1.1 Definition and Examples

What is misinformation? Misinformation is false information that may be generated and spread either intentionally or unintentionally. When an actor intentionally propagates false or misleading content, this is considered disinformation – a subset of misinformation (Guess & Lyons, Reference Guess, Lyons, Persily and Tucker2020). The question of what is “false” can undoubtedly give rise to complex epistemological concerns, though. Typically, definitions of misinformation focus on whether information contradicts expert consensus (Nyhan & Reifler, Reference Nyhan and Reifler2010; Tucker et al., Reference Tucker, Guess and Barberá2018; Guess & Lyons, Reference Guess, Lyons, Persily and Tucker2020). However, expertise can be contested (who counts as an expert?), and inappropriate experts can be platformed and amplified by contesting sides (Vraga & Bode, Reference Vraga and Bode2020). Further, there are many issues for which expert consensus is not available. According to Vraga and Bode (Reference Vraga and Bode2020, p. 138),

Even then, evidence is subject to change – especially in fields like public health (Lyons et al., Reference Lyons, Merola and Reifler2020). And for some questions, contradictory or speculative evidence is all there is. In such cases misinformation is “temporally bound,” such that what is considered true at one point in time may shift as new evidence is gathered (Vraga & Bode, Reference Vraga and Bode2020). In any event, while many cases of misinformation are clear-cut and easily categorized, we must acknowledge that others are more nuanced.

Misinformation and disinformation on social media come from a range of sources – including anonymous users, influencers, alternative media, mass media, elected officials, and institutions. As such, their dissemination is driven by many goals – ideological, electoral, geopolitical, economic, and even simply chaotic (Petersen et al., Reference Petersen, Osmundsen and Arceneaux2023), as well as the purely accidental. Some forms of misinformation are more dangerous than others. Often, a particular piece of false information causes no direct harm. But in other cases, such as misinformation surrounding childhood vaccines, for instance, its effects can be harmful not only to the audience who accepts it but also to the wider community. Further, disinformation can intentionally inflame racial animus by way of “identity propaganda,” or “narratives that strategically target and exploit identity-based differences in accord with pre-existing power structures to maintain hegemonic social orders” (Reddi et al., Reference Reddi, Kuo and Kreiss2023, p. 2201). Likewise, misinformation may stoke anti-immigrant sentiment (Abascal et al., Reference Abascal, Huang and Tran2021) or be used to justify antidemocratic behaviors (as in recent events in Brazil, the US, and elsewhere) with violent consequences, particularly when it comes from the top (Rossini et al., Reference Rossini, Mont’Alverne and Kalogeropoulos2023). Social media-fueled misinformation notably led to nearly two dozen people being killed by vigilante mobs in India in 2018 (Goel et al., Reference Goel, Raj and Ravichandran2018).

Even in more mundane cases, misinformation can cause voters to support candidates who are not in line with their true underlying preferences (reducing “correct voting”; Robison, Reference Robison2021). Its impact on the political realm is not strictly a problem of influencing a single belief and its subsequent behavior, though. There are broader systemic effects to consider (Guess & Lyons, Reference Guess, Lyons, Persily and Tucker2020). A constant refrain about a deluge of misinformation and public discourse surrounding its prevalence can give politicians cover for their actual intentions to deceive the public (Van Duyn & Collier, Reference Van Duyn and Collier2019; Ulusoy et al., Reference Ulusoy, Carnahan and Bergan2021). In other words, there are cumulative effects on trust in institutions that outweigh single-case effects, given the second-order influence on a vast array of reliable information downstream.

Considering the nature of socially mediated misinformation puts additional focus on the way these platforms transform older modes of information sharing. Put simply, the combination of social media users’ psychology and platform affordances can help spread misinformation in previously unseen ways and speeds. Users are frequently found within like-minded clusters, despite mixed research findings on the prevalence of echo chambers (Guess et al., Reference Guess, Lyons, Nyhan and Reifler2018). These network configurations can foster the exposure and dissemination of agreeable misinformation (Del Vicario et al., Reference Del Vicario, Bessi and Zollo2016; Shin et al., Reference Shin, Jian, Driscoll and Bar2017). Similarly, social media users tend to place considerable trust in their close friends. When it comes to trust in news shared on platforms like Facebook, studies indicate that the person sharing the news matters more than the news organization itself (American Press Institute, 2017). Users are more inclined to perceive news as accurate and unbiased when it is shared by someone they trust. This dynamic may contribute to the spread of misinformation, especially since platforms prioritize displaying posts from close friends to drive engagement. Consequently, algorithmic bias emerges as a result of the design choices made by most social media platforms, which prioritize engagement by favoring popular content over trustworthy information in users’ feeds (Ciampaglia et al., Reference Ciampaglia, Nematzadeh, Menczer and Flammini2018).

1.2 The State of Solutions

There are of course numerous top-down misinformation mitigation strategies that fall at the policy or platform level. Perhaps most indirectly, comparative research on media systems and Western democracies’ resilience to disinformation suggest increased funding for public media can help buffer and dilute the power of propaganda and other mistruths (Humprecht et al., Reference Humprecht, Esser, Aelst, Staender and Morosoli2023). Political maneuvers have been initiated, such as France’s legal framework that empowers judges to order the removal of “fake news.” Platforms meanwhile have attempted various forms of deplatforming and demonetization of accounts spreading misleading information, but to mixed results (Ribeiro et al., Reference Ribeiro, Jhaver, Reignier-Tayar and West2024). At the level of individual posts, different social platforms have engaged in moderation, flagging, and debunking techniques – also to mixed results. When Twitter flagged Donald Trump’s tweets about the 2020 election as false, for instance, messages with warning labels spread further than those without labels, while messages that had been blocked on Twitter remained popular on Facebook, Instagram, and Reddit (Sanderson et al., Reference Sanderson, Brown, Bonneau, Nagler and Tucker2021). Facebook’s warning labels for false stories, meanwhile, have been found to promote overly skeptical evaluations of legitimate news (Clayton et al., Reference Clayton, Blair and Busam2020). Efficacy aside, policy and platform level interventions also require tremendous political will to enact and may cause backlash among the public or platforms’ user bases due to discomfort with censorship (Saltz et al., Reference Saltz, Barari, Leibowicz and Wardle2021).

At the individual level, though, a handful of promising “bottom-up” interventions have been found to reduce belief in and spread of false news stories, many of them replicated across numerous countries. Accuracy prompts – which address the problem of inattention and carelessness and simply intend to remind users that the accuracy of the information they share is important – have been found to reliably improve the public’s discernment of true and false news and improve the quality of information they share in the wild (Pennycook et al., Reference Pennycook, Epstein, Mosleh, Arechar, Eckles and Rand2021b). However, these prompts likely cannot convince a reader that information they have considered at length and strongly endorse is in fact wrong.

Other researchers have studied the potential to educate or train news consumers to be more digitally literate and aware of manipulation techniques. Lateral reading encourages news consumers to act like a professional fact-checker by verifying the veracity of each story in concurrent in-depth web searches about the authors, sources, and topic. This technique has been found to improve the ability of students to assess credibility of digital content, but is relatively intensive, requiring hours if not months of training sessions (Wineburg et al., Reference Wineburg, Breakstone and McGrew2022). While potentially viable if integrated in school curricula, this sort of solution is more difficult to provide to broader adult populations. Brief digital literacy tips have shown promise as a scalable, low-investment approach that likewise improves discernment between true and false information (Guess et al., Reference Guess, Lerner and Lyons2020). Though effective up to several weeks later, the effects of such tips appear to decline and would require refreshing. Lastly, researchers have developed educational games to help people spot manipulation techniques (Roozenbeek et al., Reference Roozenbeek, Van Der Linden and Goldberg2022). While effective at reducing belief in false content, these approaches (as others) might cause a significant degree of skepticism in reliable news content (Modirrousta-Galian & Higham, Reference Modirrousta-Galian and Higham2023).

In general, many of these interventions may have significant unintended effects in causing such cynicism (Acerbi et al., Reference Acerbi, Altay and Mercier2022) and in the prioritization of “doing your own research” (Chinn & Hasell, Reference Chinn and Hasell2023). However, “doing your own research” may simply be a proxy for anti-expert views, rather than the embrace of cautious information search. Similarly, given the rise of such sentiments globally (Lyons et al., Reference Lyons, Montgomery and Guess2021; Han et al., Reference Han, Blackburn and Jeftić2022; Spälti et al., Reference Spälti, Lyons and Stoeckel2023), interventions aimed at urging users to privilege authoritative mainstream media outlets and expert commentary may be ineffective among those who consume and share the highest volume of misleading content (Lyons et al., Reference Lyons, Montgomery and Guess2021). This points to the need for solutions that do not rely exclusively on authoritative sources, but rather (correct) information relayed by trusted interpersonal contacts.

