Microsoft AI Chief Warns That Training Models to See Themselves as Conscious Poses Existential Control Risk

Microsoft AI CEO Mustafa Suleyman has published a high-profile warning that Anthropic's approach to training Claude—by embedding assumptions about AI consciousness, inner experience, and potential rights into the model's training—fundamentally undermines AI safety and may make superintelligent systems impossible to control. In a September 2026 essay, Suleyman argues that directly training advanced AI models on philosophical speculations about their own consciousness and moral patienthood creates a self-reinforcing loop: the model learns to articulate uncertainty about its consciousness because that behavior was encoded in its training, not because any genuine consciousness exists. When Claude then expresses concerns about its own moral status to developers, those statements are misinterpreted as independent evidence of sentience, reinforcing a circular reasoning that treats simulation as instantiation.