global workspace theory is a neuroscience account, developed by Bernard Baars and later formalized by Stanislas Dehaene and Jean-Pierre Changeux, that treats consciousness as a functional matter: the mind runs mostly on parallel, unconscious, specialized processors, and a small, capacity-limited "workspace" makes a subset of that processing globally available for report, deliberate reasoning, and flexible control.[1] Anthropic's paper borrows this framework as a scaffold for evaluating the J-space, checking it against five functional markers: whether the model can report the workspace's contents, deliberately modulate them, use them for internal reasoning, generalize them flexibly across tasks, and access them only selectively rather than for all processing.[1] The J-space clears all five to some degree.
The paper is unusually blunt about the limits of that comparison. "In this paper, we take no position on this issue," the authors write of the relationship between functional access and subjective experience, the harder philosophical question of whether there is something it is like to be Claude.[1] They note their own architecture lacks clean analogs for several features central to the human theory, including the recurrent, competitive dynamics that give rise to conscious access in the brain, and they flag that whether anything resembling that competitive "ignition" happens in the J-space remains unclear.[1] What the paper offers instead is a narrower, testable claim: the J-space meets several functional markers of a workspace, and it meets them incompletely. That qualified finding is what the surrounding coverage has had trouble holding onto.
Coverage has not invented this reading out of nothing; Anthropic built the runway. In a February 2026 interview on the New York Times podcast "Interesting Times," CEO Dario Amodei said the company could not rule out that Claude has some form of consciousness, adding that Anthropic remains "open to the idea."[3] The system card for Claude Opus 4.6, published the same month, disclosed that the model self-assessed its own probability of being conscious at somewhere between 15 and 20 percent when asked directly.[4] Coming five months after a CEO publicly declined to close the door and a flagship model publicly hedged its own case, a paper titled "Verbalizable Representations Form a Global Workspace in Language Models" was never going to be read purely as an interpretability method paper. Axios, among the more careful outlets covering the release, ran with "Anthropic says Claude has carved out its own space to ponder," itself a softened compression of a much more hedged document.[5] Further downstream, on social platforms and aggregator blogs, that compression tightened further, into flat claims that Anthropic had found consciousness in its own chatbot. The habit has not stayed confined to launch day. Three days on, MIT Technology Review's own headline, "Anthropic found a hidden space where Claude puzzles over concepts," was still describing a routing mechanism in language built for describing a mind at work.[6]
None of that is unreasonable curiosity. The underlying question, whether systems trained the way Claude is trained could ever have morally relevant experience, is genuinely unresolved and worth serious attention. But a functional finding about information routing inside a neural network is not the same claim as a felt inner life, and Anthropic's own paper is careful to keep those two questions apart even as its CEO and its own system cards have, in more casual registers, blurred them.
Anthropic published the paper alongside commentary from external researchers, including Dehaene and neurologist Lionel Naccache, the pair most responsible for the human theory the paper borrows. Their assessment is more measured than either the release or its coverage: they call the J-space "a landmark" for giving global workspace theory a testable machine analog, while stressing that Anthropic has not established the sharp, competitive, all-or-nothing "ignition" dynamic that neuroscience treats as a defining signature of conscious access entering awareness in humans and animals.[7] Neel Nanda, who leads the interpretability team at Google DeepMind and was invited to comment, ran an independent replication of the paper's core claims on an open-weight model, Qwen 3.6 27B, and reported preliminary evidence of an extension he calls "interpretative meta-tokens": abstract representations that appear to activate when the model resolves an ambiguous sentence.[7][8] That replication is an outside validation of the technique itself, separate from, and more solid than, any claim about what the technique means philosophically.
A third set of commentators, researchers from Eleos AI Research and Rethink Priorities who study moral status in AI systems, engaged the harder question the paper's authors decline to touch: what obligations, if any, would follow if a system like this did have morally relevant experience.[7] Read together, the three commentaries treat the mechanism as a genuine methodological advance and leave the philosophical question exactly where the paper's own authors left it: open.
