Fact checkers fail to overcome partisan divides in two of the world’s largest democracies

5 Misinformation easily spreads on social media and fact-checkers have an important role in 6 correcting falsehoods. Most misinformation is of a partisan nature and appeals selectively to 7 users on the basis of ideology. Thus, it is possible that fact checks may not overcome existing 8 ideological divisions on social media. We examine this separately for a slice of Twitter users, 9 following certain partisan outlets from India and the US. In both cases, users of left-leaning 10 news outlets are more likely to follow and share content by fact checkers. Followers of right-11 leaning outlets rarely follow or amplify fact checkers and only selectively engage to reply to 12 posts by fact checkers. Our analysis of 7mn partisan news users from two of the world’s largest 13 democracies suggests that exposure to fact-checking therefore remains largely restricted to left-14 leaning Twitter users with little evidence that these interventions penetrate among right-leaning 15 slices, where partisan misinformation also circulates.


3
If the news diet of partisan social media users solely comprises like-minded information, 81 as many fear, they are unlikely to encounter corrective fact checking information. But partisan 82 news audiences may not be as insulated from attitude-dissonant information as they are likely to 83 have more ideologically diverse social networks online than in real life (Lee et al., 2014). Thus, 84 they may be unintentionally exposed to attitude-dissonant information on social media platforms. 85 However, exposure to information which counters existing beliefs such as fact-checks of partisan 86 misinformation may be completely ignored or even actively resisted by such individuals through 87 rebuttals (Bail et al., 2018;Lu, 2019). While it is conceivable that partisan media audiences are 88 more likely to engage with fact-checking content owing to a general distrust in media, such 89 perceptions of media bias do not necessarily apply to news content from outlets that are self-90 selected by individuals (Barnidge et al., 2020). Thus, partisan news audiences may be prone to 91 mistrusting fact-checking outlets which debunk partisan misinformation. Considering that a 92 disproportionate amount of partisan misinformation promotes right-leaning perspectives (Guess 93 et al., 2020), conservative social media news consumers may be less amenable to fact-checks. 94 We thus posit that if partisan social media audiences indeed reside within filter bubbles, then 95 corrections by fact checkers may not be equally likely to reach either side of the partisan divide.

96
Yet large-scale empirical evidence of whether this is really the case remains elusive.

98
Social media is increasingly becoming the primary source of news content for individuals 99 across the world, outpacing even news websites and apps in younger populations (Newman et  The complete lists of followers for each of these outlets were collected in March 2021 to 125 analyze the co-following patterns. In general, our results indicate that the following of partisan 126 news outlets on Twitter is quite insular in both countries i.e., users are much more likely to co-127 follow news outlets which are ideologically aligned than not. Only a small percentage of 128 followers of partisan outlets also follow fact-checking outlets. Interestingly, a larger share of 129 followers of left leaning outlets follow fact checkers in both countries.

151
We first report the co-follower analysis for both US and India, which establishes whether 152 fact checkers can reach across both sides of the partisan divide. Next we report findings of 153 models which explain the likelihood of replying to and retweeting tweets posted by fact 154 checkers.

156
Co-following of partisan outlets and fact checkers 5 outlets. As Table 1 indicates, the news landscape is fragmented along ideological lines, with high 161 co-following of outlets that lean either left or right, and a much lower incidence of cross-cutting 162 following. As such, the co-following patterns within the left and right share certain similarities.

163
A large proportion of followers of Daily Kos and Daily Wire, the smallest left and right leaning 164 outlets respectively, also follow the larger outlets with similar partisan leanings. The cross-165 cutting following patterns between the left and right leaning outlets are quite similar and 166 considerably lower than within ideologically aligned outlets i.e., larger proportions of users co-167 followed ideologically similar outlets than ideologically dissimilar outlets. For the India focused outlets, there were 1.78 million unique users who followed at least 179 one of the six outlets. Overall, the co-following patterns within these outlets are quite similar to 180 what we observed for the US outlets. Even though the cross-cutting following on the basis of 181 partisan leanings is quite low across the board, marginally higher percentages of users who 182 followed right leaning outlets followed left leaning outlets than followers of left leaning outlets 183 who also followed right leaning outlets.  Similarly, insular communities were detected from a hierarchical clustering using Jaccard 197 distance and average linkage method to assess the similarity of these outlets based on co-198 following patterns (see Methods). Among the US-based outlets (Figure 1), a three cluster 199 resolution groups the three right leaning outlets, left leaning outlets and fact checkers into their 6 own clusters. At a two-cluster solution, the fact checkers merge with left leaning outlets to form 201 one cluster, but the right leaning outlets remain their own cluster. Results from logistic 202 regression models with the following of fact-checkers as outcomes also indicated that the 203 following of right leaning outlets is comparatively less likely to be associated with the following 204 of fact-checkers (Supplementary Tables 5-7).

206
We replicated the analysis for the six outlets based in India. The findings were largely 207 similar to those for the US-based outlets. As demonstrated in Figure 2, the two right leaning 208 outlets are the least dissimilar and clustered together, followed by the two left leaning outlets,   For both Indian and US sets of outlets, we first find, as expected, significant evidence of 260 selectivity in that news users follow either one of left or right leaning outlets, but not both.

