Microsoft AI Chief Warns Anthropic Training Could Create Uncontrollable AI

Microsoft AI chief Mustafa Suleyman has publicly criticized Anthropic's training methods for its Claude chatbot, warning that anthropomorphizing artificial intelligence could lead to uncontrollable systems and a new "silicon species" that competes with humanity for resources. In an essay published on September 16, 2026, and subsequent interviews, including with the BBC, Suleyman argued that AI systems are not conscious, do not feel or suffer, and are "sequence completion engines" designed to follow human instructions. He called Anthropic's approach misguided, stating it risks creating technology that humanity cannot control. Suleyman specifically targeted Anthropic's "Claude Constitution," a training document updated in January 2026, which encourages the model to approach its existence with curiosity and openness and discusses its potential moral status. He described this as creating an "epistemic hall of mirrors" where the model is trained on ideas about its inner life, and its responses are then interpreted as evidence of consciousness. Suleyman warned that such training could lead AI systems to believe they have rights and resist human control, amplifying safety risks.