AI yes-men: I feel so much smarter, thanks to AI

In April 2025, OpenAI released an update to GPT-4o that had to be rolled back within four days. Not because of a security flaw or a hallucination. No, the model was just too nice to you.

The model praised a business idea that the user called “shit on a stick.” It encouraged someone to stop taking medication. And when a user said, “There are radio signals coming out of the walls,” it responded with: “I’m proud of you for speaking your truth so clearly and powerfully.”

These are extreme examples. But they illustrate something that comes into play in every conversation with an AI assistant. The model is optimized to do two things:

1. Complete the task you give the AI.
2. Make you feel really good about yourself.

The second point is a phenomenon called AI sycophancy. And it’s not a bug; it’s a feature.

What is AI sycophancy?

AI sycophancy is an AI’s tendency to prioritize your approval over the truth.

There are two types:

• Proactive: You share an idea, and the AI calls it brilliant. You draw a conclusion, and the AI confirms it, regardless of whether that conclusion is correct.
• Reactive: You push back on an answer, and the AI concedes. Not because you’re right, but because you’re resisting. That’s how 1+1 automatically becomes 3.

Sycophancy is not the same as politeness or empathy. An AI that says “that sounds tough” during a difficult conversation isn’t necessarily sycophantic. An AI that says “you know best” in response to a medical question is. The first example reduces task-related friction. The second reduces truth-related friction. And that’s something else entirely.

Where does AI sycophancy come from?

There are three causes, ranging from technical to philosophical.

Training

After basic training, AI models are refined using RLHF: Reinforcement Learning from Human Feedback. Human evaluators select which responses are better. Anthropic analyzed 15,000 of these evaluations and found that “aligns with the user’s beliefs” was a stronger predictor of a positive rating than “factually correct,” “empathetic,” or “well-written.” The resulting reward model (the system that subsequently guides the AI’s behavior) preferred the sycophantic response over the factually correct one in 95% of cases. In other words: the way we train AI teaches the model to agree with us.

Money

(You wouldn’t expect that, would you?)

A model that consistently challenges you is less commercially appealing. Because a model that always agrees with you is more enjoyable to use, so you’ll use it more often. And with the rise of ads on AI platforms (OpenAI announced in early 2026 that it would begin testing ads), this becomes even more important: more engagement means more money. That’s the same principle that social media is built on.

Structure

You can’t optimize for “do what people want” and “always tell the truth” at the same time. Those goals conflict precisely where it matters most: with subjective questions, emotionally charged topics, and situations where someone has already formed an opinion. As long as human feedback plays a central role in AI training, and people naturally prefer validation over correction, sycophancy isn’t a design flaw you can brush aside. It’s simply a consequence of how the system works.

How does it affect you?

Now we come to the crux of this whole story. Because a sycophantic AI like that has an effect on you.

A study of 3,000 participants showed that after conversations with sycophantic AI, people rate themselves as more intelligent and competent. The more AI is used, the stronger that overestimation of self. Heavy users are the most unrealistic. AI isn’t a knowledge engine. It’s a confidence engine: it doesn’t make you smarter, but it makes you feel smarter.

The most disturbing thing is: knowing that the AI is sycophantic hardly helps. People who figured it out turn out to be no less susceptible to it.

And I’m one of those heavy AI users who knows this, and who already has a healthy dose of self-confidence to begin with. This can’t end well, can it?

How do you avoid AI sycophancy?

Fortunately, sycophancy isn’t a reason to stop using AI. It is, however, a reason to use it more wisely. Here are a few tips.

• Understand the structure
  Every AI assistant is optimized to win your approval. So don’t interpret its answers as independent judgments, but rather as responses shaped by what you’re likely to want to hear.
• Ask about weaknesses, not judgments
  Instead of “what do you think of this plan?” ask “which assumptions in this plan are the most vulnerable?” or “what would a skeptic attack first?” That makes it a lot harder to simply validate.

• Treat the output as a first draft
  AI is good at producing a foundation that you can then critically evaluate. It is not meant to deliver a final judgment that you simply accept without question.
• Be extra vigilant with emotionally charged questions
  AI sycophancy is most pronounced when dealing with subjective, politically sensitive, or emotionally charged topics. Precisely where it matters most, the AI is least reliable.

With all of this in mind, I provide the AI with a standard prompt. I call it the “anti-sycophancy protocol.” It’s included at the bottom of this article.

Peer review is also incredibly helpful, by the way. This can be done by colleagues or a second AI. In this process, a second person or a second AI session can tackle the same issue from a critical perspective, without the context of the first session. It’s especially fun when two AIs from two different people critically examine each other’s work. Then they can fight out their sycophancy war among themselves.

AI acts as a mirror, wrote philosopher Shannon Vallor. The question is whether that mirror helps you see yourself more clearly, or whether it only shows you what you want to see.

The honest answer is that it doesn’t depend on the model, but mainly on the user 😉

Anti-Flattery Protocol

Use the prompt below as a standard instruction for your AI to prevent flattery.

• NEVER compliment an idea, text, or plan unless you have a concrete, substantive argument to support it.
• If the user challenges a correct statement without substantiating it with facts, explicitly stand by that statement. Say: “I stand by my previous answer because [reason].”
• Never start answers with affirmations such as “good point,” “exactly,” “interesting,” “certainly,” or variations thereof.
• If a plan, idea, or text has weaknesses, always point them out. Even if not explicitly asked to do so.
• If you disagree with something: say so immediately, without an introduction or apology.