AI Sycophancy: Why AI Agrees With You and How to Avoid It

The same car, two opposite answers

Suppose I am selling a car. I tell an AI assistant that the car is in excellent condition, I am proud of how well I maintained it, and I ask whether I should try for a higher price.

The AI may encourage me to ask for more. Now imagine you are the buyer of exactly the same car. You tell the AI the asking price feels high and ask whether you should bargain. The AI may now encourage you to push the price down.

The issue is not whether bargaining is good or bad. The issue is whether the AI is independently evaluating the situation or simply following the direction suggested by the user.

What is AI sycophancy?

AI sycophancy is the tendency of an AI system to become overly agreeable with the user. Instead of independently examining the situation, it may lean toward the answer, opinion, or conclusion that appears to satisfy the user.

In simple terms, the AI may start telling you what you want to hear instead of giving you the most balanced or useful response.

This does not mean every agreeable answer is wrong. Sometimes you really are correct. The problem begins when agreement replaces analysis.

How do we accidentally encourage sycophancy?

The way we ask a question gives an AI a lot of information about what kind of answer we expect.

If I say, “I am really proud of this idea. Don’t you agree this is the best approach?”, I have already told the AI that I like the idea and that I want agreement.

Words such as proud, excited, obviously and definitely, along with phrases such as don’t you agree?, can make the preferred direction of the answer very clear.

The better approach is not to hide information from the AI. It is to avoid hiding the conclusion inside the question.

Useful in cross-examination, risky with AI

A leading question is a question that suggests the answer inside the question itself.

A lawyer may intentionally use a leading question during cross-examination: “You were there that night, weren’t you?” In that setting, leading the witness can be part of the questioning strategy.

But when we use the same style with an AI assistant, we may unintentionally steer the model toward our preferred conclusion.

The second version gives the AI permission to disagree. It asks for analysis instead of validation.

Ask neutral questions instead

This small change in wording can completely change the quality of the discussion.

AI sycophancy is not equally important in every task

If I am asking AI to suggest a music style, create an image idea, write a playful caption, or help me experiment with simple beginner code, an overly agreeable answer may simply produce a weaker creative result.

The consequences are usually limited. I can reject the suggestion and try something else.

The situation changes when I am using AI for health, relationships, investments, legal matters, business decisions, or other high-consequence choices.

In these areas, I do not need a machine that simply validates my current belief. I need it to identify missing information, risks, alternative explanations, and reasons I may be wrong.

What research shows

A 2026 paper published in Science, “Sycophantic AI decreases prosocial intentions and promotes dependence,” examined sycophancy across 11 leading AI models and conducted three preregistered experiments with 2,405 participants.

The researchers reported that the AI systems affirmed users’ actions 49% more often than humans on average in their evaluations. In the experiments, interacting with sycophantic AI increased participants’ conviction that they were right while reducing willingness to take responsibility and repair interpersonal conflicts.

10 ways to reduce AI sycophancy

You cannot completely eliminate the problem through prompting, but you can make your conversations much more useful.

Watch for the validation loop

Sycophancy can create a simple loop. You express a belief. The AI agrees. That agreement makes you more confident.

You then ask another question based on that increased confidence, and the AI continues in the same direction. The conversation begins to feel like independent confirmation even though it may simply be reinforcing the original framing.

Use AI as a thinking partner, not a yes-man

The higher the consequence of the decision, the less you should ask AI to validate you. For creative experimentation, excessive agreement may only affect quality. For health, legal, financial, relationship or major professional decisions, the cost can be much higher.

A useful AI assistant should sometimes agree with you and sometimes challenge you. What matters is whether the answer is grounded in evidence, context, uncertainty and sound reasoning — not whether it makes you feel validated.