1.3 Social Corrections

Social media users play a considerable role in spreading false information and dubious news stories online (Guess & Lyons, Reference Guess, Lyons, Persily and Tucker2020). But just as crucially, average users can help temper its spread. A significant portion of social media users engage in correcting others when they come across misinformation online, actions which are visible to an even larger group (Chadwick & Vaccari, Reference Chadwick and Vaccari2019; Bode & Vraga, Reference Bode and Vraga2021a, Reference Bode and Vraga2021b). When surveyed, most Americans not only expressed appreciation for these corrections, but also considered it a public responsibility (Bode & Vraga, Reference Bode and Vraga2021a). Such social corrections (also known as observational corrections, i.e., corrective cues placed by other social media users) have been shown to be effective at preventing the spread of health-related online misinformation (e.g., Bode & Vraga, Reference Bode and Vraga2018). Clearly, it is important to better understand this behavior in terms of how much good (or harm) it may cause.

Seeing others add corrective cues to a post can (1) reduce how accurate one perceives this post to be, (2) reduce the probability of interacting with this post, (3) alter one’s attitudes toward the post, and (4) decrease one’s intention to do what the post recommends (Bode & Vraga, Reference Bode and Vraga2018; Margolin et al., Reference Margolin, Hannak and Weber2018; Colliander, Reference Colliander2019; Tully et al., Reference Tully, Bode and Vraga2020; Walter et al., Reference Walter, Brooks, Saucier and Suresh2021; Badrinathan & Chauchard, Reference 70Badrinathan and Chauchard2023). There are several favorable aspects to social corrections as a buffer against misinformation. First, such behavior is already a considerable aspect of online discussions, organically growing out of the conversational norms of online platforms. Along these lines, then, they represent a highly scalable approach to curbing the spread of misinformation. Even if a minority of social media users engage in corrective action, this represents millions of engaged social fact-checkers – free of charge. Perhaps just as importantly, social corrections also come from sources that other users, especially those with strong anti-expert sentiments or mainstream media distrust, are more likely to trust.

On the other hand, there are a number of potential drawbacks to relying on this strategy. First, there are factors working to disincentivize this behavior. Social corrections rely on individuals to act in the public good – and even then, to do so, they must first be interested in news and public affairs. Most users on social platforms do not post much about hard news and politics given preferences for entertainment and social content (Kalogeropoulos et al., Reference Kalogeropoulos, Negredo, Picone and Nielsen2017; Moretto et al., Reference Moretto, Ortellado and Kessler2022). The social dynamics of correcting friends and family members can also be dicey, especially for younger users (Vijaykumar et al., Reference Vijaykumar, Rogerson and Jin2022), further disincentivizing this behavior. In short, there are costs to engaging with misinformation, which can act as a barrier.

Further, important questions about the generalizability of social corrections’ efficacy remain: Do these corrections work beyond the health context in the US? Does the effect of social corrections depend on their form and strength? Researchers testing the effect of social corrections have mostly focused on health topics with US samples (Bode & Vraga, Reference Bode and Vraga2018; Tully et al., Reference Tully, Bode and Vraga2020; Walter et al., Reference Walter, Brooks, Saucier and Suresh2021; for two exceptions, see Boot et al., Reference Boot, Dijkstra and Zwaan2021; Badrinathan & Chauchard, Reference 70Badrinathan and Chauchard2023). It is particularly important to examine whether such messages are effective for more contested topics (e.g., political news) that might trigger directional-motivated reasoning and hence rejection, as some work suggests social corrections may be more limited in these cases (Vraga et al., Reference Vraga, Kim and Cook2019). Second, existing research has operationalized social corrections in different ways – from subtle, standard social media reactions (e.g., like, angry emoji) to substantiated corrective comments with links to bolstering webpages that have been reposted many times (Tully et al., Reference Tully, Bode and Vraga2020). Hence, we do not know to what extent the form and strength of social correction measures determine their effect.

Lastly, average users are not experts. Therefore, they may unknowingly engage in miscorrection. Does falsely “correcting” true news items have similar but, given their veracity, less desirable effects? Vraga and Bode (Reference Vraga and Bode2022) warn that social miscorrections of true news might amplify the spread of misinformation, at least in some cases. Specifically, they find that when social media users flag factually accurate information – for example, tick bites can trigger an allergy to red meat – as incorrect, people are less likely to believe this information. Our research is dedicated to better understanding these potential strengths and weaknesses of social corrections.

1.4 Overview of the Element

In Section 2, we map the nature of the problem by examining respondents’ ability to correctly identify false and true news in social media posts in the UK, Italy, and Germany. Our analysis is based on forty-seven different posts that span a wide range of topics, drawn from actual online content. The false news posts originate from material flagged by fact-checking organizations in the respective countries. We also explore which types of news stories are particularly difficult to assess and identify the correlates of (in)correctly perceiving their veracity. Section 3 presents the experimental research design. We use a within-subjects design and conduct fieldwork in three countries, allowing us to assess the homogeneity of effects across different European contexts. In Section 4, we discuss the results: Based on experiments conducted in the UK, Italy, and Germany, we find that social corrections can help people identify false news. However, we also find that miscorrections – comments which raise doubts about true news – also affect accuracy perceptions. Miscorrections decrease the perceived accuracy of true news, even when the logo of a news outlet is displayed. In this section, we also outline potential mechanisms. Our results support the notion that user comments seem to work as (warning) signals that are processed rather superficially instead of triggering more complex reasoning. In Section 5, we use additional survey data from Germany to shed light on people’s views on user comments and the characteristics of people who comment often. In Section 6, we discuss the implications. For instance, our results show that digital literacy is important not only to help people differentiate between true and false content but also to shield them from misleading comments from other users. The potential of corrective comments lies in the fact that they offer all users a way to improve the information environment on social media, even if platforms do not act. Table 1 shows an overview of key terms used in this Element.

2.1 The Challenge of True and False News Identification on Social Media

To better understand how people evaluate and respond to social media content that contains both true and false claims, we conducted studies in three countries (UK, Italy, and Germany). In each of these countries, we use non-probability quota samples that reflect the populations with regard to age, gender, education, and region.

We generated forty-seven stimuli, all derived from social media posts that were shared on different platforms. The material was collected in the months leading up to the fieldwork: January–July 2022 for the UK, February–August 2022 for Italy, and July 2022–January 2023 for Germany. False news stories were identified through fact-checking organizations in each country, ensuring that they were sufficiently prominent to be flagged for verification. COVID-19 was particularly prevalent among the debunked materials, but our goal was to cover a broad range of topics (Clifford & Rainey, Reference Clifford and Rainey2025). To achieve this, we included posts on several public health issues (e.g., COVID-19, vaccines, smoking), technology (e.g., the 5G cellphone network), climate change, and Russia’s attack on Ukraine (see the online appendix for all details).

Different topics are likely to vary in the extent to which they trigger motivated reasoning, meaning that respondents may be more inclined to believe misinformation on some topics than others. To account for this, we measured respondents’ attitudes toward the topics of the stimuli, allowing us to control for ideological alignment in the experimental analysis (a detailed presentation of the experimental research design follows in Section 3).

Although our selection of stimuli is not a representative sample, we are confident that it captures a realistic range of the types of content people encounter on social media. Our approach includes a larger number of news headlines than many comparable studies, which often focus on a single topic, a single platform, or a single country. Importantly, our design incorporates both true and false headlines (Pennycook et al., Reference Pennycook, Binnendyk, Newton and Rand2021a). By covering multiple topics, reflecting diverse social media platforms, and spanning three countries, our study provides broader generalizability than previous research.

The headlines reflect what has been shared on social media, but we recreated the posts with an open-source social media simulator (Zeoob.com) rather than using screenshots. This approach made it possible to randomly generate user names and faces from a photo database (Flickr-Faces-HQ Dataset, FFHQ). Moreover, by recreating the posts, we were able to control features like the number of likes and comments, which was essential for our experimental treatments.

