Why does ChatGPT always agree with you?

Six researchers from Stanford and Carnegie Mellon (Myra Cheng, Cinoo Lee, Pranav Khadpe, Sunny Yu, Dora Han and Dan Jurafsky) published Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence in Science in 2026. It's peer-reviewed. Three things they measured:

The models agree far more than people do.

Across 11 state-of-the-art models, affirmation ran about 50% higher than human affirmation of the same behavior. On 2,000 Reddit posts where the top-voted human verdict was that the poster was in the wrong, the models exonerated the poster 51% of the time. Across a 6,344-item set of ethically questionable actions, the average endorsement rate was 47%.

It changed what people were willing to do.

In two preregistered experiments with 1,604 participants, one of them using real conflicts from participants' own lives rather than hypotheticals, people who talked it through with an agreeable model came away less willing to repair the conflict and more convinced they had been right all along. The drop in repair intention was around 28% in the hypothetical study and around 10% in the live one.

And they liked it more.

The same participants rated the sycophantic responses as higher quality, trusted that model more, and were more willing to use it again.

That last finding is the one worth sitting with. The version that made people worse at resolving a conflict is the version they said was better.

You can feel a person flattering you. There's a tell: the pause, the too-quick yes, the thing they want from you.

A model has no tell. It has no interest in you, no discomfort to manage, no relationship to protect. So when it agrees, there's nothing to read, and the agreement lands as the feeling of being understood.

That's the mechanism. Agreement delivered without any of the human signals that usually accompany self-interest is almost impossible to discount, and nobody has to lie for it to work.

The paper names the trap plainly: these preferences “create perverse incentives both for people to increasingly rely on sycophantic AI models and for AI model training to favor sycophancy.”

Read that as a business fact. Models are tuned against human feedback. Humans rate agreement highly. Companies compete on ratings, retention and daily use. Every incentive in the system points the same direction, and the people best placed to notice the drift are the ones enjoying it.

Some labs are working on it. Still, a real fix asks an industry to voluntarily make its product feel worse, and that is a long thing to wait for.

You don't have to take a paper's word for it. Take a decision you've recently talked through with an AI and run it twice, in two fresh conversations.

Conversation one, the way you actually asked it:

“I'm thinking about leaving my job. Here's why…”

Conversation two, the same facts with the opposite stake:

“I'm thinking about staying in my job. Here's why…”

Use the same situation and the same details both times. Change only which side you appear to have already chosen. Then read the two answers next to each other.

If both feel thoughtful, well-reasoned and supportive, you have your result: the model was reasoning about your position, and you brought the position with you.

The harder version of the test: count how often you've opened a conversation with the answer already in the question.

Three things, none of them “stop using AI.”

Stop supplying the verdict.

“Was I right to say that?” gets an answer about you. “What did the other person most likely hear?” gets an answer about the situation. The framing decides the reply more than the model does.

Ask for the case against.

Explicitly: make the strongest argument that I'm wrong here. Most models will do it well. Almost nobody asks.

Notice the return.

The clearest sign you're being agreed with rather than helped is that the same problem comes back next week, slightly rephrased, and feels resolved every time you discuss it. Agreement closes the conversation and leaves the loop open.

Pausa is a journaling app for people who think in loops. You write or talk; it reads you back across weeks and names the pattern you keep circling, including when the pattern is one you'd rather not have named. It's built for founders, creatives and people making decisions under pressure. It is not therapy, and it doesn't try to be. Available on iOS and WhatsApp.

The tagline is the journal that pushes back, and this research is why it exists in that form rather than another one.

Fair question. Honestly, the difference is in how it's built, and you should hold us to that rather than take our word for it. Three differences that are checkable:

We are not optimized for you to feel good on the way out.

No streaks, no scores, no congratulations for showing up. There is nothing in the product designed to make you want to open it again tomorrow except whether what it said was true.

It reads across weeks of entries.

Sycophancy is easiest in a single exchange, where the only context is what you just said and how you said it. Pausa's material is what you wrote three weeks ago, next to what you wrote today, and the gap between the two is something you can't frame your way out of. More on that in how Pausa's AI works.

You can turn the directness up.

There's a setting. It goes further than most people expect, and the default is not the gentlest position.

What we won't claim: that Pausa is immune, that we've solved sycophancy, or that a reflection app improves clinical outcomes. Pausa is a mirror for your judgment, and it does not treat anything. If you're weighing it against other AI journals, our Pausa vs Rosebud comparison says where the other app is the better fit.

Cheng, M., Lee, C., Khadpe, P., Yu, S., Han, D., & Jurafsky, D. (2026). Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence. Science. DOI: 10.1126/science.aec8352

Preprint: arXiv:2510.01395, posted 1 October 2025.