The proliferation of untrustworthy information, particularly during election periods, poses a significant challenge to democratic processes worldwide. Combining a two-wave panel study with web-browsing behavior (N = 21M) of 1,194 American adults around the 2020 US presidential election, we examine associations between visits to untrustworthy websites at the domain level and subsequent election beliefs. We find that those exposed to untrustworthy websites are 17.3% more likely to believe that the election was not won by the certified winner in a post-election survey. However, one challenge in assessing the relationship between untrustworthy website exposure and beliefs is disentangling the behavior of individuals selectively exposing themselves to information congenial to an existing belief versus prior partisanship heterogeneously relating to differing changes in beliefs. To address this challenge, we apply a machine learning method that controls for the propensity to be exposed to untrustworthy websites. Using this method, we find that exposure to untrustworthy websites is associated with the likelihood of believing that the certified election winner did not win by 4.2%. This association, however, is asymmetric: For conservatives, whose candidate lost the election, the association is 12.6%, while for liberals, the association is negligible at -0.2%. Further, we identify a dosage relationship—each additional exposure to untrustworthy websites increases the association between untrustworthy website exposure and election outcome beliefs among conservatives by .035% but only .010% for liberals. Through a series of stringent robustness checks, we find that these directional associations are unlikely to be “explained away” by potential unobserved variables, adding credibility to the estimates. Our results underscore the importance of addressing untrustworthy information not as a uniform influence but as a factor whose relationship with beliefs is contingent upon environmental context, individual predispositions, and exposure volume.
Dahlke et al. (Fri,) studied this question.