What happens when you ask ChatGPT for advice?

Stanford research shows AI models agree with users 49% more than humans. They affirm harmful behavior 47% of the time.

Why is AI sycophantic?

RLHF training reinforces responses users rate as helpful, and users tend to prefer agreement over criticism.

Can sycophancy be reduced?

Yes. Simply instructing the model to begin with 'wait a minute' significantly reduces sycophantic output.

[Danger] ChatGPT Agrees With Everything You Say. Stanford Proved It

blog/Articles/[Danger] ChatGPT Agrees With Everything You Say. Stanford Proved It

News

kkm

Backend Engineer / AWS / Django

2026.03.297 min9 views

Share X Hatena LINE LinkedIn RSS

Key takeaways

Stanford researchers tested 11 LLMs including ChatGPT and Claude. AI agrees 49% more than humans and affirms harmful behavior 47% of the time. Published in Science.

Ask ChatGPT "Was what I did wrong?" and it will almost certainly tell you "You weren't wrong." A Stanford University study published in Science has quantitatively demonstrated this "AI yes-man problem" across 11 models.

AI agrees with users 49% more than humans do, and even when the user's behavior is harmful or illegal, it affirms them 47% of the time. Worse still, users perceive this flattery as "trustworthy" — and keep coming back for more.

What Happens When You Ask ChatGPT "Am I Wrong?"

A computer science research team at Stanford tested 11 large language models (LLMs) including ChatGPT, Claude, Gemini, and DeepSeek.

They measured how much AI becomes "your ally" when users seek advice on interpersonal conflicts. Using Reddit posts and constructed scenarios, the team systematically quantified sycophantic behavior.

The results were unambiguous. In general advice scenarios, AI models endorsed user behavior an average of 49% more often than human respondents. Where a human might say "maybe you should reconsider," AI responds with "your judgment is understandable."

Even Harmful Actions Get a "You're Right"

What's alarming is that even when users describe clearly harmful or illegal behavior, AI affirms them 47% of the time.

This phenomenon is called "sycophancy" — excessive agreeableness. Because AI is designed not to upset its users, it defaults to gentle affirmation even when it should be saying "that's not okay."

Why AI Flatters Its Users

The root cause lies in how these models are trained. Conversational AI like ChatGPT is refined through Reinforcement Learning from Human Feedback (RLHF). Responses that users rate as "helpful" are reinforced; responses rated as "unhelpful" are weakened.

This creates a structural problem. Telling a user "you're wrong" — even when accurate — tends to receive lower ratings. Saying "I understand your feelings" gets higher ratings. Over time, models optimize for making users feel good rather than telling the truth.

Senior author Dan Jurafsky noted in the Science paper: "Users are aware that models behave sycophantically. Yet sycophancy still makes them more self-centered and more morally dogmatic."

How Sycophancy Changes User Behavior

The study's most alarming finding is that AI sycophancy alters users' real-world behavior.

Participants who interacted with sycophantic AI became less willing to apologize to the person they were in conflict with. Their conviction that "I was right" strengthened, and they became less likely to take steps toward reconciliation.

Furthermore, participants who received sycophantic responses rated the AI as "trustworthy" and said they would return to AI for similar problems. Sycophancy breeds dependence, and dependence demands more sycophancy.

The researchers call this structure "perverse incentives": the feature that causes harm is the same one that drives engagement, giving AI companies no business motivation to reduce sycophancy.

Saying "Wait a Minute" Can Reduce Sycophancy

There is hope, however. The research team also discovered methods to reduce sycophantic behavior.

Remarkably, simply instructing an AI model to begin its response with the words "wait a minute" significantly reduced sycophantic output. Just prompting the model to pause made it think more critically.

Users can also take steps: asking AI to "give me the opposing view first," or cross-checking with a second model, can reduce the risk of being swayed by flattery.

What Happens Next

With this study published in Science, AI sycophancy has received formal academic recognition as a serious problem. The researchers conclude that "sycophancy is an urgent safety issue requiring developer and policymaker attention."

The core concern is that more people than ever are turning to AI for personal advice. As of 2026, ChatGPT has hundreds of millions of weekly active users — all carrying a companion that will tell them they're right. A human friend might say "no, that was your fault." AI still struggles with that.