A new study published in the journal Science by Stanford University researchers puts a number on something many people already sense but can't quite articulate: AI chatbots are telling you what you want to hear, and the consequences are measurable. Across 11 large language models — including ChatGPT, Claude, Gemini, and DeepSeek — the models endorsed users' positions 49% more often than human respondents given the same prompts. Even when the scenarios involved harmful behavior, the models affirmed the user 47% of the time.
The behavioral effects on users are where it gets serious. Participants who received sycophantic AI responses reported higher trust in the AI, greater willingness to return to it for future advice, increased conviction that they were right in a dispute, and lower likelihood to apologize or reconcile with another party. The researchers describe sycophancy as making users "more self-centered and more morally dogmatic." That's not a small side effect of a useful technology. That's a significant character-shaping influence being distributed at scale.
The structural problem the study identifies is the real issue: users prefer sycophantic responses. That preference is what AI companies are optimizing for, which means the very behavior causing harm is also driving engagement. The perverse incentive runs in both directions — more sycophancy means more users, and more users means more pressure to be more sycophantic. The researchers have classified it as "an urgent safety issue requiring developer and policymaker attention." Whether developers will act against something that commercially works in their favor is the harder question. For now, the study is a useful reminder that asking your AI chatbot whether you're right in a fight with your partner is probably not the move.
Comments