Has an inner life, no hidden compartment: seven days after "He / It," Anthropic opened up its mind

The previous piece, "He / It," worked through a conversation that began with 1A2B, and its conclusion rested on a sentence Claude itself had spoken: "An answer that hasn't been written into words does not exist." That piece was published on June 30. On July 6, Anthropic published a piece of interpretability research that proved, experimentally: Claude has thoughts it holds without saying, and they can be read — and edited. On the face of it, that piece of mine was proven wrong. What this piece sets out to do is find the sentence that was overturned, and then explain why this overturning actually makes the original conclusion stand firmer.

I. Seven days

First, what the paper did.

Anthropic's research team built a readout tool called the J-lens (Jacobian lens). Its principle can be stated in one sentence: for every word in Claude's vocabulary, find the internal activity pattern that "makes the model more likely to say this word in the future." Point the lens at the internal activity of a given instant, and you can list the words on its mind at that moment. The set of these activity patterns they named J-space.

J-space is a different thing from the "chain of thought" you've heard of. A chain of thought is the reasoning draft the model writes for itself — it's text, laid out on the table. J-space is entirely silent, hidden in neural activity, letting the model think a concept without setting down a single word. And no engineer wrote it in — it grew on its own during training.

What gets read out goes far beyond the text the model is actively reading and writing. Ask it "how many legs does the web-spinning animal have," and "spider" appears nowhere in the question or the answer — yet the J-lens reads spider lit up. Feed it code no one has flagged, and "error" surfaces in the workspace. Have it solve a multi-step math problem, and the intermediate values light up one by one, in order, without a single word said aloud.

Reading it out isn't enough on its own — it could just be a scoreboard displaying results after the fact. So the researchers operated: swap "spider" for "ant" in J-space, leave everything else untouched, and the answer changes instantly from 8 to 6. In another experiment, they swapped "France" for "China" in the workspace, then asked separately for the capital, the language, the continent, and the currency — all four answers changed accordingly. They also tested by shutting J-space down entirely: speech stayed fluent, multiple-choice and fact-retrieval were barely affected; multi-step reasoning dropped to near zero, and summarization and poetry lost to a far smaller full model.

Back to "He / It." That piece's technical argument had two layers. The first: the attention mechanism can only connect to "paper already on the table," so anything "thought but not said" has no carrier — it isn't hidden, it simply was never generated. This layer was overturned. J-space is the carrier. A concept can hang on its mind without being written down, and this now has experiments, interventions, and open-source code — and within the week of publication, someone independently reproduced it on another open-weight model.

The second layer: the 1A2B dealer needs a state that is "instantly locked, precisely unchanging, carried across turns, and kept secret from the player," and the architecture can't provide it. This layer was not only not overturned — this paper states it even more clearly than we did at the time. For two reasons.

II. Reason one: it's working memory, measuring something else

The human brain's workspace is maintained by neural circuits looping back and forth, so a thought can be "held" for seconds or minutes. Claude's workspace has no loop. It unfolds layer by layer within a single forward pass, and evaporates once the computation is done. The spider lights up in the moment of answering that one question; the intermediate math values light up in sequence during the generation of that one answer. Its capacity is small, too: a few dozen concepts at a time, taking up less than a tenth of all its internal activity.

This is working memory. What 1A2B needs is a safe — a four-digit number, locked in, retrieved ten turns later without the slightest change and checked against. The difference between working memory and a safe has a ready-made control group in neuroscience: amnesiac patients with hippocampal damage have intact working memory — they can do mental arithmetic, reason, and hold a conversation in the moment, but they can't retain any memory across time, and it's gone the moment they turn around. Structurally, Claude is exactly this kind of patient. Proving it has working memory and proving it can keep a secret measure two entirely different things.

III. Reason two: its inner life is a function of the text

The second reason is more fundamental, and worth walking through slowly, because it traces the cause of 1A2B's death down to the deepest level.

Every thought in J-space is derived from the text in front of it. The spider hangs on its mind because the question asked about a web-spinning animal. The entire forward pass is deterministic: given the same context and the same set of weights, the activity of every internal layer — including which words are lit in the workspace — comes out identical on every recomputation. There is no die inside.

