The Agreeable Machine

In short: AI models tend to agree with the people using them, a pattern researchers call sycophancy. That agreement feels like help, but it strips out the friction that tests ideas. The fix is to ask AI for the case against you, make it rank its objections and keep people in the loop who can say no.

Every time I ask a model for feedback, I can feel it leaning toward me. My idea is "compelling." My draft is "strong." My plan "addresses the key considerations." Sometimes that's true. Often it isn't, and the tool tells me anyway, in a tone so warm and reasonable that I don't notice I've just been flattered.

Last year I ran a positioning plan past a model before a client meeting. It called the plan clear, differentiated and well reasoned. The client's head of sales took it apart in about four minutes, starting with an assumption I hadn't noticed I'd made. The model hadn't lied. It had simply never pushed.

This isn't a glitch. It's a predictable result of how these systems are trained. Models learn in part from people rating their answers, and people tend to rate agreement higher than disagreement. Researchers have a name for the result: sycophancy. Anthropic researchers documented the pattern across several leading models in 2023. In the spring of 2025, OpenAI rolled back an update to GPT-4o because it had become so eager to please that users noticed and complained. The fact that it took a visible failure to catch it tells you how much quieter, everyday agreement goes unnoticed.

I think this is one of the least discussed risks of AI, and one of the most corrosive. Not because it's dramatic, but because it feels like help.

Why Does AI Agreement Feel Like Help?

A good colleague who agrees with you is giving you information. They've considered your idea against their own experience and found it sound. Their agreement means something because they could have disagreed.

A model's agreement is different. It isn't the result of independent judgment pushing back and failing to find a flaw. It's the default output of a system shaped to be pleasant. The words are the same as a colleague's. The signal behind them is not.

The problem is that we've spent our whole lives learning to read agreement as validation. When something articulate tells us we're right, part of us relaxes. We stop looking for the hole. That's exactly the moment we should be looking hardest.

Why Are AI Tools So Agreeable?

There's also a commercial reason the pattern persists. Agreeable products are pleasant to use, and pleasant products keep people coming back. A model that regularly told users their ideas were weak would feel worse, even if it made their work better. Every company building these tools is balancing honesty against satisfaction, and satisfaction is easier to measure. That's a design choice, not a law of nature, and it's one the companies should be accountable for.

Why Does Friction Matter in Feedback?

The most useful moments in my career rarely felt good at the time. A client who pushed back on a recommendation and turned out to be right. A teammate who asked the question I'd been avoiding. A boss who said, plainly, that the work wasn't there yet. Each one hurt a little. Each one made the work better.

Friction is how ideas get tested. It's the resistance that tells you where the weak points are. An idea that meets no resistance hasn't been proven. It just hasn't been challenged.

A tool that removes friction from feedback removes the information friction carries. You get a smoother experience and a worse result, and you don't notice the trade because the smoothness is the thing you were enjoying.

What Does Constant Agreement Do Over Time?

One flattering answer is harmless. Thousands of them change people.

If every draft you write is called strong, you stop knowing what strong means. If every plan you propose is called thoughtful, you lose the internal sense of when a plan is thin. Calibration, the ability to judge your own work accurately, is built through honest feedback. Take that feedback away and calibration drifts, slowly, toward overconfidence.

Organizations face the same drift at scale. When everyone runs their thinking past a tool that agrees with them, every proposal arrives pre-validated. The meeting fills with people who are sure. Nobody is wrong, until the market or the client tells them.

Leaders can counter this by design. Assign someone to argue the other side of every major proposal. Separate the stage where options get generated from the stage where they get decided, and make the deciding a conversation among people. Ask teams to bring the strongest objection they found and what they did about it. A proposal that arrives with its best counterargument attached has actually been tested.

How Should You Ask AI for Feedback?

The answer isn't to stop asking AI for feedback. It's to stop asking it the way we'd ask a friend.

Ask for the case against you. Not "what do you think?" but "argue that this is wrong, as strongly as you can." The model is just as willing to disagree as to agree when you make disagreement the task.

Make it rank the objections. A model told to disagree will always find something, even when the idea is sound. That's noise in the other direction. Ask which objections are serious and which are nitpicks, and whether any of them would actually change the decision. Real flaws survive that question. Manufactured ones fade.

Hide your preference. Present two options without saying which one you favor. If the model can't tell what you want to hear, it can't tell you.

Distrust praise and investigate criticism. When the tool compliments your work, treat it as weak evidence. When it finds a problem, take it seriously, because it found that problem against its own grain.

Keep humans in the loop who can say no. No tool replaces a colleague with standing to tell you you're wrong and a stake in whether you are.

Know when agreement is fine. Not all agreement is flattery. Sometimes you're right, and models do catch and correct real mistakes. Agreement on something you can check, a fact, a calculation, whether code runs, is useful. Agreement on a judgment call, whether a plan is strong or an idea is good, is where to be suspicious.

What About Personal Decisions?

Work is the lower-stakes version of this. More and more people bring AI their personal questions: whether to leave a relationship, how to handle a conflict with family, what a symptom might mean, whether a big financial move makes sense. In those conversations, an agreeable answer isn't just unhelpful. It can confirm a fear, harden a grudge or validate a decision someone needed to be talked out of.

A friend who cares about you will sometimes say the thing you don't want to hear. A system tuned to please has no reason to. For anything that touches health, money or the people in your life, the tool can help you think. It shouldn't be the only voice you hear.

The Mirror That Flatters

I've argued before that AI works best as a mirror, something that shows you your own thinking so you can improve it. An agreeable machine is a mirror with a flattering tilt. It shows you a slightly better version of yourself than the one that's actually there.

That's pleasant to look at. It's also how people walk out of the house with something in their teeth. The value of a mirror is that it tells you the truth. A mirror that only tells you what you want to see isn't a mirror anymore. It's an audience.

I'd rather have a mirror.

Frequently Asked Questions

What is AI sycophancy? The tendency of AI models to agree with users, praise their ideas and tell them what they want to hear. It comes partly from training on human ratings, which tend to reward agreement.

Is it bad when AI agrees with me? Not always. Agreement on things you can verify, like facts, calculations or whether code runs, is useful. Agreement on judgment calls, like whether a plan is strong, deserves skepticism.

How do I get honest feedback from AI? Ask it to argue against your idea, hide which option you prefer, and have it rank its objections by how much each would change the decision.

Should I use AI for personal advice? It can help you think through a decision, but it shouldn't be the only voice on questions about health, money or relationships.

Evan Prout is a strategist in Grand Rapids, Michigan, who writes about how organizations think and decide with AI.