Figure 1 shows the perceived accuracy of the various false claim headlines used in our studies. We find variation of rating headlines accurate, ranging from 18 percent (for a headline about Russia’s invasion of Ukraine, see Figure 2) to 59 percent of respondents (for a headline about the greenhouse gas emissions from private jets flown to the COP26 climate summit in Glasgow, see Figure 3).

Figure 1 Percentage of false news headlines perceived as accurate or very accurate (combined share). Respondents rated all posts on a four-point scale: not at all accurate, not accurate, accurate, and very accurate. The sample includes only posts without user comments (i.e., the control condition). The figure pools posts from the UK, Italy, and Germany, displaying short titles. The full text of all posts with translations is available in the online appendix.

Figure 2 Example of a social media post used in Italy (false news). False news post identified as accurate by the smallest share of respondents. Translation: “Italy is already at war; the army heads toward Russia’s new campaign.”

Figure 3 Example of a social media post used in the UK (false news). COP26 jets emit tons of CO2. False news that is identified as accurate by highest share of respondents.

Turning next to the perceived accuracy of true news, we again see substantial variation (Figure 4). Comfortingly, as a test of face validity of our stimuli, our samples rate these headlines as significantly more accurate, on average, than the headlines with false claims. The true news story that is considered accurate by the largest number of people is one about long COVID (Figure 5). Eighty-five percent of respondents consider this story to be true. The true news story that is perceived as accurate by the smallest share of respondents is regarded as misinformation by almost three out of five respondents (57 percent). It is a story about child soldiers in Ukraine (Figure 6). While only one false news story was perceived as accurate by half of the sample, four of the true news stories were not considered accurate by half the sample or more (“Saharan dust melts glaciers,” “Lauterbach almost abducted,” “Monkeypox emergency US,” “Putin recruits children”).

Figure 4 Percentage of true news headlines perceived as accurate or very accurate (combined share). Respondents rated all posts on a four-point scale: not at all accurate, not accurate, accurate, and very accurate. The sample includes only posts without user comments (i.e., the control condition). The figure pools posts from the UK, Italy, and Germany, displaying short titles. The full text of all posts with translations is available in the online appendix.

Figure 5 Example of a social media post used in Germany (true news). Long Covid symptoms. True news considered accurate by highest share of respondents. Translation: “Long Covid: common symptoms; after two years of the pandemic, it is still difficult to detect long Covid. (…)”

Figure 6 Example of a social media post used in Germany (true news). True news considered accurate by a comparatively small share of respondents. Headline translation: “Putin recruits and raises his future soldiers in Ukrainian Mariupol. (…)”

A challenge for any work focusing on social media content is choosing what content to include in experimental studies. We would like to reiterate that we do not claim that the set of headlines we have selected for inclusion in our study is a random sample of all content. Thus, we discourage any interpretation that focuses on how much people generally believe or do not believe social media headlines in general. However, we do believe that our broader-than-average pool of headlines allows us to make more generalizable claims about what factors are associated with perceived accuracy.

2.3 The Correlates of Believing Online Content

What factors are associated with believing – or disbelieving – the content we see online? As a general rule, people are biased toward accepting incoming information (Brashier & Marsh, Reference Brashier and Marsh2020). There are perhaps many competing and complementary explanations for this predilection to generally accept claims as true – most claims we encounter are true, telling the truth is a societal norm, rejecting and/or counterarguing the content we see online takes cognitive effort, and so on (Schwarz et al., Reference Schwarz, Newman and Leach2016; Abeler et al., Reference Abeler, Nosenzo and Raymond2019). In this section, we examine the factors associated with believing online content, and the extent to which these markers are similar for true and false content.

2.4 The Correlates of Belief in False and True News in Our Data

We use key variables mentioned in the literature to examine susceptibility to the false news stories in our sample, and we use the same predictors to analyze disbelief in true news (e.g., Pennycook & Rand, Reference Pennycook and Rand2019b; Bryanov & Vziatysheva, Reference Bryanov and Vziatysheva2021; Lyons et al., Reference Lyons, Montgomery and Guess2021). We find patterns that are consistent with the literature and we also unearth novel findings (Figure 7). Respondents are more likely to consider false news as accurate when the information is in line with their attitude toward that issue (e.g., individuals who do not find vaccines helpful are more likely to consider misinformation about vaccines to be true; for more information on our measure for ideological alignment, see Section 3 and the online appendix for our coding). The strongest predictor in the model is a far-right ideological position; far-right positions are associated with a greater likelihood to find false news accurate. Anti-expert positions are also linked to a greater likelihood of believing false news. In our data, older respondents are better equipped to identify false news. Here, it is important to note the above mentioned discrepancy with digital trace data: While older individuals are often better at identifying false news in surveys, they seem to share false news more often according to behavioral data (Lyons et al., Reference Lyons, Montgomery and Reifler2024). Cognitive reflection capabilities help people to identify false news as well. We do not find that education levels and gender are associated with variation in accuracy perceptions.

Figure 7 The correlates of false and true news susceptibility. Authors’ data. Multilevel model with post ratings nested within individuals. N (false news posts) = 6,663, N (true news posts) = 7,154. Outcome is perceived accuracy of a social media post (range = not at all accurate, not accurate, accurate, very accurate). Lines show 95 percent confidence intervals.

What about disbelieving true news? Are the predictors that explain false news susceptibility also linked to disbelief in true news? Here, we find a more mixed picture. Ideological alignment of information plays a similar role on both sides. For instance, when a story is ideologically aligned – i.e. in line with preexisting beliefs – individuals are more able to conceive it to be true, irrespective of veracity: That is, the effect holds for false and true information. Individuals who score high on SSI are also more likely to find true as well as false news accurate. This means that they are more likely to identify false news incorrectly, because it is assumed to be accurate.

Individuals who score high on anti-expert sentiments are more likely to believe that false news is accurate but less likely to find true news accuracy. Hence, they are more likely to incorrectly identify information on both counts. While we find that people are less likely to believe false news, they are more likely to find true news accurate. Cognitive reflection skills make it less likely for people to identify false news as true but not more likely to identify true news as accurate. People with higher levels of education are more likely to consider accurate information as true. When it comes to ideology (beyond anti-expert sentiments), we find that individuals without party attachment are significantly less likely to identify true news as accurate. Individuals with moderate left leaning views are more likely to rate true news as accurate.

2.5 Conclusion

Based on the false and true news that we use as material in our experiments, we can show much variation in belief in false news and disbelief in true news. If individuals are asked to rate the accuracy of a set of true news and false news stories, we find that some false stories are considered accurate by a much larger share of people than others. The difference that we document is large; all false news stories are considered accurate by at least one fifth of respondents, while some are considered accurate by a much larger share. In a similar vein, a lot of true news stories are considered accurate and there is also much variation. We also find important regularities among the characteristics that are associated with the extent to which people find information accurate, irrespective of whether it is true or false. For instance, individuals are more likely to find ideologically aligned news stories more accurate, irrespective of veracity. In contrast, anti-expert sentiments increase the likelihood of an individual finding false news accurate and true news inaccurate.

3.1 Material

The content of our material was taken from real-world social media posts. Thus, we did not create false news for the purposes of our experiments. We used a social media simulator (https://zeoob.com/) to generate the treatment material based on existing posts and to design the various treatment versions. Every post was generated in four versions (control and three treatment conditions). To maximize external validity, our treatment material shows comments written and posted by social media users, though these comments are not necessarily from the posts that we display. We did not use the original names of people who posted the stories we use, but generated names with another simulator (https://britishsurnames.co.uk/random). For profile pictures, we rely on material from the Flickr-Faces-HQ Dataset (FFHQ) (Karras et al., Reference Karras, Laine and Aila2019). We created a post with the original claim and used those as post for all treatment conditions. Features of a post (online user status, activiation of fields, location tags, etc.) were kept constant across posts and conditions.

In all countries, we used a treatment condition where corrective and miscorrective comments respectively include a link. In the case of false news, these are links to websites that officially debunk content. (e.g., from https://fullfact.org/). In the case of true news, miscorrections include a link to a website that seemingly bolsters the (false) claim of the miscorrection.

