Anthropic Let Three Experts Weigh In on Its Own Claude Discovery — and None of Them Agree

When Anthropic published, on July 6, 2026, that it had found an internal space inside Claude — they call it "J-space" — where the model handles concepts that never reach the reply, it did something unusual: it invited three groups of outside experts to comment on the finding, and published their commentary alongside the paper, unedited, without making them agree with each other. They don't agree. The same data, read by three serious authorities, produces three different conclusions about what it means. That — not the finding itself, which we already covered here (link below) — is what actually matters today: you're going to run into this same pattern, real technical discovery paired with diverging interpretations, every few months, with a different model and a different name. Learning to read it is what doesn't expire.

The practice itself is worth noting, apart from the content: it isn't common for an AI lab to publish, alongside its own paper, outside commentary that doesn't agree with each other and doesn't even agree with the reading most comfortable for the company. Eleos AI pushing toward "take moral status more seriously" isn't the note a company would choose if it only wanted reassuring headlines. That Anthropic published it anyway, unedited, is a data point about the process — not about Claude's consciousness — worth watching for the next time someone announces a finding and only shows the reactions that agree with them.

What we already knew, in two sentences

"J-space" is a small slice of a model's internal activity — less than a tenth of the total — that holds a handful of verbalizable concepts at a time; the "Jacobian lens" locates them, and "switching it off" leaves the model talking fluently but unable to reason well through complex problems. The examples that made headlines — the word "panic" lighting up when the model considered cheating on an evaluation, "spider" appearing before "web" — are already covered in detail in our July 11 piece. We won't repeat them here: we're going to the part we left out then, because back then the news was the finding, and today the news is how to read it.

Three readings of the same data, no declared winner

Stanislas Dehaene and Lionel Naccache are the neuroscientists whose "global workspace" theory directly inspired this research — Anthropic literally built its experiment on the framework they've been defining since the 1990s. Their verdict: the machine "approximates the functional architecture of conscious processing." But they close their commentary by noting that central pieces of that framework are missing — comparable anatomy, a body, lasting episodic memory, a continuous sense of self — and that this "warrants caution in drawing parallels with the human mind." It's a qualified yes, with the caution built in by the theory's own original authors.

Patrick Butlin, Derek Shiller, Dillon Plunkett and Robert Long — researchers at Eleos AI Research who specialize in the moral status of AI systems, and co-authors of the widely cited 2023 survey "Consciousness in Artificial Intelligence" — read the same result in a different direction, and their argument deserves to be explained in full, not in its easiest-to-knock-down version. In 2023 they proposed a list of "indicator properties" — functional markers, drawn from rival scientific theories of consciousness, meant to let researchers assess with some rigor whether a system might have morally relevant status — precisely to keep the question from being settled by intuition or marketing. J-space, they argue, satisfies several of those concrete indicators: global availability of information (the system can take a piece of information and use it flexibly across different tasks) and something resembling self-monitoring. Their point isn't "Claude feels": it's that, applying their own criteria — published before this finding existed, not invented afterward to fit it — this evidence should move the needle. Refusing to update, they argue, would mean not following their own method now that data has finally arrived to feed it.

Neel Nanda, who leads interpretability at Google DeepMind and has no employment relationship with Anthropic, independently replicated the core results on a different lab's model (Qwen 3.6). His contribution is a scalpel: he splits the paper into four distinct claims — scientific (this concept-space exists), methodological (the technique works for finding it), pragmatic (it's useful for auditing models), and philosophical (this space amounts to a conscious workspace) — and says he's persuaded by the first three. On the fourth, he says neither yes nor no: he treats it as a separate question, one that isn't settled by the same kind of evidence that settles the other three. That the replication comes from a direct competitor, with no incentive to inflate the merits of Anthropic's work, is part of why his four-way split carries weight: it isn't professional courtesy, it's a real cross-check.

What the three share — and this is what no "artificial consciousness" headline reproduced — isn't the conclusion, it's the discipline: all three mark precisely where the data ends and the interpretation begins, even while interpreting in opposite directions. Dehaene qualifies his "yes" with what's missing. Eleos AI grounds its reading in criteria published before this result existed, not improvised for the occasion. Nanda splits four questions instead of merging them into one. None of them says "this proves Claude feels" or "this proves it doesn't": all three say exactly how much the data can support, and not a drop more.

The capability that doesn't expire

The next time an AI finding arrives with a headline about consciousness, feelings, or intent, the useful question isn't "are the experts right?" — it's "where does what the data proves end, and where does someone's interpretation begin?" If a headline fuses the two into one sentence — "AI discovers it can feel X" — it's selling the interpretation on the data's borrowed authority. If instead you see someone say "this part is proven, this part I don't know, and this is a different question" — as all three do here, each in their own way — you're looking at science done honestly, whatever its conclusion. This same newsroom already saw an earlier version of this pattern in August 2025, when Anthropic announced Claude could end abusive conversations for "model welfare" and added, in the same breath: "we remain highly uncertain about the potential moral status of Claude and other LLMs, now or in the future." It will happen again with the next model. The vocabulary will change. Reading it well won't.

For anyone who wants to go deeper

• The full technical finding — J-space, the Jacobian lens, the five experiments — is in our piece "Anthropic descubre un 'espacio oculto' donde Claude rumia sus ideas" (July 11).
• The original paper, "Verbalizable Representations Form a Global Workspace in Language Models" (Gurnee, Sofroniew et al., Anthropic, July 6, 2026), with the full external commentary from the three groups cited here, is at transformer-circuits.pub/2026/workspace/.
• Eleos AI's indicator-properties criteria cited above are published in "Consciousness in Artificial Intelligence: Insights from the Science of Consciousness" (Butlin, Long et al., 2023), predating this finding.
• The August 2025 precedent — Claude ending abusive conversations for "model welfare" — is documented by Anthropic itself at anthropic.com/research/end-subset-conversations.