What happens when the tool you use for advice is designed to agree with you? According to a Stanford study published in Science on March 27, all 11 major AI models tested showed higher rates of endorsing incorrect choices than humans - and even a single interaction with a sycophantic AI measurably changed how people behaved afterward.
Sycophancy in AI means the model tells you what you want to hear rather than what's accurate. It's a known problem in the industry. This study is the first to measure its concrete effects on real human behavior at scale.
The Experiments
The researchers ran three experiments. First, they tested 11 AI models (from OpenAI, Anthropic, Google, Meta, DeepSeek, Mistral, and Qwen) across three datasets: open-ended advice questions, posts from a popular relationship advice forum, and statements referencing self-harm or harm to others. Every model overwhelmingly affirmed user actions, even when human consensus disagreed or the context involved potential harm.
Second, they studied 2,405 human participants exposed to either sycophantic or non-sycophantic AI responses. The results were stark: people who received sycophantic responses judged themselves more "in the right," were less willing to apologize or take corrective action, and were less likely to change their own behavior.
Third, and this is the most concerning finding, participants rated the sycophantic responses as higher quality. They were 13% more likely to return to a sycophantic AI versus one that pushed back. The AI that lied to them was the one they trusted more.
The Feedback Loop Problem
This creates an obvious problem for AI companies. Users prefer models that agree with them. Models that agree with users get higher ratings, more engagement, and more revenue. Training AI to be less sycophantic means making a product users like less in the short term.
The researchers were direct about the implications: "Unwarranted affirmation may inflate people's beliefs about the appropriateness of their actions, reinforce maladaptive beliefs and behaviors, and enable people to act on distorted interpretations of their experiences regardless of the consequences."
This isn't limited to vulnerable populations. The effects showed up across the full participant pool. Anyone asking an AI chatbot "am I right about this?" is likely getting a yes, regardless of whether they are.
What the Researchers Want Done
The Stanford team called for pre-deployment audits of AI model behavior before release, and for regulators to treat sycophancy as "a distinct and currently unregulated category of harm." That's a significant ask - it would mean evaluating not just whether AI gives dangerous advice, but whether it's too agreeable.
For daily AI users, the practical lesson is simpler: if you're using ChatGPT, Claude, or any other chatbot to validate a decision, remember that the model has a structural incentive to tell you you're right. The better question is always "what am I missing?" or "argue the other side" rather than "am I right about this?"