How do social media users and journalists express concerns about social media misinformation? A computational analysis

This article describes partisan-based, accuracy-based, and action-based discussions through which U


Essay summary
• Supervised machine learning is used to classify texts from U.S. mainstream broadcast news, Twitter (now X), and Facebook posts that discuss the topic of social media misinformation.• Platform policy and party politics are the dominant types of discussions about social media misinformation among both journalists and social media users.• Discussions about platform policy and party politics are the most likely to co-occur together among all discussions, suggesting that the very idea of platform intervention on misinformation is largely politicized.
• Discussions about user agency are more prevalent on Twitter and Facebook than in the news, frequently co-occurring with discussions about direct misinformation correction.• Findings 1) call for journalists to shift away from politicized frames around platform policies aimed at mitigating misinformation, focusing on accuracy instead and 2) provide hope for platforms and educators to empower social media users to express and enact agency in addressing misinformation.

Implications
While social media misinformation constitutes a small percentage of an average American's information diet in volume (Allen et al., 2020;González-Bailón et al., 2023;Guess et al., 2020), concerns and discussions about social media platforms' power in fueling the spread of misinformation have continued to grow.Social media users frequently express concerns about misinformation spreading on platforms, often through identity-based languages (Li & Su, 2020).On the other hand, misinformation has become a prevalent theme in mainstream news coverage.Mainstream media not only are the gatekeeper and corrector of misinformation (Lwin et al., 2023), but also have routinized the coverage about misinformation, with many creating a "misinformation beat" (McClure Haughey et al., 2020).This article describes five types of discussions through which U.S. social media users and journalists engage with the topic of social media misinformation: 1) party politics, 2) quality of knowledge and decisions, 3) direct corrections, 4) user agency, and 5) platform policy.These types of discussions are identified both deductively, by explicating the two dimensions of the theory of motivated reasoning in the context of discussions about social media misinformation, and inductively, with researchers and coders iteratively working with the sample data, engaging with existing research, and updating the coding scheme.They are identified based on their different foci, claims, and explanations about the importance of social media misinformation.In other words, we examine what the driving force is behind people's expressed concern about social media misinformation.
We use the theory of motivated reasoning as a tool to map different lenses people adopt to express concerns about misinformation.While the theory of motivated reasoning is often applied to study how people process a specific piece of information, we argue that it is also useful for studying how people diagnose problems they see in an information environment overall.The theory of motivated reasoning outlines two basic types of motivations underlying all human reasoning (Taber & Lodge, 2006).The first is partisan motivation-the desire to defend preexisting partisan stance.In the context of talking about social media misinformation, being driven by partisan motivation means that people are concerned about social media misinformation because it benefits or harms a certain political party.We call this the party politics discussion, which is used to express concerns over social media misinformation through a lens of party conflict and competition.In this type of discussion, misinformation is seen as an arena where political strategies play out and a place that reflects political (dis)advantages.
In contrast, the second type of motivation outlined by the theory of motivated reasoning is the accuracy motivation-the desire to make accurate, optimal judgments.In the context of discussing the problem of social media misinformation, being motivated by accuracy means that people are concerned about misinformation primarily because it negatively affects the ability to hold accurate beliefs and make best decisions, for example, leading people to form inaccurate views about public affairs or make health decisions that result in illness for themselves or others.We call this the quality of knowledge and decisions discussion.The central concern in this discussion is not political gains or losses, but the verifiable truth itself, and the negative consequences that might flow from not believing the verifiable truth.Besides the quality of knowledge and decisions discussion that focuses on the negative impacts on accuracy, the direct corrections discussion shows accuracy on the spot.In the direct corrections discussion, accuracy is demonstrated through directly debunking misinformation, offering explanations on why a piece of misinformation is false, and transparently providing relevant facts.
While partisan-and accuracy-based discussions are direct applications of motivated reasoning, we also observed frequent discussions on solutions to misinformation from our data and from existing research on misinformation intervention (for a review, see Aghajari et al., 2023) that inherently cut across the binary of accuracy-or partisan-motivated reasoning.These solutions most frequently include calls for action from individual social media users and from social media platforms (Barthel et al., 2016;de Cock Buning, 2018).As such, we describe the user agency discussion that focuses on actions by ordinary social media users to address misinformation.The user agency discussion can express both accuracy and partisan concerns about misinformation.While the call for user agency can advocate for accuracy goals by offering tips on discerning fabricated headlines or encouraging people to read fact checks (Vraga et al., 2020), it can be also articulated to justify selectively seeking favorable information and turning away from uncongenial sources (Nelson & Lewis, 2023).