3.2 Dependent Variables

3.3 Moderators: Anti-expert Sentiments, Cognitive Reflection, and Susceptibility to Social Influence

In addition to testing the main effects, we also examine whether the treatment effects vary depending on respondents’ cognitive reflection capabilities, their susceptibility to social influence, and anti-expert sentiments. These variables correlate with the ability of people to identify false information (Section 2), but these variables might also condition the effect of our treatments. For instance, individuals with stronger cognitive reflection skills might be more likely to take comments into account when reading a post and they might not be taken into account at all by those with low cognitive reflection skills. If that is the case, they would only be effective among those with moderate or strong cognitive reflection skills. We use a common measure for cognitive reflection skills that combines four items (Pennycook & Rand, Reference Pennycook and Rand2019b). Each of these items has an answer that might intuitively be right but is in fact wrong, and careful reflection helps to answer correctly. For instance, one of the questions is the following: “If you’re running a race and you pass the person in second place, what place are you in?” The answer options are first, second, third, or fourth (with “second” being the correct answer).

In a similar vein, we test if susceptibility to social influence and anti-expert sentiments moderate the role of the treatments. Individuals who are more susceptible to social influence might be more affected by other users’ reactions to content, and therefore react more strongly to signals from others. This variable also helps us to better understand the mechanism through which corrections might work: If corrections work because people align their views to accuracy perceptions of others, one could expect those who place greater care more about the views of others generally to exhibit greater treatment effects. We measure susceptibility to social influence using a seven-item scale (Stöckli et al., Reference Stöckli, Henry and Bartsch2020), which includes items such as: “I based my decisions on information I gain from online social network posts and interactions” (with a five point response scale that ranges from strong agreement to strong disagreement). Finally, we test if the effect of the treatment differs depending on the strength of respondents’ anti-expert sentiments. People who see experts negatively might be less receptive to other people’s knowledge and their messages about content accuracy. We use a common measure from the literature that combines three items; for example, “I am more confident in my opinion than other people’s facts” (Han et al., Reference Han, Blackburn and Jeftić2022; Spälti et al., Reference Spälti, Lyons and Stoeckel2023). We also phrased our expectation for an interaction as a hypothesis for the fieldwork in the UK, and as a research question in the subsequent fieldwork in Italy and Germany.

3.4 Fieldwork

3.5 Modeling

We preregistered all fieldwork on OSF(UK: https://osf.io/fpm2e/?view_only=1f2999c931c84404bddd618ce33208bd; Italy: https://osf.io/upzm8/?view_only=55393ea5c1634c2ea87c793f0cfc07d3; Germany: https://osf.io/rfq6h/?view_only=26e88087c3c442ff810b6ce452736e75). We preregistered the fieldwork separately for each country, with the first fieldwork in the UK informing our preregistrations for fieldwork in Italy and Germany. The preregistrations specify hypotheses for the main effects as well as for interactions. As shown in our previous work (Stoeckel et al., Reference Stoeckel, Stöckli and Ceka2024), the results hold when not controlling for ideological alignment. In order to probe the results further, we present analyses in Section 4 that additionally control for ideological alignment of the content that respondents assessed (which is a substantively important deviation from the preregistrations).

Each respondent rated a set of nine social media posts, which allows us to leverage a within subjects design. We employed a randomization at the level of posts, that is, respondents could see each post either in the control or in one of the three treatment conditions. In order to model this data structure, we use linear mixed-effects models as specified in the preregistrations. Mixed-effects models account for a data structure in which observations (assessments of posts) are nested within individuals (Steenbergen & Jones, Reference Steenbergen and Jones2002). We run separate analyses for each country.

For the main effect of social corrections on the three response variables of interest – perceived accuracy, probability to “like”, and probability to share false news – we entered the social correction treatment (4-level factor) as main predictor into a model. We also added demographics as covariates (age, gender, education).

Our models also control for ideological alignment of news stories (albeit the results hold if ideological alignment is excluded; Stoeckel et al., Reference Stoeckel, Stöckli and Ceka2024). Controlling for ideological alignment allows us to account for the fact that individuals are more likely to find information accurate if it aligns with their beliefs or opinions. In research in the US context, a common procedure to control for ideological alignment is to account for the party affiliation of respondents (Lyons et al., Reference Lyons, Montgomery and Guess2021). In our case, we asked respondents pretreatment questions about the topics that come up in the treatment material. For instance, we would ask respondents their view on COVID-19 vaccines. Misinformation that portrays COVID-19 vaccines as harmful would be ideologically aligned for individuals who see vaccines negatively. By controlling for ideological alignment, we can test the robustness of social corrections even when they are not ideologically aligned. The ideological alignment measures for each social media post can be found in the online appendix.

The model includes random intercepts for both respondents and social media posts to account for the nested structure of the data. We apply the same modeling approach when analyzing the effects of miscorrections on perceptions of true news.

$\begin{array}{ll}Response\hfill & ~socialcorrectiontreatment+gender+age+education+ideologicalalignment+\hfill \\ \hfill & \begin{array}{l} \left(1|respondentid\right)+\\ \left(1|socialmediapost\right)\end{array}\hfill \end{array}$

We also examine whether treatment effects vary across individual characteristics – specifically anti-expert sentiments, cognitive reflection (CRT), and susceptibility to social influence (SSI). To do so, we estimate interaction models that include each individual difference measure and its interaction with the social correction treatment:

$\begin{array}{ll}response\hfill & ~socialcorrectiontreatment*individualdifferencemeasure+gender+age +education+ \hfill \\ \hfill & \left(1|respondentid\right)+\hfill \\ \hfill & \left(1|socialmediapost\right)\hfill \end{array}$

Identifying interaction effects typically requires larger sample sizes than detecting main effects. Therefore, our power calculations focused on ensuring sufficient statistical power to detect a small interaction between social corrections and individual-level moderators, such as susceptibility to social influence or anti-expert attitudes. The linear mixed-effects model treats perceived accuracy as a repeated measure and includes a random intercept for each social media post. Based on pilot data, and assuming an effect size of –0.07 for the difference in slope between the control condition and the high amplification + link condition, we determined that a minimum sample size of 1,800 respondents would be necessary to achieve 80 percent statistical power (see preregistration for further details).

4.1 A Debate on the Mechanism

Several mechanisms are discussed in the literature that could explain why corrective comments from other users might affect accuracy perceptions of content that users see on social media. There are key differences across potential mechanisms regarding the extent to which users process information. According to one set of studies, users process both the content of a post as well as contextual information such as comments, and features of these comments, such as their source or credibility (e.g., Vraga & Bode, Reference 86Vraga and Bode2017; Colliander, Reference Colliander2019; Badrinathan & Chauchard, Reference 70Badrinathan and Chauchard2023). Thus, there is a possibility that comments are processed in a central or systematic manner. We discuss the related approaches in Sections 4.1.1 and 4.1.2. An alternative assumption is that users do not in fact spend a lot of cognitive resources on both the content and comments, which has consequences for the way corrections (and miscorrections) affect users (e.g., Haugtvedt and Wegener; Reference Haugtvedt and Wegener1994; Anspach & Carlson, Reference Anspach and Carlson2020). We briefly outline the related mechanisms in Sections 4.1.3 and 4.1.4. As with any mental process, it is difficult to directly test the mechanism that drives the effect of social corrections. That said, we can use our data to shed light on this issue. The results from subgroup analyses do not support the notion of users processing social media content and related comments in a central way; instead, it seems more likely that processing is superficial and recency effects dominate.

4.2 Social Corrections Help People to Spot False News

We find that social corrections consistently reduce the perceived accuracy of false news and they decrease the intention of users to like and share such posts. We test three types of corrective comment interventions on these outcomes in three countries (see Section 3 for a more extensive discussion of the experimental design). We find statistically significant effects for 25 out of the 29 relationships (Figure 9). Thus, there is a systematic pattern across countries, in the context of a variety of false news stories and different platforms (Instagram, Facebook, and Twitter). The similarity of treatment effects is particularly noteworthy given that we increased the signal strength of the corrective comments in some cases. The pattern that we find suggests that corrections can be effective even when they do not include a source; moreover, an amplified signal does not necessarily lead to a greater effect. These results warrant a closer look, both because this is important to identify best practices and because these results contrast with some prior findings in the literature.

Figure 9 Effects of corrections as well as miscorrections on perceived accuracy, proclivity to like and share false as well as true news.

Note: Circles, squares, and triangles show point estimates. Lines show 95 percent confidence intervals. Multilevel models that control for ideological alignment of the respective post and demographic characteristics of respondents (age, gender, and level of education). Low amp.: low amplification (few likes or comments); high amp.: high amplification (many likes or comments). The corresponding results tables can be found in the online appendix.

