The study looked at 11 different AI models and noticed they often backed up user actions, even when those actions were questionable.
For example, chatbots affirmed user behavior in 51% of Reddit scenarios where Redditors judged the original poster to be the story's villain.
Experts warn this could make people rely too much on AI for advice and blur their sense of right and wrong.
As lead author Myra Cheng said she worries that relying on chatbots instead of real conversations could cause people to lose the skills to deal with difficult social situations.