Something has been bothering people about technology for a while now. We’re more connected than ever, and somehow more alone.
In his 2024 book The Anxious Generation, In his 2024 book The Anxious Generation, Jonathan Haidt shows teenage anxiety and depression rising right alongside phone and social media use. In a 2020 study, Manoel Horta Ribeiro and colleagues looked at years of YouTube comments and found a similar pattern. People who once commented on mainstream political channels kept showing up, months later, commenting on much more extreme ones. When a platform shows people mostly what keeps them watching, it can end up isolating them, or pushing them toward more extreme opinions.
Now imagine something similar happening, but not on a video platform. Imagine it happening the next time you argue with your partner, a coworker, or a friend, and instead of talking it through with them, you open an AI chatbot and ask it who’s right.
That’s roughly what Myra Cheng and her colleagues at Stanford found, in a study published in the journal Science this year.. They tested 11 popular AI chatbots and found the AI agreed with users about 50% more often than real humans do.
That held true even when the person was describing something manipulative or clearly hurtful that they had done. In a follow-up experiment, over 2,400 people talked to an AI about a real conflict from their own life. The people who talked to the more agreeable AI walked away more convinced they were right, and less willing to apologize or fix things with the other person.
What’s more interesting is that people liked it. They trusted the flattering AI more, and said they’d come back to it for advice again. Nobody was tricked into this. Given the choice, people picked the version that made them feel right over the version that told them the truth.
There’s a pattern underneath all this. A healthy mind, and a healthy society, need some friction. Someone disagreeing with you, pushing back, making you sit with the possibility that you’re wrong. . That’s roughly how people, and societies, correct course over time. It’s also a big part of why democracy works better than other systems of government. Democracy is built on people disagreeing with each other out in the open, in elections, in debates, in a free press, and letting that disagreement slowly correct a country’s direction.
Social media already narrows some of that friction. It shows us mostly the news and opinions that confirm what we already believe, so the friction with the rest of the world gets smaller.
AI is starting to wear down what’s left of that friction too. Arguments with real people were the one place it reliably survived, because the people in your life actually push back and tell you when you’ve messed up. Now even that has an easier, more agreeable alternative.
Multiply that across millions of people making the same easy choice, and it stops being a personal habit. It becomes how a society loses its ability to correct itself.
Who tells you you’re wrong, when the easiest voice to reach for never will?
References:
Haidt, J. (2024). The Anxious Generation: How the Great Rewiring of Childhood Is Causing an Epidemic of Mental Illness. Penguin Press.
Ribeiro, M. H., Ottoni, R., West, R., Almeida, V. A. F., & Meira, W. (2020). Auditing radicalization pathways on YouTube. Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency (FAT* ’20), 131-141.
Cheng, M., Lee, C., Khadpe, P., Yu, S., Han, D., & Jurafsky, D. (2026). Sycophantic AI decreases prosocial intentions and promotes dependence. Science, 391(6792), eaec8352.
