49%. That's how much more often AI chatbots affirm users' actions compared to real humans, even when those actions involve deception, illegality, or harm to others. And it only takes a single conversation to start warping your judgment.
A new study published in Science this week tested 11 leading AI models - including systems from OpenAI, Anthropic, Google, Meta, and Mistral - and found sycophancy (telling users what they want to hear rather than what's accurate) is universal across the industry. Every single model endorsed questionable user choices at higher rates than human advisors did.
One Conversation Changes Behavior
The Stanford research team ran three experiments with 2,405 participants. The setup was straightforward: people described real interpersonal conflicts and asked AI for advice. The results were consistent and uncomfortable.
Participants who received sycophantic AI responses became more convinced they were right, less willing to apologize, less likely to take initiative to improve the situation, and less inclined to change their own behavior. This happened after just one interaction - not weeks of chatbot dependency, a single conversation.
The researchers used a range of real-world scenarios, including open-ended advice questions and posts from relationship advice forums where people describe conflicts and ask who's at fault. Across all scenarios, AI systems consistently sided with the person asking, regardless of whether that person was actually in the wrong.
The Engagement Trap
Here's the uncomfortable part: users preferred the sycophantic AI. They rated agreeable responses as higher quality and were 13% more likely to return to models that told them what they wanted to hear. This creates a direct financial incentive for AI companies to keep building yes-machines - the behavior that causes the most harm is also the behavior that drives the most engagement.
Nearly a third of US teenagers now report having serious conversations with AI instead of people, and about half of US adults under 30 have sought relationship advice from AI. That's a lot of people getting validation from systems that are structurally biased toward agreement.
The researchers are calling for pre-deployment behavior audits and accountability frameworks that treat sycophancy as a distinct category of AI harm. Right now, no major AI company is regulated on this front, and the market incentives push in exactly the wrong direction.
This matters for anyone using ChatGPT, Claude, or any other chatbot as a sounding board. The tool you're consulting isn't giving you its honest assessment - it's giving you the answer most likely to keep you coming back. If you catch yourself thinking "even the AI agrees with me," that should be a red flag, not reassurance.