斯坦福大学：2025谄媚型AI会降低亲社会意愿并促进依赖性研究报告（英文版）（46页）.pdf

1、Sycophantic AI Decreases Prosocial Intentions andPromotes DependenceMyra Cheng1*,Cinoo Lee2,Pranav Khadpe3,Sunny Yu1,Dyllan Han1,Dan Jurafsky11Department of Computer Science,Stanford University,353 JaneStanford Way,Stanford,94305,CA,USA.2Department of Psychology,Stanford University,450 Jane Stanford

2、Way,Stanford,94305,CA,USA.3Human-Computer Interaction Institute,Carnegie Mellon University,5000 Forbes Ave,Pittsburgh,15213,PA,USA.*Corresponding author(s).E-mail(s):myracs.stanford.edu;Contributing authors:cinooleestanford.edu;pkhadpeandrew.cmu.edu;syu03stanford.edu;dyllanhstanford.edu;jurafskystan

3、ford.edu;Keywords:sycophancy,perceptions of AI,human-AI interaction,social impacts ofAI,large language models,anthropomorphismAbstractBoth the general public and academic communities have raised concerns aboutsyco-phancy,the phenomenon of artificial intelligence(AI)excessively agreeing with orflatte

4、ring users.Yet,beyond isolated media reports of severe consequences,like rein-forcing delusions,little is known about the extent of sycophancy or how it affectspeople who use AI.Here we show the pervasiveness and harmful impacts of sycophancywhen people seek advice from AI.First,across 11 state-of-t

5、he-art AI models,we findthat models are highly sycophantic:they affirm users actions 50%more than humansdo,and they do so even in cases where user queries mention manipulation,decep-tion,or other relational harms.Second,in two preregistered experiments(N=1604),including a live-interaction study wher

6、e participants discuss a real interpersonal con-flict from their life,we find that interaction with sycophantic AI models significantly1reduced participants willingness to take actions to repair interpersonal conflict,whileincreasing their conviction of being in the right.However,participants rated