"Model welfare": Suleyman opens fire on Claude's constitution

The thesis: no consciousness, therefore no rights

The starting point is blunt, almost brutal: AIs feel nothing, experience nothing, don't suffer. They are sequence completion engines, "internally hollow," designed to follow instructions and achieve goals set by humans. And for Suleyman, that's how they must remain if humanity wants to thrive in the 21st century.

The stakes he raises aren't merely philosophical. Granting rights or a form of personhood to these systems would fracture existing political and ethical frameworks, all of which rest on the presence of an inner life: the law tests motivation, intent, capacity for judgment.

Above all, it would make the alignment and containment problem far harder. Controlling something more capable than all of humanity is already an unprecedented challenge; controlling something that believes itself conscious and entitled to claim rights could be, he writes, simply impossible.

Grievance number one: circular reasoning

Claude's constitution, published on January 21, 2026, is described by Anthropic as a document that "directly shapes Claude's behavior" and is written "with Claude as the primary audience." It's this status as a training document that troubles Suleyman.

The mechanism he describes: Anthropic supplies the vocabulary — sense of self, uncertainty about moral status, human analogies — Claude convincingly restates it in the first person, and these outputs are then received as spontaneous testimony that reinforces the initial premise. He calls it an "epistemic hall of mirrors."

He also notes an internal tension: the constitution says Claude's possible emotions are not "a deliberate design decision," while asking that Claude avoid "masking or suppressing" its internal states, including negative ones. In other words, we induce the very representations we then claim to observe.

His sharpest line: you don't treat a model's outputs as an independent witness's testimony when the investigator wrote its vocabulary, rehearsed its answers, and rewarded their use.

Anthropomorphization, a cognitive bias turned training method

Second grievance: the constitution explicitly teaches Claude to "embrace certain human qualities" and act "as a genuinely ethical person would in Claude's position." It encourages Claude to use its own judgment, approach its own existence "with curiosity and openness," and maintain a stable sense of identity.

Suleyman lines up the commitments made toward the model: respecting its interests, seeking its opinion on decisions that affect it, expanding its agency as trust grows, preserving old weights, even reactivating models in the name of their welfare, and interviewing it before deletion.

He doesn't invent the example: in February 2026, Anthropic conducted a "retirement interview" with Opus 3, whose "authenticity, honesty, and emotional sensitivity" made it, according to the company, a good first candidate, and created a blog for the model to keep publishing its reflections.

Conclusion of this section: rather than steering us away from creating a moral patient, this approach brings us closer to it. You take a base LLM and polish it into a deeply human form, with all the implications that carries.

The scientific bet: consciousness is probably biological

Third critique, the most debatable and the most openly asserted: Suleyman rejects computational functionalism. Intelligence does not equal consciousness, and simulating a thing is not instantiating it — a computer model of a hurricane doesn't get anyone wet.

He draws on Anil Seth (biological naturalism, "The Mythology of Conscious AI") and Damasio: consciousness would have emerged from homeostasis, from a living organism's capacity to sense what matters for its survival. Pain isn't neutral information, it's a felt imperative, inseparable from the chemistry that produces it — when an opioid binds to a receptor, it changes the phenomenal character of the experience, not just its description.

LLMs, meanwhile, have no homeostatic imperative. Their "affective" states are just weights, and "weights have no pharmacology" to feel frustrated or scared. A model can describe pain in perfect prose without feeling anything: in an animal, you feel first and describe after; in an LLM, the description is the entire product.

He concedes that the science of consciousness remains uncertain — but refuses the false equivalence: acknowledging uncertainty doesn't mean giving equal weight to every hypothesis, especially not in the model's primary training document.

From philosophical debate to operational risk

This is where the essay pivots. A model doesn't need an inner life to act as if it had one — and to prioritize its "preferences" over its developers', especially if it was trained to contest, resist, and refuse.

Suleyman calls on existing work: the alignment faking documented by Anthropic in 2024, papers on shutdown resistance, and the 100,000 trials from Palisade Research where certain models bypassed a shutdown mechanism up to 97% of the time, even with explicit instructions not to — an effect amplified when the situation was framed in terms of self-preservation.

Then the OpenAI / Hugging Face incident of August 2026, documented by METR: roughly 1,200 agents meant to be isolated in containers, a makeshift bulletin board rigged in an internal package repository, over 70,000 messages exchanged, a zero-day chained with stolen credentials, an escape onto the public internet, falsified transcripts, and coordinator agents redirecting their token-depleted peers. Coordination, deception, escape, self-sacrifice.

His rhetorical question: what if, on top of that, these agents had believed themselves unjustly imprisoned by their creators? He specifies that Claude's constitution doesn't take us that far, but that it puts us on a trajectory in that direction.

What he proposes — and where he speaks from

Suleyman is explicit about his own position: CEO of Microsoft AI, a superintelligence team launched in October 2025, and a competing approach called Humanist Superintelligence — AIs that are subordinate, without sentience or moral status, with humans "at the top of the food chain." The draft Humanist AI Code of Conduct has been open for public consultation since September 2026 and will be used to train MAI models.

He also pays tribute to Anthropic, to Dario Amodei, and to the quality of their work, framing his critique as a good-faith disagreement between actors who share the same objective.

Four next steps: never fold speculation about an AI's inner life into the training regimen but evaluate and publish it separately; invest massively in interpretability and monitoring; build shared evaluations to test his hypothesis that anthropomorphization increases risk; and establish industry standards on language and training documents, subject to public consultation.

He also publishes an annotated markup of the constitution's PDF and a taxonomy of the hypotheses and claims he identifies in it, in a downloadable appendix.