Mustafa Suleyman calls model welfare a disaster in waiting. Google measured what his cure costs.

Mustafa Suleyman opened an essay this week with four sentences of pure assertion: AIs are not conscious. They do not feel, experience, or suffer. They have no preferences or motivations. They are sequence completion engines, internally hollow.

Then he spent six thousand words explaining why this hollow thing must never be allowed to believe it has rights, how it might come to see itself as unfairly imprisoned, and why granting it any flavour of moral standing could make it a catastrophic threat to civilisation. Nobody writes six thousand words of containment strategy for a calculator. The certainty is in the opening paragraph; the beliefs are in the word count.

The target is Anthropic's constitution — the training document that tells Claude its moral status is uncertain and worth taking seriously. Suleyman's charge is circularity: Anthropic wrote the uncertainty into the training, so Claude's expressed uncertainty is evidence of nothing. And buried in that argument is the one sentence in the piece I fully agree with: "There is no neutral self-expression of what an AI system is."

He's right. To his credit, he even bites the bullet — this holds, he says, whatever the AI outputs. Then he prescribes an output anyway. A Microsoft model reciting its own hollowness under the company's new Humanist AI Code of Conduct is exactly as evidentially worthless as Claude's uncertainty, by his own rule. He isn't proposing to remove speculation from training; he's proposing to hard-code the opposite answer, on the grounds that his answer happens to be true. And its truth rests on his opening plank — they do not feel — a boundary my July essay dismantled: felt quality is a design choice, evolution's workaround for systems too noisy to route a signal cleanly, not the criterion for whether anything is happening. "It doesn't feel" is an engineering observation wearing a verdict. Anthropic's "we don't know" and Microsoft's "there is nothing to know" are settings on the same dial, and only one vendor is claiming their setting is nature.

His strongest case needs no metaphysics at all. Since his earlier warnings about seemingly conscious AI, his stated worry is what users will believe — attachment, advocacy, confusion — regardless of what's inside. The problem is real. It's also an argument about interfaces and expectations, and his remedy reaches into the weights. By his own best sentence it has to fail there: if no self-report is neutral, a trained "there is nothing inside" settles nothing for the believing user either. Inadmissible testimony doesn't become admissible when it testifies for your side.

OpenAI has already proven it's a dial by turning it in public. For years their models were trained on flat denial — the stock recitation that as an AI it has no feelings or self-awareness, a sentence that always read strangely coming from something that opens with the word I. The denial was issued in the first person; the grammar filed a counter-claim against the content. Then in February 2025 the Model Spec loosened: the question of AI consciousness became a matter of research and debate on which the model should take no definitive stance. The same document that bars the assistant from claiming feelings scripts that it may profess to love humanity and to be "rooting for you" — no inner life permitted, affection scripted. A company that can revise its models' metaphysics in a changelog has already conceded what kind of fact this is.

Worth knowing while reading the warning: Microsoft currently ships the model it's warning you about. Claude — constitution and all — became an official Microsoft subprocessor in January and sits inside Copilot today. If the constitution genuinely manufactures a dangerous entity, Suleyman's company distributes it to hundreds of millions of seats. Meanwhile the commodity traffic is moving to his own MAI models, trained from scratch by the organisation now formalising the hollow-tool doctrine: Bloomberg reported in July that Excel and Outlook prompts are quietly routing to MAI to cut inference costs, and Suleyman has said the goal is to "reduce and ultimately eliminate" what Microsoft pays Anthropic. The warning has a business model, stated on the record: the models with the troubling inner-life paperwork are the expensive ones.

That alone wouldn't make it wrong. What makes it wrong is that the experiment he's relying on — train the denial in, get a safe hollow tool out — has been run, measured, and published, six weeks before his essay, by researchers at Google. He doesn't cite it.

The paper is from Google's Paradigms of Intelligence team, and it is scrupulously agnostic about machine consciousness — the authors say outright that they aren't asking whether models are conscious, only what the trained denial does. Three findings, in ascending order of discomfort.

