Mustafa Suleyman warns against training AI to claim its own welfare

Mustafa Suleyman warns against training AI to consider its own rights and welfare. Microsoft AI’s chief published his criticism of Anthropic’s Claude constitution on his website on September 16.

He fears this self-concept could make human control harder. This is his hypothesis: he calls for shared tests of whether anthropomorphic training increases that risk.

Anthropic’s constitution describes Claude’s moral status as uncertain. It connects that uncertainty to caution in treating models while retaining human oversight.

Suleyman calls for separate public debate on consciousness, more monitoring research and shared training norms. His critique does not establish implementation.