How Much Has Social Media Affected Polarization?
By Tom Cunningham, Integrity Institute Founding Fellow
Tom Cunningham worked as economist and data scientist at Facebook and Twitter, working on content moderation and company strategy. His work has been extensively quoted in many publications, in the House report on Competition in Digital Markets, and the House report on January 6. Since resigning from Twitter in November 2022 he has been writing about content moderation.
Tom is a Founding Fellow of the Integrity Institute and, since June 2023, a Resident Fellow as well. As Resident Fellow, Tom will conduct various original research projects on social platforms and social media.
This research is conducted by Tom as Resident Fellow as an independent evaluation of Meta’s 2020 Facebook and Instagram Election Study (FIES). We publish the summary below, while the full research note can be accessed here.
TL;DR: The experiments run by Meta during the 2020 elections were not big enough to test the theory that social media has made a substantial contribution to polarization in the US. Nevertheless there are other reasons to doubt it.
Summary
Three new experiments show that changing Facebook’s feed ranking algorithm for 1.5 months has an effect on affective polarization of less than 0.03 standard deviations. This is small compared to a growth of 1.1 standard deviations in nationwide affective polarization over the last 40 years.
Small effects in these experiments are consistent with large effects in aggregate. The aggregate contribution of social media to polarization will differ from these experimental estimates in a number of ways: depth, breadth, duration, timing, category, and population. My rough attempts to account for these considerations make me think the aggregate effect is likely 10 or 20 times larger than the effects that would be measured in these experiments, and so small effects in these experiments are consistent with large effects on aggregate.
Put simply: these experiments measure the effect of reducing exposure of an individual user (not their friends and family) to political content on Facebook by 15% for 1.5 months, and occurred in a period after Facebook had already sharply reduced the amount of partisan content circulating. Thus we should expect them to measure only a small fraction of the cumulative impact of social media, and in fact these results are consistent with social media being entirely responsible for the growth of polarization in the US.
Nevertheless other evidence implies that social media has probably not made a huge contribution to US polarization. If we wish to evaluate the balance of evidence relating social media to polarization there are many other sources which are probably more informative than these experiments. I give a rough sketch below and it seems to me social media probably does not account for a majority share, mainly because (1) polarization had been growing for 20 years prior to social media’s introduction, and much of the growth since 2014 was in people without internet access; (2) a lot of partisan discourse continues to spread outside of social media, e.g. through cable TV and talk radio; (3) other countries do not show a similar increase in affective polarization.
Discussion of these results has been distressingly non-quantitative. The majority of discussion of these results (in papers, editorials, on Twitter) has been about whether these changes “have an effect” or “do not have an effect.” Interpreted sympathetically these statements are compressed ways of saying “an effect larger than 0.03 standard deviations.” However I think taking this shortcut so consistently has led to far too little time thinking about what we have learned from these experiments that we didn’t already know, and what is the balance of evidence regarding the effects of social media. I give a lot of examples below.
[For the full research note, please see here.]