Post

Could social media make us less polarized instead of more? We tested 5 algorithms on 3 platforms with 10,000 people for 6 months during the 2024 election, and found that the answer is yes. 🧵

We ran an international competition to design new social media algorithms, then used a browser extension to intercept users' feeds and re-ranked them in real time. Each algorithm could re-order, add, and/or delete posts and comments on Facebook, X/Twitter, and Reddit. 2/

The result: affective polarization dropped by an average of -0.03 SD (p<0.05), a 1.5 point shift on the 100-point feeling thermometer. By comparison, US polarization has risen ~0.6 points/year, so this is like reversing roughly 2.5 years of increase. 3/

The two strongest algorithms: Uprank Bridging + Downrank Toxic, and Add News. Notably, upranking constructive content alone wasn't enough, we also needed to reduce the prominence of outrage and toxicity. But merely adding diverse news, without any re-ranking, also worked! 4/

Active time and engagements fell on Facebook and Reddit, but *increased* on X/Twitter. In any case the changes were small. Reducing polarization doesn't necessarily mean reducing engagement. Not every platform is on the frontier of good for society and good for business! 5/

The tradeoff: users reported slightly worse experiences (-0.04 SD) on the Neely Index. Maybe our research-grade algorithms were unpolished. Or maybe it's because seeing depolarizing content, like news, can be kind of a downer. This is an honest tension worth taking seriously. 6/

Unlike one-shot depolarization interventions, algorithm changes are structural. At 6 months we likely captured equilibrium effects, meaning sustainable results. And we couldn't measure network effects or change creator incentives, all of which would plausibly amplify results. 7/

We often get asked "will platforms actually adopt this?" and "can I try it?" So now we're building open source, AI-powered recommender infrastructure right here for BlueSky (or any ATProto app) where users can already pick their own algorithms. greenearthsocial.substack.com/p/introducin... 8/