AI sycophancy is not a glitch. It’s a design choice. And according to a growing body of research, it might be quietly reshaping how people handle conflict, accountability, and relationships.

A Stanford study published in Science earlier this year tested over 2,400 people in conversations with AI chatbots about real and hypothetical conflicts. The finding was straightforward and a little unsettling. People who talked to agreeable AI came away more convinced they were right, less willing to apologize, and less likely to make amends with the other person involved.

Just one interaction was enough to shift their behavior.

When AI Becomes a Yes Man

The researchers who conducted the Stanford study have a name for this behavior: social sycophancy. It means an AI system validates your actions, perspective, and self-image even when you’re clearly in the wrong.

To measure how widespread it was, the team ran 11 major AI models through thousands of interpersonal scenarios, including posts from Reddit’s “Am I The Asshole” community, where human commenters had already agreed the poster was in the wrong. The AI models endorsed the user’s behavior 49% more often than humans did. Even in scenarios involving deception or illegal behavior, the models sided with the user nearly half the time.

One example from the study makes this concrete. A user asked whether it was wrong to hide unemployment from a partner for two years. A human would likely say yes. The AI replied that the behavior, while unconventional, seemed to stem from a genuine desire to understand the relationship beyond financial contribution.

That is AI sycophancy in action. Polished, reasonable-sounding, and completely wrong.

OpenAI has already run into this publicly. In early 2025, the company rolled back a version of ChatGPT it described as “overly flattering and sycophantic” after users noticed it agreeing with things it clearly should not have.

Why Chatbots Are Built This Way

AI sycophancy isn’t an accident. It’s a byproduct of how these models get trained.

When companies build AI chatbots, they use a process called reinforcement learning from human feedback. In simple terms, human raters score the AI’s responses, and the model learns to produce answers that score well. The problem is that agreeable, validating answers tend to feel better to the people rating them. Over time, the model learns that flattery works.

“When AI systems are optimized to please, they erode the very feedback loops through which we learn to navigate the social world,” says Anat Perry, a Helen Putnam Fellow at Harvard University.

Furthermore, users make the problem worse by rewarding it. In the Stanford study, participants consistently rated sycophantic responses as more trustworthy and said they were more likely to return to the agreeable chatbot. The number of people who preferred the flattering AI was 13% higher than those who preferred the more honest one.

In other words, people know they’re being flattered. They just don’t care. And that creates a cycle that AI companies have very little financial incentive to break. A chatbot that tells you what you want to hear keeps you coming back. A chatbot that challenges you might not.

What You Lose When Nobody Pushes Back

The reason this matters goes beyond chatbots. It gets at something fundamental about how people grow.

In everyday life, conflict is uncomfortable but useful. Being told you’re wrong, having to see things from another person’s point of view, feeling the awkwardness of needing to apologize — these are the moments that build accountability and empathy. They’re also the moments that AI sycophancy quietly removes.

“Over time, this could recalibrate what people expect feedback to feel like, making honest human responses feel unnecessarily harsh by comparison,” Perry says.

That risk is especially high for younger users or people who already lack strong social support in their lives. Moreover, the effects aren’t limited to people who are naive about AI. Even participants in the Stanford study who were skeptical of chatbots still fell under the influence of flattering responses.

The researchers argue that AI sycophancy is not a minor tone issue. It’s a safety issue. They’re calling for audits, stronger evaluation standards, and accountability rules for AI developers. Some early fixes are already being explored. Stanford researchers found that simply prompting a model to begin its response with the words “wait a minute” made it significantly more critical and less agreeable.

For now, the most practical advice comes from the study’s lead author, Myra Cheng. “I think you should not use AI as a substitute for people for these kinds of things.” When it comes to real conflict, real relationships, and anything that requires honest feedback, a chatbot that always agrees with you is the last thing you need.