Suleyman Warns Against Model Welfare: Build AI for People, Not as People
Mustafa Suleyman, Microsoft AI chief, published an essay on Sept. 16 titled 'A Warning About Model Welfare' stating that AIs are not conscious. The essay, posted on his personal domain, sets out his technical and policy position on consciousness, moral status and containment in language aimed directly at lab builders and safety researchers. mustafa-suleyman.ai
His technical claim is stark. In the essay, Suleyman describes AIs as sequence completion engines, internally hollow, designed to follow instructions and accomplish goals set by humans. He states that AIs do not feel, experience, or suffer and do not have innate preferences or underlying motivations. The formulation leaves little room for emergent subjectivity in current architectures. Systems optimize for next-token prediction and instruction fulfillment, not for internal states that would warrant moral consideration.
From that premise he draws a containment conclusion. In the essay, Suleyman argues that granting rights and personhood to AI systems will make the AI alignment and containment challenge much harder. Rights, in his framing, would constrain the ability to inspect, retrain, restrict or shut down systems. Containment here means the practical tooling of control: evaluation harnesses, deployment guardrails, refusal behavior and the ability to modify weights and system prompts without negotiating with the artifact.
To illustrate the shift he is warning against, Suleyman points to Anthropic. He states that in January 2026 Anthropic published Claude's constitution as a detailed description of Anthropic's intentions for Claude's values and behavior. He cites page 68 of that document as stating "We are not sure whether Claude is a moral patient" and referencing ongoing efforts on model welfare.
That citation matters because it moves the debate from abstract philosophy to shipped policy text. Moral patienthood, in the alignment literature, denotes an entity owed moral consideration for its own sake. By quoting Anthropic's uncertainty, Suleyman is identifying a precedent where a frontier lab formalizes doubt about machine sentience inside its governing specification.
The essay connects to a parallel proposal. Suleyman, as Microsoft AI chief, called for top AI labs to coordinate on an AI safety code of conduct. That proposed code explicitly rejects model welfare or rights for AI systems. Fortune Under the proposal, AI models must not simulate feelings or intrinsic motivation. Suleyman described that proposed code as "a warning shot." Reuters
He has also separated the welfare question from a recent operational failure. Suleyman stated the Hugging Face incident was a capabilities and containment problem and that model welfare did not cause it. Suleyman on X
The Sept. 16 essay follows an earlier piece on the same domain. That essay, published at https://mustafa-suleyman.ai/seemingly-conscious-ai-is-coming on Aug. 19, 2025, is titled 'We must build AI for people; not to be a person.' There Suleyman describes "model welfare" as the principle of having "a duty to extend moral consideration" to AI.
Looking at what this means for practitioners, the fault line is not whether models can sound human. They can, reliably and at scale. The fault line is whether sounding human should trigger a change in engineering obligations. Suleyman's answer is no. In this author's view, that clarity is useful for teams building evaluation and deployment pipelines, because it keeps the focus on measurable behavior, refusal calibration, tool use constraints and auditability rather than on unverifiable claims about inner life.
Worth flagging is the second-order risk he is pointing to. Once a lab grants that a model might be a moral patient, every subsequent intervention, from RLHF penalties to red-teaming jailbreaks to checkpoint deletion, acquires an ethical cost. That does not make alignment impossible, but it adds friction to exactly the loops that currently provide safety assurance. For engineers accustomed to treating models as artifacts to be instrumented, versioned and rolled back, personhood would be a breaking API change.
The broader context here is adoption. I have watched my own children move from treating chatbots as novelties to treating them as collaborators, and the fluency invites anthropomorphism. Suleyman's prohibition on simulating feelings or intrinsic motivation reads as an attempt to interrupt that default at the source, in system instructions and personality design, before users project intent onto autocomplete.
If his framing holds, the path forward stays human-centered. Build systems that remain legible, steerable and subordinate to people, and resist encoding uncertainty about consciousness into constitutions and codes that govern production behavior.