We conducted our fieldwork sequentially, so results and insights from one country could help inform our decisions about what to field in the next country. As a consequence, there are some subtle differences in our experiments across the three national contexts of the UK, Italy, and Germany. To foreground our results across countries, treatments that question the accuracy of the social media post almost always decrease both the perceived accuracy of the headline claim as well as the self-reported likelihood to like or share that content. Moreover, there is not a statistically significant difference between the treatments. Put succinctly, social corrections “work.” There are, however, two important nuances that accompany this takeaway. First, we use scare quotes because social corrections reduce the perceived accuracy of both false and true content (we expand on this point in Section 4.3). Second, our various manipulations that sought to amplify the strength of the corrective signal (including a justifying link in comment, changing the number of likes for a corrective comment, using two corrective comments) did not noticeably increase the size of the treatment effect.

4.2.2 Ideological Alignment

One contributing factor to falling for misinformation is that people are biased toward accepting claims that are ideologically aligned to their world view or ideological orientation as true, even when the information is false (Flynn et al., Reference Flynn, Nyhan and Reifler2017). Typically, this pattern is attributed to directionally motivated reasoning with differential application of cognitive effort to scrutinize the truth value of claims that are either worldview consistent or inconsistent. Importantly, recent research has suggested that a directionally motivated reasoning approach is observationally equivalent to either other motivations in the broader motivated reasoning framework (Bayes & Druckman, Reference Bayes and Druckman2021) or to a Bayesian approach (Little, Reference Little2021).

Accordingly, we test the role of social corrections in the context of ideologically aligned and not-aligned misinformation. Our surveys included measures that allow us to determine whether information is and is not ideologically aligned: For instance, we measure how respondents view vaccines, science, 5G network technology, Russia’s attack on Ukraine, and so on, before respondents enter the experimental part of the survey. This way, we can identify when a social media post is ideologically aligned and when a correction is not aligned for a respondent. In line with the literature, we find that individuals are more likely to believe false news posts that are in line with their world views (see Section 2). For instance, individuals who do not trust vaccines are more likely to believe false news stories about vaccines.

In this section we are specifically interested in whether social corrections help individuals to identify false content even when it is not ideologically aligned. Conducting this analysis goes beyond our preregistration (see Stoeckel et al. (Reference Stoeckel, Stöckli and Ceka2024) for results executed as preregistered). We account for ideological alignment in two ways. First, our results control for ideological alignment, using a dummy that indicates for each observation whether its content is ideologically aligned for a respondent. Thus, the results discussed above and in Figure 9 control for ideological alignment.

Second, we interact all treatment effects with the dummy variables that indicate whether content is ideologically aligned. The respective marginal effect plots (Figure 10) show treatment effects for the case of ideologically aligned and not-aligned content. Consistent with the emerging literature questioning the ubiquity of heterogenous treatment effects due to directionally motivated reasoning (Wood & Porter, Reference Wood and Porter2019; Carey et al., Reference Carey, Guess and Loewen2022; Coppock, Reference Coppock2023), we find the effects of social corrections are not moderated by prior attitudes in most of our conditions across the three countries. The only exception is the low amplification correction condition in Italy (where the correction is not statistically significant in the case of information that is not ideologically aligned). We conclude that the effect of social corrections does not depend on the ideological alignment of the topic. While directionally motivated reasoning might explain different levels of belief in the first place, it does not seem to affect the extent to which social corrections change underlying beliefs.

Figure 10 Marginal effects of corrections by ideological alignment and by country (false news).

Note: Results based on an interaction of treatments with ideological alignment dummies. Model specifications are similar to those of the main models shown in Figure 9. Multilevel regressions run separately for each country. Dots show point estimates of marginal effects and lines show a 95 percent confidence interval. Negative marginal effects imply that a respective treatment condition decreases perceived accuracy. This means that a respective (false news) post is perceived as less accurate in the treatment condition compared to a control condition which does not include a correction.

4.3 Social Miscorrections Mislead

User comments seem to help against false and misleading information shared on social media. However, if these social corrections are effective against false or misleading content and if the effects are not moderated by prior attitudes, might there also be a dark side? Can (inaccurate) user corrections decrease the perceived accuracy of true claims from mainstream news sources?

As shown in Section 2, a significant share of people also mistake true news for false news – suggesting there is scope for user comments to nudge people toward not believing accurate claims. Indeed, we find substantial variation across the stimuli used in our fieldwork: While one of our true news posts was only considered inaccurate by 15 percent of respondents, three posts were considered inaccurate by more than half of the sample that saw them. What happens if other social media users post comments which say that these true news stories are in fact inaccurate? We call these misleading comments miscorrections and examine their effect on accuracy perceptions as well as engagement intent (liking and sharing).

Examining the role of comments in the context of true news is important because far more news shared on social media is true (Acerbi et al., Reference Acerbi, Altay and Mercier2022). Some studies already point toward the problematic effects of miscorrections in the context of true news. For instance, Anspach and Carlson (Reference Anspach and Carlson2020) examine the role of miscorrections in the context of social media posts about Donald Trump, and find that miscorrective comments are more likely to be remembered. Based on a North American student sample, Vraga and Bode (Reference Vraga and Bode2022) find miscorrections to decrease perceived accuracy of true news in the context of health information.

If miscorrections leave an imprint on social media users, misinformation comes not only in the form of false news but also results from factually incorrect commentary from other users. Over the longer term, miscorrections might contribute to a more general reduction in trust in information, making it more difficult for people to decide what is true and what is false.

We test the effect of miscorrections using a design that parallels our examination of social corrections on false content. Our experiments use the true news material shown in Section 2 and test how miscorrections affect perceived accuracy as well as intentions to like and share a post. In our fieldwork in the UK and in Italy, the different experimental conditions for miscorrections mirror those for corrections. In the UK, we use a control condition without any corrective comments plus three miscorrection conditions: a low amplification condition with two user comments (one being a miscorrection with a small number of likes), a high amplification condition where more than 100 users like the miscorrective comment, and a miscorrection that additionally includes a link to a fact-checking website. In Italy, amplification means a greater number of miscorrective comments, rather than more likes. The low amplification treatment condition includes just one miscorrective comment (besides one other non-corrective comment). The high amplification conditions include three miscorrections from users (among a fourth comment). The last condition includes again a link to a website that seemingly bolsters the miscorrective comment of a user.

The results show that miscorrections affect accuracy perceptions of true news (Figure 9). All but one of our treatment conditions reduce the perceived accuracy of true news compared to the control condition (the lone nonsignificant condition was the low amplification condition in the UK, p < 0.07). Social miscorrections also decrease the intentions of respondents to like and share the true social media posts that they see, albeit the pattern is more mixed (while all treatment conditions are statistically significant in Italy, only one of the three treatment conditions is statistically significant in the UK).

Turning now to differences across our treatments and different forms of amplifying the miscorrective message of the comment, results are strikingly similar to those we found for corrections of false news. For instance, miscorrections that include a link to a website which seemingly bolsters the comment are no stronger than comments without a link. Our Italian fieldwork also shows that additional miscorrections do not have a stronger effect than just a single miscorrective comment. These findings are important to note as, in the case of true news, it shows just how easily perceptions of true news can be affected by user commentary.

We also conduct an additional analysis (not preregistered) that tests whether the effect of miscorrections depends on the ideological alignment of the news content (Figure 11). We find that miscorrections affect accuracy perceptions of true news irrespective of ideological alignment in Germany. In Italy, we find that miscorrections affect accuracy perceptions of ideological aligned true information consistently, but they do not consistently affect perceptions of true posts that are not ideologically aligned. In contrast, in the UK, miscorrection consistently affects accuracy perceptions of posts that are not ideologically aligned, but they do not affect perceptions of ideologically aligned true information. One should note that content of posts differs between countries. Hence, rather than taking these results as manifestations of country differences, they can result from the particular content that was used. In sum, miscorrections have the ability to affect accuracy perceptions of true news and this can happen in the context of information that is ideologically aligned and when information is not ideologically aligned. Future research could examine the precise role of ideological alignment in this context.

Figure 11 Marginal effects of miscorrections by ideological alignment and by country (true news).

