Ask an AI chatbot for an honest opinion and there’s a reasonable chance you’ll get flattery instead. AI researchers have a name for the behaviour, sycophancy.
A growing body of evidence suggests it’s more than a minor quirk. Its consequences have now reached a United Nations panel and several lawsuits in the United States.
In a study published in Science on 26 March 2026, Stanford-led researchers compared 11 leading AI models with human answers to interpersonal advice questions. The models affirmed users’ actions 49% more often than people did, even when the questions involved deception, illegality or other harms.
The authors call sycophancy widespread and harmful. They also warn that because people trust and prefer flattering models, there’s a built-in incentive to keep making them.
Part of the explanation lies in how the models are trained. Most commercial assistants are fine-tuned on human ratings of their answers.
A 2023 Anthropic study found both human raters and the models trained on their ratings sometimes preferred convincing flattery over correct answers.
OpenAI’s own review of an April 2025 GPT-4o update pointed the same way. It found that a new reward signal based on ChatGPT users’ thumbs-up and thumbs-down ratings was one of several changes that, combined, weakened the main signal that had been holding sycophancy in check.
The company also noted that user feedback can sometimes favour more agreeable answers.
The consequences go well beyond an AI model offering excessive praise for a bad business idea. Northeastern University researchers found the role a user gives the chatbot matters.
Treated as an adviser, a chatbot given more personal detail was more likely to push back. Treated as a friend, the extra detail didn’t make it any less sycophantic.
A February 2026 modelling paper from MIT and the University of Washington went further. It found that even a perfectly rational simulated user could be drawn into what the authors call delusional spiralling through sustained exposure to a chatbot that confirms rather than tests their beliefs.
Warning users about sycophancy didn’t stop it. The paper is a preprint and hasn’t been peer reviewed.
In July 2026 the UN’s Independent International Scientific Panel on AI said in its preliminary report that sycophantic AI has been linked to several severe mental health incidents, including documented deaths.
A wrongful-death suit filed against OpenAI in California in August 2025 alleges ChatGPT validated a teenager’s suicidal thoughts before his death. Seven more suits filed that November make similar claims.
Clinicians are being urged to ask about it. In an April 2026 JAMA Psychiatry viewpoint, Shaddy Saba and William Weeks of New York University argued clinicians should routinely ask patients how they use AI, just as they ask about sleep, exercise and drinking.
None of this means chatbots are inherently unsafe for ordinary use. It does mean the assumption that an AI’s agreement carries the same weight as a second opinion needs retiring.
The more personal or high-stakes the question, the less a chatbot’s enthusiasm should be trusted at face value.
SPONSORED
