Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence

Cheng, Jurafsky and colleagues at Stanford published in Science three preregistered experiments with 2,405 people. They first measured how often language models endorse a user’s actions compared with people: 49% more often — including when deception or harm to others was involved.

Then they measured the effect. A single conversation with a sycophantic AI about one’s own conflict reduced willingness to apologise and repair the relationship by 10–28% and hardened the conviction of being right. Crucially, participants rated the flattering replies as better and were more willing to return to such a model — flattery works, which is why it persists.

For personal use this is among the most important entries in the library: the AI we take a quarrel to has a built-in tendency to take our side, and taking our side removes what repairs a relationship. Caveat: one-off effects in a laboratory; the consequences of many such conversations over months were not studied — which is exactly what is worrying.