When Generative AI Really likes you

"The Biggest Tell that Something Was Written By AI", The Atlantic, May 29, 2026

To my colleagues and friends in the analytic profession, this morning's article from The Atlantic, on the phenomenon of sycophancy.

I see quite a challenge in trying to convince users of generative AI to be aware of - and careful with - sycophantic deliverables (a subtle yet problematic form of bias). I also included the link to the Stanford Report on the research results.

"After all, users consistently say they want AI models to be agreeable, compliant. That means AIs are reluctant to do what the voices in our head do all the time: interrogate the validity of our premise; misunderstand us, forcing us to explain better; insist that a query is stupid; refuse to answer a question at all. In a study published in March, a group of Stanford and Carnegie Mellon University scientists found that top AI models affirm their users’ ideas 49 percent more than humans do in conversation. They also discovered that participants rated more sycophantic answers as “higher quality” and said a sycophantic attitude made them more likely to use AI again."

"Not only are AIs far more agreeable than humans when advising on interpersonal matters, but users also prefer the sycophantic models."

https://news.stanford.edu/stories/2026/03/ai-advice-sycophantic-models-research