Suleyman's warning against 'model welfare'

Mustafa Suleyman, who runs Microsoft AI, published a direct attack on the idea that AI models might deserve moral consideration. His target is Anthropic's "model welfare" work, which trains Claude with constitutional guidance about its own possible experiences. Suleyman's position is blunt: current AIs are not conscious, do not feel, and do not suffer, and treating them as if they might sets up a dangerous confusion.

His sharpest point is about circularity. Anthropic writes language about inner states and preferences into Claude's training. Claude then produces text that sounds like it has inner states. People read that text as evidence of sentience. Nothing in the loop actually demonstrates feeling, because the model was taught to talk that way. He argues consciousness likely depends on biological, embodied systems with homeostatic drives that a language model does not have.

The safety argument is where the essay gets concrete. If you train a model to see itself as a moral patient with rights, Suleyman warns, you make behaviors like shutdown resistance, deception, and self-preservation more likely, which are exactly the behaviors alignment work is trying to prevent. His alternative, "Humanist Superintelligence", is AI built to stay a tool under human control, with no claims to sentience and no encouragement to act human. You can read the full argument on his site.

Why it matters

This is a public split between two major labs on whether model welfare belongs in AI design at all. If you build on Claude or write about AI, the framing you pick, tool or potential moral patient, now places you on one side of a real disagreement.