Why social media algorithms send you posts you don’t like

Ziv Epstein, MIT*

Do your social media accounts feed you content reflecting your beliefs and principles? Our research published in the Proceedings of the National Academy of Sciences shows that algorithms may prioritize content that clashes with your values. Because algorithms heavily weigh posts you reply to, users tend to comment more on content they oppose than content they agree with.

Our study of X shows that although its algorithm promotes content contradicting the values of Democratic and Republican users, it does so more for Democrats.

Social media platforms use algorithms to select posts from a pool of content. On X, posts appear on “For You” pages. Algorithms predict whether you will engage with content by liking or commenting on it, then use those predictions to tailor future content.

But do those posts reflect what you value? Some users care about preserving traditions or keeping society safe. Others prioritize free expression or protecting nature. Most care about all those things to varying degrees. A feed aligned with someone’s values would reflect those priorities.

Psychologists use established surveys to measure values. To measure values expressed in posts selected by a platform, we built a tool using psychological classifications of human values and applied it to the feeds of 715 U.S.-based X users.

We found that X’s algorithm was most likely to amplify posts about upholding tradition, following rules or keeping society safe. It was most likely to demote posts about caring for people, concern for people far away, dependability or protecting nature. Compared with values users expressed in their own posts, the algorithm was more likely to promote conflicting content than aligned content.

Why? We first checked whether users followed accounts diverging from their values, but most accounts they followed were aligned. We also examined whether people engaged only with posts they opposed, but they engaged with plenty reflecting values they shared.

The catch lies in the interactions. People mainly respond by liking content. Less often, they write replies, and when they do, they frequently respond to posts clashing with their values.

Here is the smoking gun: X treats those rare replies as a much weightier signal than numerous likes. It therefore learns most strongly from replies and serves users posts reflecting the values of content they commented on – often values conflicting with their own.

This tendency was stronger among self-identified Democrats than Republicans. Our evidence indicates Democrats object more than Republicans to content they reply to, creating a stronger feedback loop of clashing posts. The content amplified was more than four times more misaligned for Democrats than Republicans.

What comes next? In other research, we developed a way for platforms to ask users what they value and sort feeds accordingly. Users proved quite good at identifying whether sample feeds aligned with their values.

Value-aligned feeds might offer a path out of echo chambers that unmediated exposure to opposing views does not. Our research suggests engagement-optimized algorithms – essentially handing people the opposition and leaving them to sort it out – may increase polarization. Surfacing content that bridges political lines while addressing users’ values could foster autonomy and constructive conversation.

People should have a greater say in the information platforms show them. If designers, the public and policymakers create tools serving that goal, platforms may better support the values and actions people care about. [Abridged]

*with Farnaz Jahanbakhsh, University of Michigan,
and Michael Bernstein, Stanford University