Social media algorithms are harmless, or are they?

This AlgorithmWatch article critically reviews a set of Meta-partnered academic studies on the 2020 US election that compared users shown algorithmically ranked feeds against those shown reverse-chronological feeds. The studies relied mainly on short-period, self-reported surveys alongside behavioral data supplied by Meta, using opt-in volunteers rather than a fully randomized population.

The article highlights a contradiction between the two data sources: survey responses found no significant shifts in polarization, views on immigration, COVID-19, healthcare, election knowledge, or behaviors like protest attendance and campaign donations, while the actual behavioral data showed that algorithmic feeds increased exposure to uncivil content, reduced political content from moderate sources, and substantially increased time spent on the platform; removing the algorithm cut exposure to like-minded sources at twice the rate of cross-cutting sources.

The piece concludes that the studies' design and reliance on short-term self-reported measures may have obscured real algorithmic effects on behavior, and suggests that more comprehensive, independent behavioral analysis could reach different conclusions about the platforms' impact on democratic discourse.