Embedding Societal Values into Social Media Algorithms

Social media influences what we see and hear, what we believe, and how we act-but artificial intelligence (AI) influences social media.By changing our social environments, AIs change our social behavior: as per Winston Churchill, "We shape our buildings; thereafter, they shape us."Across billions of people on platforms from Facebook to Twitter to YouTube to TikTok, AI decides what is at the top of our feeds (Backstrom 2016; Fischer 2020), who we might connect with (Guy, Ronen, and Wilcox 2009), and what should be moderated, labeled with a warning, or outright removed (Gillespie 2018).These AI models change the social environment around us by amplifying or removing misinformation and radicalizing content (Hassan et al. 2015), by highlighting or suppressing antisocial behavior such as harassment (Lees et al. 2022), and by upranking or downranking content that might harm well-being (Burke, Cheng, and Gant 2020).How do we understand and engineer this sociotechnical ouroboros (Mansoury et al. 2020)?As the traditional critique goes, these challenges arise because social media AIs are optimized for engagement (Backstrom 2016; Narayanan 2023).But this is not the full story: to help manage undesirable outcomes of engagement-based algorithms, platforms have long augmented their algorithms 1 with nonengagement outcomes (Eckles 2021).For instance, to help defeat clickbait, platforms such as Facebook began surveying users for their opinions on specific posts, and then building models that could predict and downrank posts that people dislike, even if they are likely to click on them (Backstrom and Mosseri 2015).To ensure that all users receive feedback, platforms designed algorithms weighing the effect of user feedback on other users who might otherwise get few replies (Eckles, Kizilcec, and Bakshy 2016).To diminish the prevalence of content that violates community standards, such as misinformation and gore, platforms built algorithms and paid moderation teams to flag and remove this content.This battery of surveys, moderation, downranking, peer effect estimation, and other models are all now components of many platforms (Eckles 2021).1.In this commentary, we refer to "AI" and "algorithm" interchangeably to refer to machine learning procedures that learn to predict from large-scale data.We are primarily concerned with social media algorithms focused on ranking and recommendation, especially feed algorithms, but we note that social media AIs play many other roles as well, including content moderation, (de)monetization, misinformation tagging, political content tagging, and toxicity judgments.