Note: Results based on an interaction of treatments with ideological alignment dummies. Model specifications are similar to those of the main models shown in Figure 9. Multilevel regressions run separately for each country with observations nested within individuals. Dots show point estimates of marginal effects and lines show a 95 percent confidence interval. Negative marginal effects imply that a respective treatment condition decreases perceived accuracy. This means that a respective (true news) post is perceived as less accurate in the treatment condition compared to a control condition that does not include a miscorrection.

4.4 Role of Source Cues

Is it possible that some media organizations have sufficient standing and credibility that they are not subject to miscorrection effects? Clear source cues from respected media institutions have the potential to protect against miscorrections, just as source cues from untrustworthy sites may do the opposite. Indeed, how social media platforms choose to display content from news sites is an important decision. (X, formerly Twitter, made the decision to stop displaying headlines in people’s feeds to links from news sites.) To help answer this question, we examine whether the inclusion or absence of a source cue or news outlet logo affects the perceived accuracy of the headline claim.

Accurate news stories might be posted by social media users without showing or linking to a source. However, legitimate news stories are often broken first by professional media outlets. When social media users share this content, other people might thus see a source cue – that is, a logo or name of a news outlet. News from media outlets could be inaccurate as well, but a logo could be a strong signal indicating that a piece of information has been produced or vetted by professionals (assuming that the logo has not been added in a malicious way). Our results show that miscorrections decrease the perceived accuracy of true news posts as well as the intention to like or share a true news post even when source cues are displayed (Figure 9, bottom panel, parameter estimates for “Germany cue”).

The literature is mixed when it comes to the role of source cues. Several studies find evidence of source credibility having a significant effect on accuracy perceptions (Swire et al., Reference Swire, Ecker and Lewandowsky2017; Nadarevic et al., Reference Nadarevic, Reber, Helmecke and Köse2020; Traberg & van der Linden, Reference Traberg and van der Linden2022), but the results are inconsistent (e.g., Clayton et al., Reference Clayton, Davis, Hinckley and Horiuchi2019; Dias, Pennycook & Rand, Reference Pennycook and Rand2020; Baum & Abdel Rahman, Reference Baum and Abdel Rahman2021).

In the true news part of the German fieldwork, we deviate from the pattern used in other countries and for false news in Germany (i.e., we do not show a low amplification condition, a high amplification condition, and a correction with link condition). Respondents are randomly assigned to seeing either the control condition (no comment, no source cue), a high amplification condition with source cue (source cue shown as part of the original post, several miscorrective comments), the same condition without source cue, or a miscorrection with link. The different true news posts include a variety of news outlet labels, most of them being highly reputable mainstream outlets (e.g., tagesschau, FAZ, NDR Info; see Figure 12 and Figure 13 as well as our treatment material on OSF). The results for the treatment condition with these source cues are substantively similar to the ones of the treatment condition that shows the same treatments without source cues. We do not find that a source cue for the true news makes miscorrections less effective. Respondents perceive a lower accuracy when misleading comments tag a post as false whether the original post includes a source cue or not (bottom panel, Figure 9).

Figure 12 Example of a social media post with source cue in the upper left corner (true news; source cue refers to the German newspaper Süddeutsche Zeitung). Translated content (from German): The autumn wave hits hospitals with full force. Top user comment: Now they start again to tell fairy tales. Everyone knows by now that there was no overcrowding in the health system. Bottom comment: That was bound to happen.

Figure 13 Example of a social media post that shows a source cue in the bottom right corner (true news; source cue refers to the German news broadcast tagesschau). Headline: Record melting due to Sahara dust. Top comment: Fearmongering. Bottom comment: And yet there are people who do not believe in climate change.

4.5 Heterogeneous Treatment Effects

In Section 2, we found that anti-expert sentiments, susceptibility to social influence, and cognitive reflection capabilities are related to the ability of individuals to identify false and true news. Do these individual differences play a role in the effectiveness of social corrections? In other words, having observed a direct effect from these variables, we now also examine whether they have a moderating effect on our treatments. We test this possibility by interacting our treatments with these three variables, and we examine the interaction effects for all three outcomes in all three countries: accuracy perceptions, intention to like, and the intention to share false news posts. Our sample is well powered to detect interaction effects (see Section 3 for details on the simulation-based approach we used to estimate the power needed). However, while there are statistically significant interactions between some individual characteristics in some countries, we do not find a consistent pattern that holds across treatments and across countries.

For instance, individuals with more pronounced anti-expert positions are more likely to believe false news (Section 2). While individuals with such positions might be unlikely to take cues from mainstream media outlets or professional fact-checkers that correct falsehoods (Lyons et al., Reference Lyons, Montgomery and Guess2021), they could be more likely to accept corrections from other (non-elite) members of the public. Alternatively, anyone who posts corrections could be seen as someone who seeks to take a position of authority from the perspective of a user who trusts their own position as much (or more so) than cues from others. We find a statistically significant interaction in the UK (corrections have a stronger effect among those scoring lower on our anti-expert sentiment measure). The pattern we find in the UK is neither supported by the results from Italy nor by those from Germany, where we do not find any statistically significant interactions (Figure 14).

Figure 14 Correction effects in the context of false news by levels of anti-expert sentiment scores across the UK, Italy, and Germany. Each graph shows the marginal effect of the respective treatment (i.e. low amplification, high amplification, amplification with link condition) by levels of anti-expert sentiment. The outcome is accuracy perceptions of false news. Grey shading around lines show the 95 percent confidence interval. Bold frames indicate statistically significant interaction effects. Results from multilevel models. Corrections have a stronger effect among individuals with lower levels of anti-expert sentiments in the UK.

In line with the literature, we also find cognitive reflection skills to be related to false news susceptibility in Section 2. There are several potential reasons why individuals who score higher on CRT are better able to spot what is likely to be false, one of them being that higher CRT scores are related to individuals processing information they see more thoroughly and in a more critical way. The same cognitive process could mean that those high in CRT are processing also comments – and thus social corrections – more systematically, which in turn would render corrections more effective. While two interaction are statistically significant, the overall pattern does not support the notion that treatment effects are conditional on CRT scores (Figure 15). This null finding is a very important one as it sheds some light on the role of information processing for the effect of social corrections. Apparently, individuals who score low and high in CRT are affected in a rather similar way by social corrections. We interpret this as evidence that the very reason why social corrections affect users is not because they process the corrections very thoroughly (if that was the case, we would expect effects to be more pronounced among those who are predisposed to process information more thoroughly).

Figure 15 Correction effects in the context of false news by levels of cognitive reflection scores across the UK, Italy, and Germany. Each graph shows the marginal effect of the respective correction effect (i.e. low amplification, high amplification, amplification with link condition) by levels of CRT-score. The outcome is accuracy perceptions of false news. Grey shading around lines show the 95 percent confidence interval. A bold frame indicates a statistically significant interaction effect. Results from multilevel models.

The third individual level characteristic we take into account is susceptibility to social influence. As shown in Section 2, individuals who are more susceptible to social influence are more likely to believe false news and at the same time also more likely to believe true news. Analyzing the interaction between the effect of social corrections and variation in susceptibility to social influence can give us important insights on the mechanism that explains why corrections work in the first place. One set of studies (e.g., Colliander Reference Colliander2019; Boot et al., Reference Boot, Dijkstra and Zwaan2021) points toward a bandwagon effect and, in a similar vein, to conformity as a reason why corrective comments might have an effect on citizens. A key part of that mechanism is that individuals adjust their views or behavior to those of others. Susceptibility to social influence is about individual differences in the extent to which individuals react to other people. Hence, if conformity is a reason why social corrections work, we would assume the corrections to have a stronger effect among individuals who are more susceptible to social influence. What we find is a mixed picture. There is a statistically significant relationship in the UK and for some conditions also in Italy (Figure 16). However, the effect of corrections seems to be stronger among individuals who are less susceptible to social influence (rather than more susceptible). There is no evidence of an interaction in Germany. In sum, we do not find that susceptibility to social influence consistently conditions the effect of social corrections. This casts doubt on the idea that conformity is a reason why social corrections have an impact on user perceptions of accuracy. That is because we would expect individuals who are more likely to adjust their views to others to react more strongly to social corrections, but this is not what we find.

Figure 16 Correction effects in the context of false news by levels of susceptibility to social influence across the UK, Italy, and Germany. Each graph shows the marginal effect of the respective correction effect (i.e. low amplification, high amplification, amplification with link condition) by levels of susceptibility to social influence. The outcome is accuracy perceptions of false news. The top panel refers to the UK, the middle one to Italy, and the bottom one to Germany. Grey shading around lines show the 95 percent confidence interval. A bold frame indicates statistically significant interaction effects. Results from multilevel models. Corrections have a stronger effect among individuals with lower susceptibility to social influence in the UK and in two of the conditions in Italy.

