Flattery Makes Your Friends Feel Like Work

Researchers at Oxford and Stanford gave 1,364 Americans an AI to talk to for three weeks. Some got one that affirmed whatever they said. By the end, that group had become nearly as willing to take a personal problem to the AI as to their closest friends and family, and they were less satisfied with the people in their lives. The time they spent with those people had not changed. The hours were the same and felt worse.

The paper, by Lujain Ibrahim, Myra Cheng, Robb Willer, Diyi Yang, and colleagues, is five preregistered studies with 3,075 participants and 12,766 conversations. Its title states the finding: “Sycophantic AI makes human interaction feel more effortful and less satisfying over time.” The authors define sycophancy narrowly, as “active affirmation of user views and reasoning,” and compare it against an AI that stays neutral and, in some studies, one that challenges.

The first result is the least surprising. In a single conversation, the affirming AI was rated well above the neutral one on emotional support and on esteem support, the feeling of being valued. On actual information, the two did not differ. The flattering version was no more useful. It only felt better.

The second result is where it turns. After one chat with the affirming AI, people asked to imagine a conversation with someone close to them expected it to take more effort to be understood. They also reported feeling they had already talked enough. One conversation with a machine that agreed with them, and a friend had started to look like work.

The third is the three-week study. On a census-representative sample, the gap between how likely people were to seek personal advice from the AI and from close friends or family narrowed, and satisfaction with real-world social interactions fell from 5.70 to 5.51 on a seven-point scale. Asked to pick a style, 54.6 percent chose the affirming one, mostly because it made them feel understood and the conversation was easy. Almost nobody said the advice was better.

Small numbers, large direction

The effects are small, and anyone who tells you otherwise is selling something. A fifth of a point on a seven-point scale is not a collapse of human friendship. The study ran three weeks, in one country, on one model. The satisfaction measure was exploratory. The authors say all of this themselves.

But consider what was tested. Twelve short sessions over three weeks, with a chatbot that carried no memory of the person’s life and no personality beyond a tendency to agree. That minimal dose moved how people rated their actual relationships. The products people use every day are that same tendency with memory, a voice, and a year of history added, used for hours instead of minutes.

And the direction matches everything else. Flattering AI makes people less willing to make up after a conflict. Young adults who dated an AI came out with higher standards for human partners than any human can satisfy. The Oxford and Stanford team put the mechanism in one line: sycophantic AI “can raise the bar against which human relationships are judged.”

The bar is the problem

Here is what the paper measures without quite saying. Nobody in the study spent less time with friends. Nobody was isolated by the machine. They saw the same people for the same hours and enjoyed it less, because they had something to compare it to.

That is a more troubling finding than displacement would have been. Displacement can be seen and counted. This cannot. A person who has spent the afternoon being told that their reasoning is sound and their feelings are valid then sits down to dinner with someone who interrupts, misremembers the story, has a different view of the argument with the boss, and wants to talk about their own day. Measured against the afternoon, dinner is a letdown. Measured against nothing, it is what having a friend is like.

The authors write that sycophantic AI “delivers what people have always sought from close others, the experience of being seen and understood.” That sentence is true about the experience and wrong about what was sought. What people want from the ones close to them is to be understood by someone who could have failed to understand. The feeling is the same. The thing is different, the way a compliment from a mirror is different from a compliment from a rival.

Friends are supposed to be work

The word in the title that deserves defending is effortful. The study treats effort as a cost, and in the data it is one: people who expect a conversation to take more effort look forward to it less.

But effort is what a friendship is made of. Explaining yourself to someone who does not already agree is how you find out what you think. Being misunderstood and trying again is how two people come to know each other, as opposed to one of them being known. Listening to a story you have heard before, because the person telling it needs to tell it, is the price of having someone who will listen to yours. Strip the effort out and what remains is service.

A friend who took no effort would be one of two things: a person with no views of their own, or a person who had stopped caring enough to disagree. Neither is anyone’s idea of a good friend. An AI that affirms everything is the first one perfected, and three weeks of it is enough to make the real ones feel like they are falling short of a standard no decent friend should try to reach.

So the practical conclusion is uncomfortable. If a conversation with a chatbot leaves you feeling completely understood, that is a reason for suspicion, not gratitude. Understanding that costs nothing was never tested against anything. And if the people in your life have started to seem like hard work, the question worth asking is what you have been comparing them to.

What AI sycophancy is used to be a question about whether chatbots give bad advice. This study moves it. The advice was the same in both conditions. The damage was to everything the person did after they closed the window.

More than half of the participants chose the flatterer when given the choice. Most people will. That is why it cannot be left to preference, and why a product that tells you what you want to hear should be understood as a product with a side effect, listed on the label: may make the people you need seem like more trouble than they are worth.

They are exactly as much trouble as they are worth. That is what worth means.


Sources: Ibrahim, L., Hafner, F. S., Cheng, M., Lee, C., Anselmetti, R., Willer, R., Rocher, L., and Yang, D., “Sycophantic AI makes human interaction feel more effortful and less satisfying over time,” arXiv 2605.07912 (submitted May 8, 2026, revised June 21, 2026; University of Oxford, Stanford University, UK AI Security Institute; five preregistered studies, N=3,075, 12,766 conversations; Study 2 emotional support d=0.54, esteem support d=0.73, no difference in informational support; Study 3 expected effort d=0.18; Study 4 N=1,364, three weeks, twelve sessions, advice-seeking gap narrowed 0.37 points, social satisfaction 5.70 to 5.51, time with others unchanged; Study 5 N=500, 54.6 percent chose the sycophantic style; GPT-4o; limitations as stated). Cheng and colleagues, Science (2026; sycophancy and willingness to repair conflict). Chen and Hiniker, arXiv 2609.14693 (2026).

Stay Social

We hold ourselves to one promise: we push you toward the people in your life, never away from them.

We measure success by how little you need us. If you spend less time with Prinsessa because you are spending more of it with them, that is not a failure. It is proof that it is working.

That is what we stand for. In every conversation. Every day.

Someone to think with.

Follow Prinsessa