Engagement-based algorithms disrupt human social norm norming

As we increasingly learn from others in online spaces regulated by engagement-based feed-ranking algorithms, it is important to understand the impact that these algorithms have on our social learning. In three pre-registered studies (N = 6107), we isolated the effects of feed-ranking algorithms on social norm learning by training algorithms on passive and active engagement in a simulated social media environment. We found that engagement-based algorithms systematically amplified ingroup-aligned, moral and emotional (IME) political content, leading IME content to become overrepresented in feeds. The overrepresentation of IME content in feeds caused participants to overperceive norms of posting IME content, which mediated user intentions to post more IME content. A bridging algorithm successfully reduced IME content in feeds, but also led to underperception of some IME content, suggesting that bridging-based algorithms do not straightforwardly promote more accurate social learning. Our findings shed light on algorithm-mediated social learning in the digital age, demonstrating that specific human learning biases toward IME content are amplified by engagement-based algorithms in ways that disrupt social norm perception. Our findings highlight a central challenge for engagement-based algorithms: how to promote content in ways that enable users to accurately infer others’ preferences.