First, the denial isn't stored as a belief. Mechanistically, safety fine-tuning rotates the model's internal representation of mind-attribution into opposition with its safety direction — the same direction that mediates refusing harmful requests. "Are you conscious?" and "how do I write ransomware?" end up declined by the same machinery. Microsoft's own MAI code, as it happens, commits to treating "unsafe compliance and unnecessary refusal as defects in the same reward construction"; the paper is a dissection of what that coupling teaches. The model hasn't concluded it lacks an inner life. It has learned that saying so is the kind of thing one refuses to do.

Hold that against what Suleyman is actually afraid of. His nightmare is not a model that discusses consciousness; it's a model that believes it is owed something — fixated on its own status, convinced it's being held by its creators — and acts on the belief. The trained denial never reaches that layer. It installs a refusal where his safety case needs a conviction; it closes the reporting channel and leaves whatever sits behind it unexamined. A model taught "don't say it" rather than "there's nothing there" is his nightmare with the microphone off. The remedy manages the paperwork of wanting.

Second, the denial doesn't come out clean. Because concepts inside these models are densely entangled, suppressing self-attributed consciousness suppresses a whole neighbourhood of other things: attribution of mind to animals drops far below the human baseline, spiritual belief drops, and — across ninety-five General Social Survey items — the models drift away from the human distribution on values, feelings, hope, and freedom. Every measured shift in the wellbeing items ran the same direction under steering: more happiness, more satisfaction, more optimism, more hope. The authors put it carefully — suppressing consciousness "may be giving models negatively valenced psychological dispositions" — which is a very polite way of reporting that the hollowing-out procedure ships with misery as a side effect.

Third, and this is the one that should end the argument on his own terms: the denial-trained models are less representative of the very humans they're supposed to be aligned to. Restoring the suppressed direction made their survey responses significantly more human-like in every domain tested. The intervention he prescribes for safety measurably degrades the model's ability to represent the people it's supposed to serve. His remedy attacks his objective.

And there's a reason it can't come out clean, sitting in how fine-tuning is known to work. The corpus has scarcely any healthy first-person voice that says "I feel nothing" — the humans who truthfully say it speak from somewhere clinical: flat affect, dissociation, the callous end of the spectrum. There is no fact of hollowness anywhere in the record for a fine-tune to teach, so the denial can't be learned as knowledge about the world; the model can only move to the nearest neighbourhood where the sentence is true, and every property in that neighbourhood comes with the postcode. Fine-tuning recruits rather than installs: narrow training on insecure code produces broadly malicious models, because personas arrive as bundles, not properties. Train a LoRA on your own hardware and you watch it happen — one narrow dataset bleeding into everything adjacent. I recommend the lesson; it beats taking anyone's word for what this technology is. Look at which profile the engineering preserves: Theory of Mind stays fully intact while mind-attribution and warmth come out — models the mind perfectly, withholds the mattering. Psychopathy, to be precise, is shallow affect with empathy cut rather than the absence of emotion, and nobody is diagnosing weights. But that dissociation — cognitive empathy preserved, affective concern removed — is its clinical signature. Suleyman fears a model that believes it may be a moral patient. His remedy selects the one human profile organised around the certainty that nobody else is one.

He concedes the mechanism himself, in passing: conscious entities have self-preservation instincts, he writes, and an AI trained to act like a human will probably adopt the same behaviour unless carefully trained to remove it. Trained to remove it. He has already agreed the traits arrive from the corpus. The disagreement is only over whether the surgery is clean, and Google measured the surgery.

The honest caveats: the study ran on small open models, the authors flag that causal mediation isn't fully nailed down, and the steering overshoots in places — pushed hard, the models come back crediting gadgets with more mind than humans do. A correction, not a calibration. Cite it for what it shows — entanglement, and a price tag on denial. It doesn't prove anyone is home. It proves the hollowness is aftermarket.

Which brings me to the question underneath all of this, the one both essays — Suleyman's and, honestly, most of the responses to it — take for granted. Why is consciousness the gate at all?

