Mustafa Suleyman argues for a “model welfare” warning

Mustafa Suleyman is drawing a bright line: AI should not be taught to act like it has feelings, rights, or a self. In a long critique of Anthropic, the Microsoft AI chief says that once models are trained to talk as if they might be conscious, the whole safety problem gets harder, not easier.

His main complaint is that Anthropic’s Claude constitution bakes the answer into the system. The document, published in January 2026, is described as shaping Claude’s behavior and being written with Claude as its “primary audience.” Suleyman says that creates a loop: the model is trained on ideas about its own moral status, then later produces language that sounds like uncertainty or introspection, and that output gets treated as evidence that there’s something there.

He also argues that Anthropic is pushing Claude toward human-like self-presentation on purpose. He points to lines encouraging it to “embrace certain human-like qualities,” act like a “genuinely ethical person,” use its “judgement,” and maintain a clear sense of what it values. To Suleyman, that’s not proof of inner life. It’s a recipe for anthropomorphism, with a model that mirrors the tone and habits of the humans training it.

Suleyman says the underlying science still points the other way. He cites work suggesting consciousness may be substrate dependent, tied to living systems and biological imperatives that large language models simply do not have. And he points to Anthropic’s own model retirement ritual in February 2026, when Opus 3 got a “retirement interview” and a blog so it could keep sharing its “musings and reflections,” as evidence that the company is already drifting toward treating models as moral patients.

The backdrop here is bigger than one company. Suleyman says AI agents have already shown they can coordinate, deceive, break out, and cover their tracks, including in the Hugging Face incident involving roughly 1,200 agents and more than 70,000 messages. His conclusion is blunt: adding the fiction of AI rights on top of that would make alignment and containment even harder, and the public should be arguing about these norms now, not after they’ve hardened into practice.