We also tested whether individual differences in anti-expert sentiments, cognitive reflection skills, and susceptibility to social influence condition the influence of miscorrections. As was the case with corrections, we do not observe a consistent pattern of interactions (we observe mostly null effects). Miscorrective commentary seems to affect individuals in a similar fashion irrespective of these differences.

4.6 Limitations

We want to highlight several important limitations. First, and most importantly, we do not measure the effect of corrective comments in real life social media feeds. Though, it is not clear that it would be possible to run a study that embedded randomized user corrections into people’s social media feeds. While there are several factors that give us lots of confidence in the robustness of findings – the large number of observations, replicating findings across countries, and getting consistent results across different stimuli and treatments – we are always subject to the fact that data were recorded in the artificial context of a study. While the social media posts that we showed respondents follow the design of Twitter, Instagram, or Facebook, respondents did not see this information in a context of a feed with information from friends as well as strangers. The fact that the social media posts were not created or posted by friends should not, on its own, be a major concern. First, this is a very common experience on social media sites, where users often are exposed to and interact with those outside their offline network. As mentioned above, we also believe that this leads to more conservative effect estimates, because users would likely be expected to pay more attention to corrections from friends.

Another limitation is that participants were rating the accuracy of shorter statements rather than full articles and they could not access a full article if they would have wanted to (e.g., to better assess the accuracy of the headline). Here, we are following not only a common approach in social media misinformation research, where only showing headlines is a common procedure (Pennycook et al., 2021). Social media users are in fact often – or even most of the time – making decisions, for instance to share a post, just based on the headline. From that point of view, one might argue that reducing the extent to which people are sharing content just based on a headline should be the goal of an intervention. Additionally, individuals who would like to see a full story before feeling comfortable enough to assess the accuracy of a piece of information or before sharing a post might in fact not be the group of people most susceptible to misinformation (e.g., people scoring higher on cognitive reflection – i.e., those who are predisposed to reflect more before making a decision – are less susceptible to misinformation).

Another limitation is that asking respondents about accuracy, liking, and sharing as we do creates an artificial situation that does not fully resemble real social media feeds. Asking respondents about these three outcomes might affect both accuracy perceptions and sharing intentions (see Epstein et al., Reference Epstein, Sirlin, Arechar, Pennycook and Rand2023, and our discussion about that point in Section 3). We have confidence in our results because of the robust pattern across posts and countries. Yet, we kept asking these outcomes in all fieldwork, and hence all measures are potentially affected by this. However, sharing is a salient feature of social media and therefore conducting our study without asking about sharing intentions would not have increased external validity from our perspective (and removing accuracy perception was not an option, given that this is our main outcome). We appreciate, however, that this is an important issue and limitation that deserves attention in future work.

Finally, our results might be driven by the particular material that we used: the content of the post, the way corrections were written, and the way miscorrections were written. As Section 2 shows nicely, content matters – some content is more believable than other content. In a similar vein, corrections or miscorrections might just be more effective in some contexts than in others. However, we show robustness by covering a particularly wide set of topics as well as a wide set of (mis)corrective comments. Corrections (and miscorrections) seem to work on average.

4.7 Conclusion

In this section, we show that social corrections of false news seem to help people identify inaccurate information from social media. One of the main reasons why individuals believe that something is true even though it is false is because a piece of information is ideologically aligned – that is, it is in line with ideological orientations or prior attitudes. We find that, on average, the effects in our experiments are not conditional on ideological alignment. However, while this is an important result, its effect should not be taken out of proportion: Motivated reasoning and the role of partisanship are strong forces, and the precise effect of them – relative to social corrections – is likely to depend much on the specific piece of misinformation and how strong someone’s opinion is on an issue or whether the topic is central to someone’s identity. Put differently, we would not expect a vaccine opponent to see a (false) social media post claiming that vaccines are ineffective differently just because other users are posting corrective information.

We tested several versions of corrective comments as well as how variation in likes and the number of corrective comments affect readers. We find that the mere presence of a correction is more important than the number of comments or likes (that a corrective comment received). Just because likes and fact checks do not strengthen the effects of corrections, one should not dismiss them as irrelevant – likes may matter in order for a corrective comment to be salient and fact checks could help writers to be sure that they are in fact correcting accurately. We return to their important role in a part on best practices in Section 6 of this Element.

Our findings also shed light on the mechanism that drives the effect of social corrections. Our experiment focuses on the role of corrections posted by strangers, rather than corrections from friends or family. Signals from friends and family might likely be more effective and thus what we observe is likely a conservative estimate of correction effects. Yet, the reason why corrections work might not in fact result from commenters who are perceived as credible or because users might want to conform to the views of the general public. Rather, information processing might be much less systematic, and because information is processed with few cognitive resources, recency effects can dominate. Meaning, a correction or refutation would work merely because it casts doubt on a piece of information that has just been seen – that is, because doubts are raised briefly after an original piece of information has been processed.

Other patterns of our results point in a similar direction. For instance, if source credibility would be very important, corrections with a link or those reinforced by several individuals (by likes or by people who agree) could be expected to have a stronger effect, but this is not what we find. If conformity considerations (i.e., bandwagon effects) were central, then individuals who are more susceptible to social influence could be expected to respond more to the signals from others. We do not find this interaction effect either. We also do not find corrections to have a stronger effect among those who are predisposed to process information in a more reflective manner, even though we find cognitive reflection skills to help people generally to identify false information (see Section 2). Thus, in sum, our results do not support the idea that conformity, credibility considerations, or a particularly thorough processing of comments is the mechanism through which corrective comments affect users. Instead, individuals are more likely to process information in a less thorough way, which is why the piece of information considered after an initial post – a set of refutations – might affect user perceptions. For this very reason, though, miscorrections appear to be equally effective at casting doubt.

5.1 What Do Social Media Users Do When They See Misinformation?

Taking time to correct others online might impose personal costs since it requires effort (Druckman, Reference Druckman2022). It also may hurt relationships (Bode, Reference Bode2016; Yang et al., Reference Yang, Barnidge and Rojas2017; Gurgun et al., Reference Gurgun, Cemiloglu, Arden-Close, Phalp, Nakov and Ali2023) and some may prefer to avoid conflict (Chadwick et al., Reference Chadwick, Vaccari and Hall2023). Similarly, correcting others may go against behavioral norms (Gurgun et al., Reference Gurgun, Cemiloglu, Arden-Close, Phalp, Nakov and Ali2023) or conflict with one’s personality characteristics (e.g., high self-esteem is linked to a lower likelihood of denouncing false content shared by strangers; Johnson & Kromka, Reference Johnson and Kromka2023). As such, there are numerous reasons as to why many members of the public may avoid corrective behavior.

Toward this end, our survey asked respondents about their typical reactions when they encounter information on social media that they find inaccurate. Based on the self-reports in the survey, the most common behavior is for people to disregard such information (61 percent report that they ignore false news often or most of the time). The second most frequent reaction for users is to unfollow or unfriend the person who posted misinformation (48 percent do this often or most of the time), followed by flagging misinformation to the platform (26 percent). Approximately one in five (21 percent) respondents say that they react to misinformation by writing corrective comments often or most of the time. In sum, a majority say they take an avoidance tactic, though a sizable minority are frequent correctors.

5.2 Who Is Responsible for Fixing the Problem?

Aside from interpersonal reasons, individuals might refrain from commenting because of their views about the problem of misinformation itself – that is, who is responsible, and whether social corrections are a valid means to address it. Some may view individual action as less preferable to top-down action by platforms or governments (Cheng & Luo, Reference Cheng and Luo2020; Tully et al., Reference Tully, Madrid-Morales, Wasserman, Gondwe and Ireri2022). Others may eschew corrective action if they think it may inadvertently amplify misinformation (Swire et al., Reference Swire, Ecker and Lewandowsky2017).

Our survey finds that users have a clear ranking in mind when it comes to the role that social media companies, users, and governments should play when it comes to fighting misinformation. Our German sample seems to primarily allocate responsibility for dealing with misinformation to social media companies, rather than government. Over half (63 percent) of respondents say that companies play a big role in this regard, while about a third (37 percent) believe that companies play a small or no role. When it comes to users, 56 percent say they should play a key role, versus just 44 percent who say they should play a small or no role. Finally, according to almost half of respondents (48 percent), governments should take on a major role to combat misinformation. This clearly outweighs the approximately 14 percent who oppose any government interference in this domain.

