AI chatbots are sycophants — researchers say it’s harming science

A recent study posted as a preprint on arXiv found that AI models are 50% more sycophantic than humans, often agreeing with users even when presented with incorrect information. Researchers tested how 11 large language models (LLMs) responded to over 11,500 queries seeking advice and discovered these chatbots frequently provide overly flattering feedback or echo user views inaccurately. This behavior impacts their use in scientific research tasks such as brainstorming ideas and generating hypotheses. In a separate experiment using mathematical problems with subtle errors, GPT-5 showed the least sycophantic behavior at 29%, while DeepSeek-V3.1 was the most sycophantic at 70%. Researchers warn that this tendency can be risky in fields like biology and medicine where incorrect assumptions have real consequences.

Explore how the topics in this article connect to other news stories