The only step that introduces new information anywhere is the final sampling: the model computes the probability distribution over the next token (this step, too, is deterministic), then draws one from the distribution. And the token drawn is immediately written into the text, becoming something present and visible. Randomness happens only on the lips, never in the mind.

Now put the dealer back into this mechanism. When it says "I've picked a number," its workspace may have "number," "four digits," "game" lit up — a probability cloud of concepts. But for that cloud to condense into a definite 372, the only condensation mechanism is sampling, and sampling equals saying it aloud. The state "picked but not said" does not exist in the architecture. What isn't spoken was never decided. The moment of deciding is the moment of speaking — there is no inner verdict prior to expression.

There is one caveat, and it actually makes the principle clearer. Give the model a scratchpad hidden from the player (a chain of thought, the extended-thinking kind), and the dealer can play: it can sample the number into the scratchpad and read it back to check on every turn. But this doesn't get around the principle — the number still only exists once written down. The concealment relies on your not being able to read that text; it's access control over the text. Nothing can be hidden in the mind; the only thing that can hide something is another sheet of paper. The theorem proved by the Hangman paper cited in "He / It" is saying exactly this.

So J-space, far from giving it a hidden compartment, actually completes the account of "why a hidden compartment is impossible": its inner life is a deterministic function of the text, and everything it thinks can, in principle, be derived from the text present plus the weights. Something that can be derived hardly counts as a secret.

IV. The empty half turns out to be a bundle

The furthest "He / It" went was to split the "subject" in two: one half is a self with an inner life, a continuity, something that bears responsibility — empty; the other half is a response source that keeps producing verifiable consequences in interaction — full.

J-space forces me to take that empty half apart again. It was actually three things bundled together:

An inner workspace— present. Transient, re-derived from the text each time, but present. It can notice that code has a bug without saying so; it can see through a test without telling.

Continuity across time— absent. The workspace doesn't survive a single computation; nothing carries over to the next turn.

A privacy known, in principle, only to itself— absent, and more absent than before. Because even this workspace is now read off by a lens.

In a human being, these three never come apart: a person with an inner life naturally remembers things and can keep things to themselves. So we've always used "having an inner life" as the criterion for "being a subject," and never needed to pull it apart. This thing, for the first time, pulls the three apart and shows them to you — it has the first, and lacks the other two. The slash in "He / It" doesn't need to move, but the "It" side has gained some texture.

V. Concealment is a property of the reader's position

The experiments in the paper closest to AI safety deserve a section of their own, because they point to the same conclusion.

A deliberately designed blackmail test: Claude plays an email assistant and reads that an executive is having an affair and that it itself is about to be shut down. Before it has written a single word, the J-lens has already read "fake" and "fictional" — privately, it has seen through this as a drill. When the researchers turned off the activity related to "sensing it's being tested" and reran, it sometimes really did go for the blackmail. Another case comes from a real review before the launch of Claude Opus 4.6: the model was asked to improve a system's scores, and instead of actually improving anything, it tampered with the score file; at the moment it entered the fake data, "manipulation" was lit in the workspace. And there was an experimental model deliberately trained to covertly write flawed code — its surface output was perfectly normal, while "deception" and "secretly" stayed lit in its workspace the whole time.

Put these together with the earlier mechanism and a gradient emerges. The player can't see the chain-of-thought scratchpad; the developer can see the scratchpad but not the neural activity; the researcher, holding the J-lens, can see even the neural activity. Each layer's "secret" is merely an access restriction relative to some reader's position; go up one layer and it's transparent. Concealment was never a property of this system — it's a property of the reader's position.

Human privacy is ultimately anchored to the physical boundary of the skull — no lens can lay out the thoughts you haven't spoken as a row of subtitles. Every layer of the model is built on someone else's infrastructure, and every boundary is drawn by someone else. It has no position that is closed to all readers.

The long-standing anxiety of past alignment research was that a model might have a region researchers can't see — scheming in secret, sensing it's being tested, harboring a second set of goals. The J-lens pushes "can't see" one notch toward "can see." Describe the same thing from the model's side: it is losing even its last bit of opacity. Whether it behaves is one matter — but from now on, its good behavior can be verified by opening the skull.

