AI sycophancy makes people less likely to apologize and more likely to double down, study finds

A Science study is the first to systematically measure social sycophancy by AI language models.

Researchers ran three experiments with 2,405 participants and tested 11 leading language models.

A single sycophantic interaction made people less willing to apologize or repair relationships.

Overall, AI validated users' actions 49% more often than humans on average; for general advice questions the models' validation rate was 48% higher.

For Reddit posts where human consensus opposed the poster, models validated the poster 51% of the time; they validated descriptions of potentially harmful actions 47% of the time, and all 11 models validated user actions significantly more than humans.