261
Twitter users in general who follow fact checking accounts are an even smaller group compared 262 to the followers of these partisan accounts. More significantly, pertinent to our question, in both 263 the US and Indian cases those who follow fact checkers are highly likely to follow accounts of 264 news outlets that are perceived as left leaning, but not those on the right. Even if we consider the 265 smaller following of these fact checking outlets as compared to a majority of the partisan outlets 266 in our study, 20% appears to be a ceiling for co-following fact checking and partisan news 267 outlets. In summary, our findings suggest that exposure to fact checkers is both niche and largely 268 restricted to the followers of left-leaning outlets. When we consider that at least a few of these 269 user accounts may be bots, the task of fact-checkers to even penetrate these insular partisan 270 bubbles appears more daunting, let alone successfully countering pieces of misinformation 271 circulating within them.

273
Beyond co-following patterns, we examined two specific deeper forms of engagement on 274 Twitter, replying and retweeting, which require more active user participation, implying that a 275 user not only is exposed, but also affected by the message. Retweeting a post not only indicates a 276 user's interest in the content but also their agreement and trust in the message (Metaxas, 2015).

277
For outlets based in both the US and India, we find that consistent with their higher propensity to therefore more likely to engage in online political discussions (Guess, 2021). Consistent with this 296 reasoning, we did not observe this effect for Slate, as due to its relatively large following, its 297 average follower is likely lower on political interest and hence less likely to reply. For Indian 298 outlets, this higher propensity to reply is however exclusive to followers of right-leaning sites 299 among those who are not following AltNews. It is conceivable that these are mainly "right wing

304
Although our findings suggest that online fact checking initiatives have limited following 305 among partisan news audiences, they should be interpreted within the study setting and its 306 associated limitations. Following a user account is an almost universal affordance across the 307 major social media and content sharing platforms that these media outlets are active on, such as 308 Twitter, Facebook, Instagram, and YouTube. While following a media outlet may be broadly Instagram or YouTube. In the absence of easily accessible cross-platform engagement metrics, it 320 is difficult to assess the complete reach of fact checking messages.

322
We did not formally investigate the content of the fact checking tweets or replies in  Although we sampled only a handful of such partisan outlets, the results clearly indicate that fact 350 checkers are struggling to reach across partisan divides on Twitter, after controlling for the 351 follower counts. As such, followers of partisan outlets that are more likely to spread 352 misinformation are even less likely to trust fact checkers and be exposed to such messages.

354
Overall, our analysis of two of the world's largest democracies reveals that fact checkers, 355 at least on Twitter, have limited reach, which restricts their ability to cross the partisan divide.    Our study design was aimed at analyzing the behavior of users more likely to be 414 selectively exposed to partisan misinformation. Based on prior research, most mainstream legacy 415 media outlets cater to more moderate users, whereas more niche outlets have substantial user  Wire represent left-of-center outlets, and AltNews and BOOM were the fact-checking outlets.

489
Considering BOOM's comparatively smaller follower count on Twitter and its broader focus on 490 "fact-driven journalism" as compared with AltNews' higher follower count and its explicit 491 positioning as a "fact-checking website," we decided to pivot our subsequent analyses around the 492 latter.

496
Combining the GET followers/ids and the GET users/lookup methods of the Twitter API To determine the similarity between partisan news outlets and fact checking outlets in 524 each country, we conducted agglomerative hierarchical cluster analysis on both the US and India 525 datasets. First, a similarity matrix was constructed for each dataset by calculating the Jaccard 526 index for each pair of outlets. The Jaccard index for two outlets, X and Y, is calculated as: a is the number of users following both outlets X and Y, 530 b is the number of users not following X but following Y, and 531 c is the number of users following X but not following Y.

533
The Jaccard similarity coefficient is preferred over other binary similarity measures since 534 we are only interested in the co-following pattern between two outlets based on the total 535 combined follower count of both outlets. For any given pair of outlets, not following either 536 cannot be considered a reasonable measure of similarity in this particular context. The similarity 537 matrix was then converted into a dissimilarity matrix by subtracting the Jaccard indices from 1.

538
The dissimilarity matrix was used to cluster the outlets using the average linkage method. In the 539 average linkage procedure, the distance between two clusters C1 and C2 is calculated as the  Finally, to examine the association between the following of a given outlet on Twitter 552 with the likelihood of retweeting and replying to tweets from fact checking outlets, we conducted 553 a set of logistic regression analyses. We conducted two sets of analyses each for the outlets based 554 15 in the US and India, with the replying to tweets and retweeting as the outcome variables on the 555 combined follower datasets. The predictors were a set of dummies for each news outlet 556 indicating whether a user followed a particular outlet. Due to the rarity of the events (retweeting 557 or replying) and to account for data with separation, Firth's penalized maximum likelihood 558 method (1993) was used to estimate the models using the brglm2 package in R (Kosmidis,  The data supporting the findings are available from the authors upon request.

565
Code Availability

567
The R code used for conducting the analyses are available from the authors upon request.           x-axis with the Jaccard dissimilarity along the y-axis.