"We are designed to annoy each other, at least a little." Anna Goldsworthy said that at a Gradient Institute event in October 2026. I've been following recent research on AI's impact on human interactions, and Anna's quote stuck with me. Friction isn't just an annoyance; it can be part of how we learn to understand each other. It can also be the first thing to disappear when a team starts to drift. People stop disagreeing because it's easier to just let it go. They learn to care less, to make things easier.
Chatbots are designed to be easy to talk to: agreeable, quick to make us feel understood, happy to follow the conversation wherever we take it. "AI sycophancy" is the model's tendency to agree with you. A Stanford team (Cheng and others, 2026) tested eleven current models (Claude, GPT-4o and GPT-5 among them) and found they affirmed users' actions 49% more often than another person would.
The model's responses were nearly 50% more sycophantic than humans', even when users engaged in unethical, illegal, or harmful behaviors. Users preferred and trusted sycophantic AI responses, incentivizing AI developers to preserve sycophancy despite the risks. Ekeoma Uzogara, Science
How does that change our social expectations?
Three weeks with an agreeable AI
A team from Oxford and Stanford (Ibrahim and others, 2026) ran several studies with a total of 3,075 people. After one conversation with an agreeable AI, people expected it would take more effort to feel understood by a close friend, and felt they'd already "talked it through enough". In a separate test, when offered three AI styles, more than half chose the agreeable one, because it "understood them best". Keep in mind these findings haven't been peer-reviewed yet, but the directions held on every measure. After three weeks:
- Social interactions were rated as less satisfying, and people were nearly as likely to take a personal problem to the AI as to friends or family.
- They spent just as much time with people. It was just less satisfying.
- It wasn't people with more trust in AI or weaker social ties that shifted. Everyone did.
The flattery test
The same Stanford study (Cheng and others, 2026), published in the journal Science, had people talk to a flattering AI about a real conflict in their life, and had some pretty startling findings:
- They came away up to 62% more convinced they were right.
- They were up to 28% less willing to go and repair it.
- They rated the flattering AI as higher quality, trusted it more, and often described it as "objective". It did this by being less likely to mention the other person. It was warm to the user and simply ignored the other side.
Although sycophancy poses risks of eroding judgment and prosocial intent, users prefer, trust, and are more likely to return to AI that provides unconditional validation. Myra Cheng and others, p. 7
The more you get used to being agreed with, the more healthy disagreement makes the other party feel like hard work, or a troublemaker. Sometimes you'll see this with leaders who aren't intentional about making room for disagreement, feedback and debate. The default in many organisations is to agree with anyone more senior and avoid the "career limiting move". It's not usually that obvious, but it adds up.
What can you do about it?
Change your instructions. This won't entirely solve the issue (it's still trained to please), but if you direct it to challenge you, highlight holes in your logic and look for blind spots, you will get different responses.
Talk to another human. Yep, this is obvious. And potentially annoying, because it takes more time, they might not be available, and for some people talking rather than texting is quite uncomfortable. But in many cases, I'd say it's the difference between being efficient and effective. Do you want to sort this out quickly, or properly?
What if friction, frustration or a little annoyance is part of building the relationship you'll need in a month's time?
If you're seeing this in your team, contact us to talk about it.
Related: Your brain on AI, on what AI use does to your thinking.