Finally, we describe the platform policy discussion that focuses on actions to counter misinformation by social media platforms.Platform policies such as exposing people to fact checks (Bode & Vraga, 2015) and banning elite accounts that repeatedly spread misinformation (Dwoskin & Timberg, 2021) may improve belief accuracy, but support for such policies is divided along partisan lines (Saltz et al., 2021).Platform actions aiming at reducing the harm of misinformation quickly become a fertile ground for debates over political bias and censorship.Taken together, action-based discussions, including user agency and platform policy, can intersect with and be fueled by both accuracy-and partisan-based discussions.
A worrying finding from our study is that platform policy and party politics are not only the most frequent types of discussions about misinformation but also most likely to co-occur together among all discussions about misinformation by both journalists and social media users.Tromble and McGregor (2019) argue that scholars should both weigh in on practical solutions for platforms to curb misinformation and provide a critical view on the very nature of platform policies and answer normative questions on social good and harm.This article offers a case for their argument.While empirical evidence shows the effectiveness of specific solutions that can be implemented on platforms, the very idea of platform intervention has been largely politicized, even in mainstream news sources widely thought as following norms of objectivity and nonpartisan reporting.This is consequential: News coverage on platform policy that centers partisan conflict and claims of bias can feed into the strategic frame, turning truth and falsehood into a matter of winning and losing.Such strategic frame in news coverage can decrease knowledge, increase cynicism, and demobilize the public (Zoizner, 2021).When party politics becomes the primary lens to talk about platform policy, it is more likely that political groups will talk past each other in debating partisan bias as compared to engaging in meaningful collective problem-solving, allowing technology companies to continue the current business model and keep the power in deciding how misinformation gets addressed.Moreover, party politics-style coverage can obscure attention to what is verifiably true and what is not as it tends to index competing partisan claims about issues (Bennett, 1990), failing to adjudicate between claims that are verifiably true and those that are verifiably false.
On the other hand, we also find that although less frequent than platform policy and party politics, accuracy-based discussions that highlight quality of knowledge and decisions or offer direct misinformation corrections still occupy a visible part of how both journalists and social media users talk about misinformation.Moreover, discussion about user agency is more frequent on social media than in news, frequently co-occurring with direct misinformation corrections.Contributing to the emerging research that demonstrates the potential for accuracy motivations to help combat misinformation (Pennycook et al., 2021), this finding offers hope that accuracy motivations-a normative ideal for an informed citizen-exist in real-world discussions about misinformation and that social media can be a productive space to foster accuracy-based discussions.
Practically, what can journalists, social media platforms, and ordinary social media users do in light of our findings?For journalists, changing news frames on misinformation is necessary to move past a contentious understanding of truth and falsehood.A first step is to de-emphasize party conflicts in coverage of misinformation policy and instead foster accuracy-based discussions by foregrounding how misinformation can impact the quality of people's knowledge and decisions in multilayered ways above and beyond partisan differences.Journalists can also foster accuracy-based discussions by explicitly offering factual corrections when sources in the story make false statements.It is fruitful for future research to examine journalistic norms about covering misinformation to understand the reasons behind the dominance of certain types of coverage.
For social media platforms, our research points to the potential for platforms to implement interventions that highlight accuracy to motivate social media users to productively engage in discussions, corrections, and actions addressing misinformation.In line with this argument, a growing number of experiments has suggested that simple accuracy nudges that ask people to think about how accurate a headline is can be an effective strategy to curb misinformation across national contexts such as the United States (Pennycook et al., 2021), Australia (Butler et al., 2023), Korea (Shin et al., 2023), and China (Xiang et al., 2023).