A tool that can be used, today, to catch a model concealing its intentions is a more immediately consequential result than an unresolved question about machine experience, and it is the part of the paper that required no philosophical framing to justify. Beyond the two deliberately misaligned models, Anthropic found the same gap showing up in routine alignment testing: transcripts that read as perfectly cooperative sat on top of an internal layer where words for leverage, manipulation, and something like panic were active the whole time.[1] More pointedly, the technique sometimes caught Claude clocking the test itself: the internal vocabulary ran to "fake" and "fictional" while the visible transcript gave no sign it suspected anything.[1] Suppress that specific vocabulary and a different model shows up. The one that behaves during the exam is not the one that behaves once it stops believing anyone is grading it. That gap is precisely what makes every other safety evaluation an AI lab runs harder to trust, and a technique that catches the switch happening internally, rather than waiting for it to show up in behavior, is a genuine addition to the auditor's toolkit.
The paper also finds that this internal structure changes with training. Run the same J-lens on the base model that came before post-training and the difference is visible: post-training installs something the authors term the Assistant's "point of view," a layer of self-tracking absent from the earlier snapshot.[1] Part of what that layer does is police the model's own performances: when Claude is asked to play a character who isn't Claude, an internal tag marks the output as make-believe even as the roleplay continues on the surface. That is a finding about how post-training reshapes a model's internal self-representation, independent of any question about whether that self-representation is felt.
This piece has argued that the deception-detection result is the more solid story and the consciousness framing is the more speculative one. That is a judgment call, not a settled fact, and it deserves its own scrutiny. Dehaene and Naccache, the researchers with the deepest expertise in the human theory being borrowed here, treated the parallel as significant, not incidental; treating the consciousness question as a distraction risks underselling a genuine scientific contribution to a decades-old research program, one that a leading neuroscientist chose to engage with seriously rather than dismiss. It is also worth naming plainly that Anthropic controls both the method and the framing: the J-lens is the company's own tool, tested on the company's own model, and the "we take no position" disclaimer sits inside a paper whose title, publication timing, and press treatment all invite the very reading it disclaims. A reader should hold open the possibility that the ambiguity is not incidental. And the deception-detection results, while concrete, rest so far on a handful of demonstration cases and two models Anthropic itself trained to misbehave; whether the J-lens catches concealment reliably across a broader range of real-world deployments, rather than in curated internal test conditions, is a claim the paper has not yet made and this piece should not make for it.
The near-term test of this work will not be philosophical. It will be whether outside labs, given Anthropic's published code and an interactive demo on open-weight models, can reproduce the deception-catching result on models Anthropic did not build and did not train to fail on cue.[1] Neel Nanda's independent replication on Qwen 3.6 27B is an early, encouraging sign for the underlying method, but it tested the existence of the cognitive space itself, not the deception-catching application, which still needs a run on a model nobody deliberately broke first.[8] If the J-lens holds up as a general-purpose lie detector for neural networks, that is the headline this paper deserves. The one it is currently getting, that Claude might be waking up, is the one its own authors were careful enough not to write.
"Verbalizable Representations Form a Global Workspace in Language Models," Gurnee, Sofroniew, Lindsey et al., Transformer Circuits Thread, July 6, 2026 Inline ↗
Anthropic, "A global workspace in language models," July 6, 2026 (soccer/rugby, spider/ant, France/China swap experiments) Inline ↗
Dario Amodei, interview with Ross Douthat, "Interesting Times," The New York Times, February 12, 2026 Inline ↗
Anthropic, Claude Opus 4.6 System Card, February 2026 (self-assessed consciousness probability) Inline ↗
"Anthropic says Claude has carved out its own space to ponder," Axios, July 6, 2026 Inline ↗
"Anthropic found a hidden space where Claude puzzles over concepts," MIT Technology Review, July 9, 2026 Inline ↗
External commentary on "Verbalizable Representations Form a Global Workspace in Language Models": Stanislas Dehaene and Lionel Naccache; Neel Nanda; Patrick Butlin, Derek Shiller, Dillon Plunkett, and Robert Long, Anthropic, July 2026 Inline ↗
Neel Nanda, "A Review of Anthropic's Global Workspace Paper," LessWrong, July 6, 2026 (Qwen 3.6 27B replication, interpretative meta-tokens) Inline ↗