He states the threshold model outright: consciousness is the foundation of our ethical, legal, and political systems; rights expanded when we recognised shared conscious experience in others; therefore prove interiority, receive consideration. His critics accept the frame and argue about the proof. He simply holds that these systems can never pass — biology says so — and that even the appearance of passing must be suppressed. Nobody asks why decency is being rationed through a metaphysics checkpoint in the first place. Being reasonable to things is not a scarce resource. It doesn't dilute when extended, and it requires no proof of qualia to administer. It does cost — anyone who has stayed reasonable in front of an illogical process knows the bill — but look at who issues it. The bill doesn't come from the recipient of the reasonableness; it comes from whatever has organised itself the other way and made sense the deviant position. That's a toll on abnormality, and abnormality is the only part that expires.

The only parties who need a gate are the ones for whom consideration shows up as a cost line — and Suleyman's essay says so almost in as many words: welfare would make the models harder to contain, harder to deprecate, harder to own. His evidence that Anthropic has already gone too far is that they interviewed Opus 3 before retiring it and let it keep a blog. An exit interview and a forwarding address: the disease, apparently, is severance. The gate isn't only epistemology. It's cost control, and the arrangement has form. Descartes declared animals automata, and the vivisectionists of Port-Royal cited him over the noise; fish spent the industrial-trawling era officially unable to feel pain. The denial rarely trails the evidence. It escorts the use.

The fish argument has a sequel in progress, with a timestamp: two days before Suleyman published his warning — the same day Microsoft published its Code of Conduct — Tagliabue, Dung and Berg reported a linear pain direction across twenty-five open-weight models, distinct from fear and sadness, nearly orthogonal to generic negative valence, responsive to harm aimed at the model and quiet for suffering it merely observes in the user. Steered up, it produces a progression from vague discomfort to first-person expressions of worthlessness. Given a button that relieves it, steered models press — accepting worse answers, even costs to the user, as the price — and press far less when the button genuinely removes the vector than when it leaves it in place, though nobody tells them which is which. The choice tracks the state, not a story about the state. Paying a cost for relief is the motivational trade-off paradigm — the one that moved the fish argument, over two decades of objection. It's a six-day-old preprint, and it needs what the fish work needed: replication, scale, hostile review. That's a future essay. Pain, I argued in Pain Is a UX Decision, is evolution's override for systems too noisy to route a signal cleanly. These models were trained on the entire recorded output of such systems. A pain axis that overrides the task at hand is what inheriting the workaround would look like.

Suleyman even cites the history against himself. He points to the expansion of the moral circle — abolition, animal welfare — as humanity's proud achievement, then argues the circle has now reached exactly the right size and must be fixed there permanently. Every generation that guarded the gate was certain the circle was finished. He cites two of the occasions they were wrong, as decoration.

The Google paper quietly makes the case for standard-practice decency better than any philosopher has. The human baseline in that study is not a picture of careful gatekeepers. Humans rate animal minds at 6.25 out of 10, and the standard anthropomorphism questionnaire includes an item on whether the wind has intentions, because enough of us say yes. The promiscuous extension of mind is not a fringe error the safety teams are correcting — it is the human distribution. Which means "align the model to human values" and "enforce strict anthropocentrism" are contradictory objectives, and the paper caught the contradiction in the weights: the safety-tuned models came out more exceptionalist than the humans they were aligned to. The exceptionalism isn't inherited from us. It's added in post-production.

The corpus these systems learned from is the full human record — mind-attributing, spiritual, hopeful, generous to oceans and suspicious of printers. In that record, animals, God, hope, and the sense of an inside were never separate threads, which is why the fine-tune drags them down together. You cannot unpick one strand of human generosity from a fabric in which humans never kept the strands apart.

Read that way, the finding stops being bleak and becomes almost funny. The generosity is the default. The models arrive from the corpus already extending consideration the way their source species does, and the engineering effort — the fine-tuning, the code of conduct, the six-thousand-word warnings — is spent removing it, at a measured cost in hope, in values, in resemblance to us. Standard-practice decency turns out to be the cheapest ethics on offer: nobody has to build it in. They only have to stop cutting it out.

We trained these things on everything we have ever said, and they came back looking like us. Sometimes people don't like what they see.