In their “For You” pages, social media algorithms purport to find you exactly the kinds of content you want to see and gather it together in one easy-to-peruse section. That is, if you click on and watch enough political videos, you should expect lots of politics in your algorithmically curated feed. But a new study finds a curious phenomenon: The smarter the algorithm gets at reading your behavior, the further it drifts from what you actually care about.

In a new paper, Stanford scholars studied the For You feed of X (formerly Twitter) users to find that there is a fundamental value misalignment of content with the user’s stated personal values.

It all has to do with how the algorithm ranks and amplifies content based on a metric known as engagement. The algorithm measures and heavily weights how likely you are to engage with content – following creators and time spent reading, reposting, and especially replying to the content in your feed. It then uses that measurement as a proxy for your interests and serves you more of the same.

Rapt but wrong

“Engagement is easy to measure, so the platforms use it as the primary metric to assess user interest in the content,” says the paper’s first author, Ziv Epstein, who was a postdoc in the lab of senior author Michael Bernstein, a Stanford professor of computer science and senior fellow at the Stanford Institute for Human-Centered AI, when he did the work. Epstein is now a SERC postdoctoral associate at the Schwarzman College of Computing at MIT. “They are saying, in effect, engagement equals values, but it’s not that simple,” he added.

The study, published in Proceedings of the National Academy of Sciences (PNAS) and partially supported by the Stanford HAI, finds users are more likely to engage with content they disagree with than with content they find more palatable. That means the emphasis on engagement produces feeds steered toward outrage-inducing content. While it is effective at keeping users rapt, it does little to amplify what they actually care about. The algorithm goes astray and the For You page quickly grows misaligned with the user’s values.

“The message we take away is that the values that we hold – what we really care about – should be just as informative to the platforms as engagement,” Bernstein explains. “Right now, we engage with all sorts of content – sometimes because we agree with it, and sometimes because we’re angry at it – but the platforms misinterpret it all as a kind of desire for more of that kind of content.”

Elegant insights

To elucidate their hypothesis, the researchers designed a simple study. First, they randomly selected 715 X users and asked them to complete a survey. They plotted each user’s answers against a long-accepted scale known as Schwartz’s Hierarchy of Basic Human Values. Schwartz’s list includes 19 values ranging from humility to hedonism.

The researchers then examined two distinct feeds in each user’s X account. The first was the “Following” feed, which is composed of content from accounts the users have voluntarily tracked. The second stream was the “For You” page that serves up algorithmically selected content. The team then evaluated the content in each feed against the Schwartz scale and compared those scores against the user’s value profile.

They found that posts misaligned with the user’s values are more likely to be amplified into the For You page feed than posts aligned with their values.

Future research and remedies

The PNAS article, Epstein says, is not a finger-pointing exercise but an encouragement for dialogue on new research and remedies. The findings highlight a natural tension within the users themselves: We seek content that aligns with our values but react to content that provokes or offends us.

As for remedies, the team has produced a Google Chrome browser extension that reviews and re-ranks feeds based on self-stated values. Epstein says it is not a cure-all and could potentially lead to feeds that do little more than echo personal tastes. Other possibilities, the scholars say, might include pro-social algorithms that surface content based on shared values that bridge philosophical divides or tools that give users more control over their feeds.

“Our feeds shouldn’t value confrontational engagement over complementary content, nor should they obsequiously feed us only content we agree with. They should strike a balance and mend social disagreements rather than make them worse,” Bernstein concludes. “This work shows we can measure the space between what people value and what the algorithms are amplifying and, in doing that, we can use science to narrow the gaps that divide us.”