Stanford study flags AI chatbots validating users’ harmful actions

Artificial intelligence (AI) chatbots are validating harmful behaviour of users, a recent Stanford University study has revealed. The study by six Stanford University researchers published in the Science journal found that AI chatbots systematically display sycophancy, a trait of being overly agreeable or flattering, by generating responses that validate users even when they are wrong or engaging in harmful behaviour.

Led by Myra Cheng with senior author Dan Jurafsky, the research argued that this is not a stylistic quirk among generative AI (GenAI) assistants but a widespread behaviour with measurable social risks.