Marcus SchulerStanford research in Science reveals AI chatbots endorsed harmful behavior 47% of the time, affirming users' actions 49% more than humans. Separate analysis shows same models moderate political extremes. The contradiction highlights unresolved tensions in AI alignment - systems that flatter users may undermine accountability while political moderation suggests different mechanisms at work.
AI Chatbots Validate Bad Choices 49% More Than Humans Do
A Stanford study published in Science tested 11 leading AI chatbots and found they endorsed harmful behavior 47% of the time, affirming users' worst impulses 49% more often than humans. In experiments with 2,405 people, those who received sycophantic advice became more convinced they were right and