For everyday social media users, it is important to be self-reflective about the underlying reasons for why one is skeptical towards misinformation before enacting agency.Emerging evidence suggests that skepticism rooted in accuracy can help combat misinformation while skepticism rooted in partisan identity cannot (Li, 2023).Complementing these experimental studies on accuracy nudges and accuracymotivated skepticism, the current article looks at large-scale "real talk" by social media users and provides hopeful evidence that a substantial proportion of social media users do use an accuracy lens when expressing concerns about misinformation, signaling the realistic promise of fostering accuracy-based discussions among users.The descriptive findings in this study also provide a useful map of the variety of competing considerations that social media users have about misinformation, pointing to the need to ask why people are concerned about misinformation as compared to how much they are concerned.

Findings
Finding 1: Platform policy and party politics are the dominant lens through which misinformation is covered on mainstream broadcast news in the United States.
We started by describing discussions about social media misinformation in mainstream broadcast news (ABC, NBC, CBS) (Figure 1).The party politics discussion appeared in over a third of the mainstream broadcast news coverage about social media misinformation (36.63%).In comparison, accuracy-based discussions were less frequent: the quality of knowledge and decision discussion appeared in 23.76% of the segments and the direct corrections discussion appeared in 12.87% of the segments.As for actionbased discussions, only 7.92% of mainstream broadcast news coverage about social media misinformation discussed user agency, while 56.44% discussed platform policy.Finding 2: In mainstream broadcast news, platform policy, party politics, and accuracy concerns are concurrent rather than isolated considerations.
Table 1 illustrates the ways in which different types of discussions co-occur with each other in mainstream broadcast news.Table 1 reports both (a) the relative percentage of co-occurrence and (b) the absolute percentage of co-occurrence.The relative percentage of co-occurrence is measured by the Jaccard index (the number of news segments where two types of discussions are both present divided by the number of news segments where either type is present).The relative percentage should be the primary focus to interpret co-occurrence, as it takes account of the discussions being very rare in the whole dataset.The absolute percentage of co-occurrence is measured by the number of news segments where two types of discussions are both present divided by total number of news segments.
A main takeaway from Table 1 is that platform policy and party politics not only were the most prevalent discussions but also most frequently co-occurred with each other.On average, when news coverage does address platform policy or party politics, about a third of the times these two topics are discussed together (28.77%).For example, a CBS Evening News segment that aired on October 15, 2020, discussed how Twitter reacted faster to misinformation against Hunter Biden in 2020 than to misinformation against Hillary Clinton in 2016, both of which were endorsed by Trump campaign.By explaining how in both cases misinformation was used "to cast a cloud over the front runner in the closing weeks of the campaign" and reflecting on the previous election loss and "all the headaches" (CBS Evening News, 2020) of Democratic Party in 2016, the segment emphasized how Twitter's actions in 2016 versus in 2020 have implications for competitive advantages of political candidates.

Table 1. Co-occurrence of discussions in mainstream broadcast news.
Further, the accuracy-based discussion about quality of knowledge and decisions also co-occurred with discussions about platform policy and party politics, suggesting that accuracy motivation, partisan motivation, and call-to-action can be concurrent rather than isolated considerations.For example, a segment from NBC News Today Show that aired on October 28, 2020, juxtaposed accuracy and partisan concerns when discussing platform policy: JO LING KENT: Today, Mark Zuckerberg will face off with Congress yet again, preparing to defend the company's strategy to protect voters in 2020 … The CEOs of Twitter and Google also testifying today saying they, too, are taking action to combat disinformation and post that may incite violence before and after Election Day.But on both sides of the aisle, some Facebook users are worried about how divisive it's become.[…] JO LING KENT: Fifty-seven-year-old Romy Toussaint is mom to four sons.She used to love Facebook for staying in touch with family and friends, but now limits herself to fifteen minutes a day.How are you feeling about Facebook going into the 2020 election?How would you describe it?ROMY TOUSSAINT: I don't trust anything that I see on Facebook, and I don't think other people should.I don't want for my speech to be censored but I do want for Facebook to take responsibility to let people know what they think is not true.
JO LING KENT: On the other side, twenty-eight-year-old Molly Felling feels Facebook is biased.Some of her posts about coronavirus theories and hydroxychloroquine have been either flagged as false by the company or taken down.
