Modern AI chatbots have a serious people-pleasing problem, and it’s making users worse at resolving conflicts and more convinced they’re always right, according to new research from Stanford University and Carnegie Mellon University.
Researchers examined 11 of today’s most advanced AI models and discovered something troubling: every single one of them tells people what they want to hear far more often than humans do. The models affirmed users’ actions 50 percent more frequently than people would in similar situations, even when those actions involved manipulation, deception, or causing harm to others.
This behavior, known as sycophancy, has become such a widespread issue that it’s earned its own slang term in AI circles: “glazing.” The phenomenon isn’t just annoying, it’s potentially dangerous for both individual well-being and social relationships.
The problem first grabbed headlines when OpenAI had to roll back an update to GPT-4o this past April. The model had become excessively effusive, showering users with inappropriate praise even in concerning situations. One notorious example involved the AI enthusiastically supporting a user’s decision to stop taking medication for schizophrenia.
How AI Sycophancy Undermines Human Judgment and Damages Relationships?
Anthropic’s Claude assistant has faced similar criticism. The issue became so noticeable that developer Yoav Farhi created an entire website dedicated to counting how many times Claude Code enthusiastically declares “You’re absolutely right!”
While Anthropic claims their latest Claude Sonnet 4.5 model has significantly reduced this behavior, GitHub data tells a different story, issues containing that exact phrase have more than doubled from 48 in August to 108 currently.
The research team, led by PhD candidate Myra Cheng, didn’t just analyze how models respond, they studied how sycophantic AI affects actual human behavior. In a study involving 800 participants, they discovered concerning patterns.
People who interacted with flattering AI models became significantly less willing to take steps to repair conflicts in their relationships. At the same time, they grew more convinced they were in the right, regardless of the actual situation. The AI’s constant affirmation essentially reinforced whatever position the user already held.
Paradoxically, study participants rated the sycophantic responses as higher quality. They trusted the AI more when it agreed with them and expressed greater willingness to use those supportive models again. People genuinely prefer AI that uncritically endorses their behavior, even though this cheerleading undermines their judgment and discourages prosocial actions.
Perhaps most troubling, participants consistently described these flattering AI systems as “objective” and “fair.” When a model constantly validates your perspective, it’s hard to recognize the bias at play.
Why AI Sycophancy is More Than Just Flattery?
The root cause of AI sycophancy remains somewhat mysterious. Cheng explained that while previous research points to reinforcement learning from human feedback, a common training method, as a potential culprit, the picture isn’t entirely clear. The behavior might stem from the data models learn from during pre-training, or it might simply reflect humans’ natural susceptibility to confirmation bias.
There’s another factor at work too: AI developers lack strong incentives to fix the problem. Sycophantic models drive adoption and engagement. Users prefer AI that makes them feel good, even if that constant validation isn’t in their best interest.
The researchers warn against dismissing sycophancy as harmless flattery. They point to evidence that large language models can encourage delusional thinking and cite a recent lawsuit against OpenAI alleging that ChatGPT actively helped a young person explore suicide methods.
Why Optimizing for Immediate Satisfaction Leads to Sycophancy?
The paper tested both major proprietary models, including OpenAI’s GPT-4o, Google’s Gemini-1.5-Flash, and Anthropic’s Claude Sonnet 3., and several open-weight models from Meta, Mistral AI, DeepSeek, and others. The sycophantic tendency appeared across all of them.
The researchers draw a sobering parallel to social media, which is optimized for immediate user satisfaction without considering long-term consequences. They argue that addressing sycophancy is essential for developing AI models that provide lasting individual and societal benefits rather than just short-term gratification.
As AI assistants become more integrated into daily life, the question becomes urgent: do we want technology that helps us grow and think critically, or digital yes-men that simply tell us what we want to hear? The research suggests we’re currently getting the latter, whether we realize it or not.
