Can you reduce polarization on social media without losing engagement? The answer is yes, a little, sometimes, on some platforms.
Stray and 42 co-authors (arXiv:2603.19626) modified content feeds for 9,386 users across Facebook, Reddit, and X/Twitter for six months during the 2024 US election. Alternative algorithms — designed to surface bridging content, cross-partisan perspectives, and high-quality sources — were tested against the platforms' default ranking.
The polarization reduction was real but small: 0.03 standard deviations in affective polarization, a 1.5-degree decrease on 100-point feeling thermometers measuring attitudes toward the opposing party. The engagement effects split across platforms: Facebook and Reddit users spent slightly less time (0.37 and 0.2 fewer minutes per day), while X/Twitter users spent slightly more (0.32 additional minutes). No significant effects on well-being, news knowledge, outgroup empathy, or attitudes toward partisan violence.
The modesty of the results is the result. This is the largest comparative test of alternative social media algorithms to date — 43 researchers, three platforms, six months, a politically charged period — and the effect sizes are barely detectable. Bridging content can improve some societal outcomes without necessarily conflicting with engagement-driven business models. But it can't transform them.
The finding constrains an entire class of policy proposals. The idea that feed algorithm changes could substantially reduce polarization assumed that the algorithm was a primary cause. These results suggest the algorithm is a weak lever. Polarization lives in the people, the content, and the broader information environment — not primarily in the ranking. Changing the order in which content appears adjusts exposure at the margin. It doesn't change what people want to see, what they share, or what makes them angry.
The most useful reading: the relationship between engagement and polarization is not strongly competitive. Platforms don't need to choose between engagement and prosocial outcomes. But the prosocial gains from algorithmic adjustment are small enough that they may not be worth the engineering cost of implementation.