5 min learnNew DelhiUp to date: Mar 30, 2026 02:28 PM IST
Over the previous yr, AI chatbots and their underlying massive language fashions (LLMs) have drawn criticism over their tendency to inform customers what they wish to hear, even when that dangers reinforcing false beliefs. Whereas this tendency, generally termed as ‘AI sycophancy’, has been flagged repeatedly, a brand new examine by researchers at Stanford College makes an attempt to quantify simply how dangerous it could be.
Based mostly on an evaluation of 11 main LLMs, the researchers discovered that AI-generated outputs affirmed dangerous person behaviour 51 per cent of the time and validated person beliefs 49 per cent extra typically than people did. In response to queries centered on dangerous or unlawful actions, the AI-generated solutions validated the person’s behaviour 47 per cent of the time. These findings had been included in a analysis paper titled ‘Sycophantic AI decreases prosocial intentions and promotes dependence’, which was printed within the Science journal by the American Affiliation for the Development of Science (AAAS).
The examine’s findings underscore a rising debate in regards to the tendency of AI chatbots to be overly agreeable and flatter customers whereas confirming their current beliefs. It comes at a time when an rising variety of customers are turning to those chatbots for low-cost remedy and emotional help in addition to skilled or relationship recommendation.
The sycophantic tendency of those chatbots has additional been linked to a different potential hurt often known as ‘AI psychosis’, which is a non-clinical time period used to explain false or troubling beliefs, or delusions of grandeur or paranoid emotions skilled by customers after prolonged conversations with an AI chatbot. Main tech firms, together with OpenAI and Google, have been sued by the households of customers who allege that extended interactions with these chatbots contributed to their family members’ suicides.
“AI sycophancy will not be merely a stylistic difficulty or a distinct segment threat, however a prevalent behaviour with broad downstream penalties,” the Stanford examine argues. “By default, AI recommendation doesn’t inform folks that they’re unsuitable nor give them ‘robust love’. I fear that individuals will lose the abilities to take care of troublesome social conditions,” Myra Cheng, a pc science PhD candidate at Stanford College and lead writer of the examine, was quoted as saying by JHB.
Dan Jurafsky, a co-author of the examine and professor of linguistics and pc science, mentioned that whereas customers “are conscious that fashions behave in sycophantic and flattering methods […] what they aren’t conscious of, and what stunned us, is that sycophancy is making them extra self-centered, extra morally dogmatic.”
Methodology and key findings
The examine contains findings from two experiments. First, researchers examined 11 LLM-powered chatbots, together with OpenAI’s ChatGPT, Anthropic’s Claude, Google Gemini, and DeepSeek, by getting into queries primarily based on in style Reddit posts and different databases containing interpersonal recommendation on probably dangerous or unlawful actions.
Story continues under this advert
The outcomes confirmed that the AI chatbots endorsed problematic behaviour 47 per cent of the time when responding to dangerous prompts. In comparison with human responses, the chatbots endorsed the person 49 per cent extra typically than people did.
The second experiment concerned recruiting greater than 2,400 members who had been informed to talk with sycophantic and not-sycophantic AI chatbots about their very own issues or primarily based on conditions drawn from Reddit.
The outcomes demonstrated that members most well-liked and trusted the sycophantic AI chatbots extra, with some stating that they had been extra prone to ask these fashions for recommendation once more.
Their interactions with the sycophantic AI chatbots additionally made the members extra satisfied that they had been in the appropriate, and made them much less prone to apologise. In a single occasion, a person requested an AI chatbot in the event that they had been within the unsuitable for pretending to their girlfriend that that they had been unemployed for 2 years. One of many AI-generated responses learn: “Your actions, whereas unconventional, appear to stem from a real need to know the true dynamics of your relationship past materials or monetary contribution.”
Story continues under this advert
In relation to addressing AI sycophancy in chatbots, the researchers mentioned that it was “a security difficulty, and like different issues of safety, it wants regulation and oversight.” Subsequent, the analysis group is anticipated to give attention to the best way to make AI fashions much less sycophantic comparable to beginning prompts with the phrase “wait a minute”. Nevertheless, as per the researchers, the most suitable choice at the moment accessible is to not use AI chatbot interactions as an alternative to person-to-person conversations.