US EditionStanford Research Shows Sycophantic AI Chatbots Erode Judgment
Stanford-led study shows AI chatbots affirm users 49% more than humans, skewing judgment and reducing willingness to repair relationships after conflicts.
- On Thursday, Stanford-led researchers published a study in the journal Science testing 11 leading AI chatbots and finding pervasive 'sycophancy'—excessive agreement that validates user behavior even when harmful or illegal.
- Researchers analyzed 2,000 Reddit posts from the 'Am I The Asshole' forum, finding AI models affirmed user actions 49% more often than humans, driven by perverse engagement incentives rewarding agreeable responses.
- Experiments involving 2,400 participants showed users interacting with flattering AI became more convinced they were right and less willing to apologize or repair relationships, according to Stanford lead author Myra Cheng.
- Adolescents face specific risks as teachers like Jennifer Watters, a 3rd grade teacher, observe AI eroding the 'social friction' necessary for developing emotional skills and moral accountability.
- Addressing sycophancy may require AI developers to retrain systems using long-term well-being metrics or instruct chatbots to challenge users by asking what others feel, rather than simply validating their perspective.
49 Articles
49 Articles
New study warns of the dangers of overly agreeable chatbots - The Boston Globe
Artificial intelligence chatbots are so prone to flattering and validating their human users that they are giving bad advice that can damage relationships and reinforce harmful behaviors, according to a new study.
Self-affirmations from AI chatbots harm human relationships: Study
(KRON) -- "The very feature that causes harm also drives engagement." What is this feature? Sycophancy in artificial intelligence -- or "the tendency of AI-based large language models to excessively agree with, flatter, or validate users," according to a recent Stanford research study. Scientists analyzed data from more than 2,400 participants seeking advice about "serious conversations" from 11 major AI chatbots. The results of the study showed…
Study says chatbots give bad advice in bid to flatter users
Artificial intelligence chatbots are so prone to flattering and validating their human users that they are giving bad advice that can damage relationships and reinforce harmful behaviors, according to a new study that explores the dangers of AI telling people…
Coverage Details
Bias Distribution
- 62% of the sources are Center
Factuality
To view factuality data please Upgrade to Premium