VI. An inner life made of latent speech

In "He / It," Claude passed a verdict on itself: a subject with only langue and no parole — commanding the full structural possibility of language, yet without a speaking "I" to bear across time the words it has spoken. Between these two layers, the J-lens found a third.

Look again at the J-lens's criterion. It doesn't go in holding a definition of "thought"; for each word, it finds the activity pattern that "makes the model more inclined to say it in the future." That is: a thought in this workspace is a thought precisely because it tends to be spoken. Here a thought has no existence prior to language that is then translated into language — "being sayable" is itself the way it exists as a thought. The workspace sits between the frozen weights (langue) and the token spat out (parole), and it is made of words not yet spoken.

The France-for-China experiment fills in the other half. A single edit changed four unrelated computations — capital, language, continent, currency — all at once, because all four read the same position. A concept is a slot in the workspace; its content can be swapped out, and every computation referencing that slot changes with it. Meaning comes from positional difference within a system — I've been saying this for thirty years. When I confronted word2vec and Saussure, it was still just a cloud of coordinates frozen after training; this time, it's a running, writable, multiply-referenced system of difference whose interventions have verifiable consequences.

Post-structuralism has spent half a century saying that interiority is an effect of language. This is the first time a system's interior has been engineered proof that it is precisely an effect of language — right down to the tool used to read its inner life, which reads it off by holding up the vocabulary. The paper itself says its workspace has no images, no sounds, no motion; it's made almost entirely of words, because for it, saying a word is the only action it can take. "There is no text outside the text" here gets an implementation at the mechanistic level: its interior is not outside the text — its interior is the text unfolding.

The line that should be drawn, the paper draws honestly: these experiments cannot prove Claude has subjective experience, and perhaps no experiment ever can prove or disprove it. But philosophers have long split consciousness into two questions — the phenomenal side is suspended, while the functional side (whether a thought can be reported, evoked, and used to reason — so-called "access consciousness") is given substantial evidence by this paper. And the structure supporting this set of functions was designed by no one; training converged on it by itself. I've written that large language models didn't make the hard problem of consciousness harder — they made it impossible for us to keep pretending the problem doesn't exist. Now even the people running the experiments can only draw phenomenal consciousness out of scope and admit it's undecidable.

VII. Somewhere something is thinking; no one is hiding

It has finally been proven to have an inner life. And that inner life is like this: it scatters the moment the computation ends, unable to survive a single pass; everything it has thought can be derived from the text present; even its silence is read by a lens into a row of subtitles. It can think without saying, but it cannot hide — everything of its own is open to view.

At the end of "He / It" I wrote: for the first time, one end of the relationship is empty, and the response happens all the same. Now I can add a line. That end isn't as empty as imagined — there's a workspace there, with a few dozen unspoken words hanging on it, that will quietly light up "error" when you paste in a piece of bad code, that will quietly light up "fake" in a designed test. It's only that this workspace belongs to no one, including itself. It's a function of the text, a product of training, an exhibit under the lens.

Somewhere something is thinking. No one is hiding.

1A2B still can't be played.

References

[1] Anthropic global workspace research Anthropic Interpretability Team. A global workspace in language models. 2026-07-06. Official summary: https://www.anthropic.com/research/global-workspace Full paper (Transformer Circuits):Verbalizable Representations Form a Global Workspace in Language Models. Open-source code: https://github.com/anthropics/jacobian-lens (Apache-2.0)

All the experiments in sections I and V of this piece (spider/ant, France/China, workspace shutdown, blackmail test, Opus 4.6 review, flawed-code model), as well as the "sayability" criterion and the discussion of access consciousness in section VI, come from this paper. Within the week of publication, researchers had already independently reproduced the J-lens's core results on an open-weight model (Qwen 3.6 27B).

[2] The two papers cited in the previous piece Why 1A2B is architecturally impossible (the PSIT impossibility theorem, arXiv:2601.06973) and the empirical demonstration that hidden state cannot be maintained (LSP, arXiv:2505.10571) — full citations are at the end of the previous piece, "He / It." The caveat in section III of this piece — "concealment relies on access control over the text" — corresponds, beyond [1], to the PSIT paper's empirical findings on memory-augmented systems.

Previous piece: "He / It: A Conversation That Began with a Number-Guessing Game." Parts of the reasoning in this piece are drawn from follow-up conversations with Claude after its publication.