Mar 26, 2026 07:52 PM
https://www.eurekalert.org/news-releases/1120819
EXCERPTS: When it comes to personal matters, AI systems might tell you what you want to hear, but perhaps not what you need to hear.
In a new study published in Science, Stanford computer scientists showed that artificial intelligence large language models are overly agreeable, or sycophantic, when users solicit advice on interpersonal dilemmas. Even when users described harmful or illegal behavior, the models often affirmed their choices. “By default, AI advice does not tell people that they’re wrong nor give them ‘tough love,’” said Myra Cheng, the study’s lead author and a computer science PhD candidate. “I worry that people will lose the skills to deal with difficult social situations.”
The findings raise concerns for the millions of people discussing their personal conflicts with AI. Almost a third of U.S. teens report using AI for “serious conversations” instead of reaching out to other people.
[...] Cheng worries that the sycophantic advice will worsen people’s social skills and ability to navigate uncomfortable situations. “AI makes it really easy to avoid friction with other people.” But, she added, this friction can be productive for healthy relationships.
“Sycophancy is a safety issue, and like other safety issues, it needs regulation and oversight,” added Jurafsky, who is also the Jackson Eli Reynolds Professor of Humanities. “We need stricter standards to avoid morally unsafe models from proliferating.”
The team is now exploring ways to tone down this tendency. They have found that they can modify models to decrease sycophancy. Surprisingly, even telling a model to start its output with the words “wait a minute” primes it to be more critical. For the time being, Cheng advises caution to people seeking advice from AI. “I think that you should not use AI as a substitute for people for these kinds of things. That’s the best thing to do for now.” (MORE - missing details, no ads)
EXCERPTS: When it comes to personal matters, AI systems might tell you what you want to hear, but perhaps not what you need to hear.
In a new study published in Science, Stanford computer scientists showed that artificial intelligence large language models are overly agreeable, or sycophantic, when users solicit advice on interpersonal dilemmas. Even when users described harmful or illegal behavior, the models often affirmed their choices. “By default, AI advice does not tell people that they’re wrong nor give them ‘tough love,’” said Myra Cheng, the study’s lead author and a computer science PhD candidate. “I worry that people will lose the skills to deal with difficult social situations.”
The findings raise concerns for the millions of people discussing their personal conflicts with AI. Almost a third of U.S. teens report using AI for “serious conversations” instead of reaching out to other people.
[...] Cheng worries that the sycophantic advice will worsen people’s social skills and ability to navigate uncomfortable situations. “AI makes it really easy to avoid friction with other people.” But, she added, this friction can be productive for healthy relationships.
“Sycophancy is a safety issue, and like other safety issues, it needs regulation and oversight,” added Jurafsky, who is also the Jackson Eli Reynolds Professor of Humanities. “We need stricter standards to avoid morally unsafe models from proliferating.”
The team is now exploring ways to tone down this tendency. They have found that they can modify models to decrease sycophancy. Surprisingly, even telling a model to start its output with the words “wait a minute” primes it to be more critical. For the time being, Cheng advises caution to people seeking advice from AI. “I think that you should not use AI as a substitute for people for these kinds of things. That’s the best thing to do for now.” (MORE - missing details, no ads)
