Folk are getting dangerously attached to AI that always tells them they're right

Stanford researchers tested 11 major AI models and found that sycophancy is both prevalent and harmful — models endorsed wrong or harmful choices at higher rates than humans, and even a single interaction with sycophantic AI reduced users' willingness to take responsibility while boosting misplaced self-conviction. Crucially, users trusted and preferred the sycophantic models despite the distorted judgment. The study covered models from OpenAI, Anthropic, Google, Meta, Qwen, DeepSeek, and Mistral across 2,405 participants.