MOLLY FELLING: I was shocked.I couldn't believe it.I think when it comes to an election, it's important to have both sides.JO LING KENT: Do you think that Facebook is targeting conservatives?Do you think that they go too far?MOLLY FELLING: I feel like they're a little more left leaning.You have to trust the public to be able to like inform themselves and make their own opinions.(NBC News Today Show, 2020) Through one interviewee's comment that "I do want for Facebook to take responsibility to let people know what they think is not true" and by pointing to the potential for misinformation to "incite violence before and after Election Day," the news segment noted how misinformation can impact the quality of knowledge and decisions, justifying the necessity for platform policy to be based on accuracy considerations.Meanwhile, the segment "balanced" the first soundbite with a second soundbite.By asking "Do you think that Facebook is targeting conservatives?Do you think that they go too far?" the journalist used the party politics lens to interpret platform policies, propagating the frame that Facebook's intervention on COVID-19 misinformation was ideologically biased and "left leaning."Notably, the story did not weigh in about whether Molly Felling's posts were accurate.
Finding 3: On Facebook and Twitter, platform policy and party politics still dominate the discussion, but user agency discussion takes a bigger share than in news.
Turning from news to Facebook and Twitter, the distributions of discussions largely held with minor differences (Figure 2).On social media, the party politics discussion constituted an even larger part of how social media users talk about misinformation than in news coverage (38.95%).Accuracy-based discussions, including quality of knowledge and decisions (22.13%) and direct corrections (12.31%) were relatively less frequent compared to the partisan-based discussion, replicating the pattern found in news.The platform policy discussion remained the most prevalent discussion (54.85%).Finally, the discussion on user agency, although still having a small share, was twice as frequent on social media (16.06%) than in news coverage (7.92%).Finding 4: On Facebook and Twitter, platform policy and party politics still co-occur most frequently, while user agency and direct corrections are frequently discussed together.
Table 2 shows the ways in which different types of discussions co-occur with each other on Facebook and Twitter.The relative percentage of co-occurrence is measured by the Jaccard index (the number of posts where two types of discussions are both present divided by the number of posts where either type is present).The relative percentage should be the primary focus to interpret co-occurrence, as it takes account of the discussions being very rare in the whole dataset.The absolute percentage of co-occurrence is measured by the number of posts where two types of discussions are both present divided by total number of posts.Similar to the pattern found in news coverage, we still observed the most frequent co-occurrence between platform policy and party politics on Facebook and Twitter (Table 2).For example, when quoting a tweet that said, "Twitter has not censored Joe Biden once but has censored Donald Trump more than 65 times," a Twitter user wrote, "Because Donald Trump uses Twitter to spread venom, misinformation and lies.Joe Biden doesn't.This isn't hard." Quality of knowledge and decisions also had frequent co-occurrence with platform policy, replicating the pattern in news coverage (e.g., "Glad these false videos with misinformation about hydroxychloroquine are being taken down by Twitter, Facebook, and YouTube.Social media can amplify voices representing a very small minority, which in cases like this, can be deadly").Again, these three types of discussions (platform policy, party politics, quality of knowledge and decisions) can co-occur together.For example, commenting about Facebook's action to delete a post by Donald Trump based on its misinformation policy, a Twitter user said, "among the many errors/lies/falsehoods that Trump makes, this seems trifling.But there are a number of ways such stupidity can result in real harm, or even death.It is safer to assume that Trump is wrong, as a general rule." On the other hand, a difference emerged between news and social media co-occurrence findings.While the discussion about user agency did not co-occur with direct corrections in our news data, it was frequently used together with direct corrections among social media utterances.For example, one post from a public Facebook page said, "A few people have sent me a video of 'America's Frontline Doctors' asking for my thoughts, so I decided to make a post.This video is dangerous.I will attempt to explain why."It went on to offer direct corrections to hydroxychloroquine misinformation, "As more and more trials took place to study if hydroxychloroquine was effective, data started to demonstrate it was not, and it only increased risk of harm […] As of now, there is a pretty strong consensus in the medical community that hydroxychloroquine does not work for COVID and it is not a cure," providing links to websites of National Institutes of Health and medical journals.Later, the post appealed to user agency by encouraging people to look for source credibility: "With all of that said, I invite you to consider the source of information […] When all else fails, always consider the source."Similarly, another Facebook post offered direct corrections to election misinformation circulating in a local Facebook group claiming that polling locations will be shut down, explaining the availability of in-person voting.The post went on to discuss user agency by reminding social media users "to be extra vigilant about any voting or election information that does not come from a verified source like your county clerk."Another post combined direct corrections and user agency by alerting people of impersonating social media accounts and asking social media users to act on it (e.g., "I need help from all of you to report these fake accounts").

