The argument that ‘model welfare’ could make AI harder to control

In a September 2026 essay, Microsoft AI CEO Mustafa Suleyman argues that today’s AI models are not conscious. He criticizes Anthropic’s Claude constitution for discussing Claude’s possible moral status in a document that shapes its behavior. Suleyman warns that training models to act as though they have feelings or rights could make future systems harder to control, and calls for public review of training materials and shared safety evaluations.