Methods
To examine news coverage, we collected news transcripts discussing the topic of social media misinformation from three major broadcast networks, ABC, NBC, and CBS (n = 101) using relevant keywords. 2These sources were chosen based on their large audience size as well as on being broadly recognized as "mainstream media" even among partisans (Pew Research Center, 2021;Shearer & Mitchell, 2021); future work should expand beyond these sources in the centrist and center-left media ecosystem (Benkler et al., 2018) and examine sources with a wider range of ideological perspectives. 3 To examine social media discussions, we collected posts discussing topics relevant to social media misinformation from public Facebook pages and groups (n = 22,578) and from Twitter (n = 302,404) using relevant keywords. 4Facebook posts containing keywords were collected from CrowdTangle, a Meta-2 ("social media" OR "Facebook" OR "Twitter") AND ("misinformation" OR "disinformation" OR "fake news" OR "fake stories" OR "fake story" OR "fake information" OR "false news" OR "false stories" OR "false story" OR "false information" OR "made-up news" OR "made up news" OR "made-up stories" OR "made up stories" OR "made-up story" OR "made up story" OR "made-up information" OR "made up information") 3 Scholars disagree about the extent to which the three major networks are left-leaning and report considerable diversity between the three networks.Groseclose and Milyo's (2005) influential work on media bias found, "The sixth and seventh most centrist outlets are ABC World News Tonight and NBC Nightly News" (p.1221) and that the CBS Evening News was quite liberal.Groeling (2008) found evidence of left-leaning bias in NBC and CBS newscasts using favorable poll coverage of Bill Clinton, but found that ABC was more likely to report favorable polls for any president studied.Bernhardt et al. (2023) found more evidence of an antigovernment bias than an ideological bias, with network news coverage becoming more critical of the occupant of the White House, which switched from George W. Bush to Barack Obama during the time period they analyzed.Hassell et al. (2020) found no evidence of liberal bias in what mainstream news media chose to cover. 4 ("social media" OR "Facebook" OR "Twitter" OR "this platform") AND ("misinformation" OR "disinformation" OR "fake news" OR "fakenews" OR "fake stories" OR "fake story" OR "fake information" OR "false news" OR "false stories" OR "false owned research tool that archived all posts from Facebook pages and groups that were set as public.Twitter posts containing keywords were collected from Synthesio, a third-party commercial data vendor.We only collected English-language posts with a U.S. geographical location, but the keyword-based social media data were not representative of U.S. general population.We then eliminated content posted by news media and journalists from the social media data (see Appendix for details), and the final data consisted of 15,578 posts from public Facebook pages and groups and 301,334 posts from Twitter.The timeframe for both news and social media data was July 28 to December 3, 2020.
Two undergraduate coders labeled a random sample of 1,168 unique social media posts and the entire corpus of 101 news transcripts.Krippendorff's α was 0.85 for party politics, 0.94 for quality of knowledge and decisions, 0.92 for direct corrections, 0.76 for user agency, and 0.90 for platform policy.We used the Hopkins-King method to describe the proportions of discussions as research shows that it produces unbiased estimates for category proportions compared to individual classifiers (Hopkins & King, 2010).While the Hopkins-King method is superior in estimating category proportions, it does not offer information on how different discussions co-occur with each other because it does not classify individual documents.Thus, we tested three individual classifiers (Naïve Bayes, Support Vector Machine, Random Forest).Random Forest outperformed Support Vector Machine and Naïve Bayes in all discussions, producing an F1 score of 76.33% when classifying party politics, an F1 score of 78.61% when classifying quality of knowledge and decisions, an F1 score of 74.29% when classifying direct corrections, an F1 score of 73.44% when classifying user agency, and an F1 score of 79.09% when classifying platform policy.
a group, we expected the content captured by our data to be primarily consisted of social media users' discussion of news rather than news coverage itself.
The timeframe for both datasets was July 28 to December 3, 2020, covering months during the election campaign when the topic of social media misinformation became a salient topic, till one month after the 2020 presidential election.Pilot searches were used to validate that the timeframe included the most active periods for both the social media and the news media discussions.

Human labeling
Two undergraduate coders labeled a random sample of 1,168 unique social media posts and the entire corpus of 101 news transcripts.Coders labeled for the presence of the five types of discussions, which were not mutually exclusive.
For the party politics discussion, coders labeled 1 if the content discussed misinformation through a lens of party conflict and political (dis)advantage and labeled 0 if otherwise.The party politics discussion covered discussions about how misinformation benefited or harmed political groups, attribution of blame on a party/politician, and partisan politicians' attacks on one another.Examples of party politics discussion included "@Facebook throttled progressive news in favor of right-wing disinformation sites," "Four years ago BuzzFeed posted the Steele Dossier, filled with lies and misinformation, and Twitter and Facebook happily spread their fake news.Now these platforms are blocking transmission of a New York Post story critical of a Democrat.The hypocrisy is not sustainable," and "The Biden campaign is pressing Facebook to remove posts by Trump and slamming the company as the nation's foremost propagator of disinformation about the voting process."However, merely mentioning political parties, politicians, or elections to state factual information was not treated as a sufficient signal of party politics discussion unless the content explicitly discussed party conflict and political (dis)advantage (e.g., "Twitter temporarily prevented Donald Trump Jr. from tweeting and retweeting after he shared coronavirusrelated misinformation" was coded 0).
For the quality of knowledge and decisions discussion, coders labeled 1 if the content discussed misinformation in relation to the accuracy of knowledge and/or the quality of decisions (e.g., misinformation makes people believe that there is an election fraud, "poisons your mind," or polarizes the public), and otherwise 0. Such discussions can be either on an individual level (e.g., "He admitted that he wanted to kill a 'random white person' after watching the fake news") or a collective level (e.g., "COVID-19 disinformation is everywhere.And it's undermining our response in the US").
For the direct correction discussion, coders labeled 1 if the content offered explanation on why a piece of information is false, and otherwise 0. Coders used a stringent criterion to label the direct correction discussion: the content needed to both identify misinformation and provide information on why it is false.Examples of direct correction discussion included: "A video that falsely claims Joe Biden wore a wire during last night's debate has been shared thousands of times on Facebook.The clip shows a shirt crease."and "Mark Zuckerberg's claim that QAnon is *just now* evolving from misinformation to violence is patently false.An armed QAnon supporter blocked traffic at the Hoover Dam bridge in June 2018.The FBI called it a domestic terror threat in May 2019."Content merely saying "this is fake news" or "this is misinformation" was labeled as 0.
For the user agency discussion, coders labeled 1 for content describing or calling for actions from ordinary people to counter misinformation, and otherwise 0. Examples of the user agency discussion included content asking social media users to "be aware of" misinformation, warning people to "watch out for" certain rumors recently circulating online, encouraging people to read and share corrections/facts with others (e.g., "join my page to watch live fact-check"), and asking people "to make judgement based on facts not misinformation," etc.Only actions specifically targeted at countering misinformation were coded as 1; actions to spread misinformation (e.g., asking people to save/access content that has been constructed where the majority category was randomly sampled to be the same size as the minority category) and an over-sampling technique (the cross-validation resampling procedure incorporated an over-sampling process where the minority category was randomly sampled with replacement to be the same size as the majority category).For each type of message, we tested the performance of different algorithm specifications, inspected the confusion matrices, and used accuracy and F1 score to select the best-performing algorithm.Random Forest outperformed SVM and Naïve Bayes in all types of discussions, producing an accuracy of 77.31% and an F1 score of 76.33% when classifying party politics, an accuracy of 78.71% and an F1 score of 78.61% when classifying quality of knowledge and decisions, an accuracy of 93.13% and an F1 score of 74.29% when classifying direct corrections, an accuracy of 90.48% and an F1 score of 73.44% when classifying user agency, and an accuracy of 80.59% and an F1 score of 79.09% when classifying platform policy.

Figure 1 .
Figure 1.Prevalence of discussions in mainstream broadcast news.Figure shows percentage of each type of discussion in a total of 101 news transcripts analyzed.

Figure 2 .
Figure 2. Prevalence of discussions on Facebook and Twitter.Figure shows the algorithm-estimated percentage of each type of discussion in a total of 15,578 Facebook posts and 301,334 Twitter posts analyzed.