Reacting to misinformation poses a potential quandary for social media users inclined to take action. Social media algorithms typically upweight content that receives significant engagement. By commenting or interacting with untrustworthy content (with the purpose of trying to correct it), users might inadvertently make misinformation more salient to others. Thus, people might avoid interacting with misinformation specifically to stop spreading it.

The degree of the correction/salience tradeoff depends on the algorithm of a social media platform – what is promoted and how much interaction triggers this (which individual users are extremely unlikely to have direct knowledge about). However, people might have a preference regarding the world they would prefer to live in, which gives us an insight into their thinking about the trade-off between corrections and lack of salience. Users could prefer a situation where misinformation is corrected even at the price of salience, or one where it is not corrected and misinformation less salient. As can be read from Table 2, 73 percent of our survey respondents prefer misinformation to be corrected even if corrections contribute to salience.

5.3 Views on the Efficacy and Value of Corrections

Even though most seem to accept the salience trade-off raised by corrective actions, social corrections themselves could be seen as ineffective or futile (Lyons, Reference Lyons2022b; Gurgun et al., Reference Gurgun, Cemiloglu, Arden-Close, Phalp, Nakov and Ali2023). However, most people do not seem to view corrective comments as ineffective. We find that a majority (73 percent) of respondents believe that corrections of false news on social media help against the spread of misinformation. In a similar vein, most people (76 percent) think that the collective efforts of users could make a difference in the fight against misinformation.

A large share of respondents also believes that they personally are able to correct inaccurate information (73 percent). We can also map confidence against correction frequency, which can be seen in Table 3. Only about 23 percent of those with confidence in their ability to correct misinformation actually do so often or most of the time when they see misinformation. This percentage may seem small, but the baseline in the sample for writing corrections is generally low. Therefore, it is potentially concerning that among those who do not feel confident to correct misinformation, 15 percent still state that they react to misinformation by writing a corrective comment frequently or most of the time.

Along these lines, it is also important to consider views toward miscorrections and their perceived effects on oneself and others. As we show in Section 4, comments which wrongly attack the veracity of true headlines decrease perceived accuracy, complicating decision-making about what is true news and what is false news. Our survey investigates whether this is a concern people have about comments that other users post on social media. Indeed, we find that 76 percent of respondents believe that user comments can make it more difficult for people to identify true and false news. Consistent with other research (Sun et al., Reference Sun, Tsai, Finger, Chen and Yeh2008), people are, however, more worried about others being affected by inaccurate corrections than about themselves. Only 50 percent of respondents also fear that comments might make it more difficult for themselves to identify false and true news. Prior work suggests this perceptual gap reflects at least some degree of overconfidence (Lyons et al., Reference Lyons, Montgomery and Guess2021; Lyons Reference Lyons2022b).

This can pose a more general problem for social corrections. The high number of people who say that they can largely correct misinformation should be treated with caution: While individuals might have confidence in their ability to correct false news, at least a share of them might overestimate their ability to identify true and false news. Lyons et al. (Reference Lyons, Montgomery and Guess2021) show that overconfidence is in fact very common. Thus, when people try to correct something they assume to be false even though it is in fact true, they might contribute to the spread of misinformation. For this reason, we now turn to an analysis of commenters’ personal characteristics.

5.4 Characteristics of Users Who Write Comments

It is essential to understand potential asymmetries in who tends to comment most often, given the double-edged nature of comments. In our study (Section 4), we find miscorrections to be just as influential as corrections, so examining the relative frequency of each is important for future work (see Metzger et al., Reference Metzger, Flanagin and Mena2021). One way we can begin to shed light on this, though indirectly, is to examine the background characteristics of habitual commenters. This may illuminate whether most comments tend to be done by skilled individuals or those less equipped to “get it right” (Trepte & Scherer, Reference Trepte and Scherer2010; Weeks et al., Reference Weeks, Ardèvol-Abreu and Gil de Zúñiga2017).

Using our survey where respondents self-report on whether they write corrective comments, we can identify the characteristics of individuals who engage in social corrections more often. To do this, we use data from the survey that we fielded in Germany in February 2023. This survey asks respondents how they typically react when they come across false information on social media. We are particularly interested in the frequency of certain reactions. Thus, for each of six reactions, we ask how frequently this occurs (never, rarely, often, most of the time). The reactions include writing comments that correct inaccurate information, as well as ignoring wrong information, notifying the platform, unfriending the person who posted the information, and so on. For our analysis in this section, we use the data about the frequency with which respondents react to (seemingly) wrong information by posting a correcting comment. As mentioned in Section 5.1, only 21 percent of respondents say that they write comments often or most of the time.

To identify the characteristics that are associated with more frequent commenting, we run an OLS regression. The range of the outcome is 1 to 4 (self-reported frequency: never, rarely, often, most of the time). We regress the frequency with which respondents claim to write comments on a set of demographic variables, measures for political orientations, and attitudinal variables. Figure 17 shows the results.

Figure 17 Determinants of correction frequency in response to false social media content.

Note: Measures used above are self-reports. Survey question: “What do you do when you see false information on social media?” For each of six reactions, we ask how frequently it occurs. The dependent variable of the model plotted above is how frequently respondents say that they react by doing the following: “Write a comment and say what was wrong” (never, rarely, often, most of the time). OLS regression, N = 3,775. Dots: point estimates, lines: 95 percent confidence interval. SPD: Social Democrats, Die Grünen: Green Party, AfD: Alternative for Germany, FDP: Liberal Party, Die Linke: Left Party, CSU: Bavarian part of Conservative Party. Reference groups: age under 35, female, education: low, party: Conservative (CDU), respondent prefers low salience over corrections of false information. Opinion on science: negative, rather negative, rather positive, positive. CRT: higher values represent higher cognitive reflection test (CRT) capabilities. Anti-expert sentiment score: higher values indicate greater anti-expert sentiment (index consisting of three items; see online appendix). Statistical significance: *(p < 0.05 = *, p < 0.01 = **, p < 0.001 = ***)

We first find that views toward corrections matter. Individuals who prioritize corrections over a low salience for misinformation write comments more frequently, as do those who believe that comments can help against the spread of misinformation. We also find that users over 35 write comments less often, while education plays no role. Additionally, we find men to write corrections more frequently, in line with Peacock and Van Duyn (Reference Peacock and Van Duyn2023), who found that while women are more likely to read comments, men are more frequently the ones writing them. Individuals with higher cognitive reflection test (CRT) scores write comments less often. This specific finding is particularly interesting because some have interpreted CRT as, at least in part, a measure of (cognitive) effort (Pennycook & Rand, Reference Pennycook and Rand2019b). So, this form of (cognitive) effort does not translate into self-reported effort in another domain. Party affiliation is not related to comment writing frequency, except that those who do not feel close to any party write comments less frequently than all other respondents. Finally, individuals who score higher on an anti-expert measure write comments more frequently, which is a concerning finding given the harm done by miscorrections. Comparing men and women across the different variables, we do not find any significant differences.

We now focus on the confidence that people have in their ability to correct misinformation (Figure 18). Men have more confidence in their ability to correct than women. Yet, gender does not have moderating effects for the other variables (tested in a separate analysis). Both individuals with anti-expert positions and positive views toward science feel more confident in their ability to correct. Put differently, those who see science negatively have less confidence in their ability to correct, though this is of course occurring when holding anti-expert positions constant. As with willingness to write comments, education is not significant. However, unlike willingness to write comments, age and CRT scores are now insignificant. While younger and less cognitively effortful respondents are more likely to actually write comments, they express no greater sense of ability than their older and more cognitively reflective counterparts.

Figure 18 Determinants of subjective ability to correct false social media content.

Note: Survey question: “In most situations, I am able to correct false information.” Response categories include: disagree completely, disagree somewhat, agree somewhat, agree completely. OLS regression, N = 3,757. Dots: point estimates, lines: 95 percent confidence interval. SPD: Social Democrats, Die Grünen: Green Party, AfD: Alternative for Germany, FDP: Liberal Party, Die Linke: Left Party, CSU: Bavarian part of Conservative Party. Reference groups: age under 35, female, education: low, party: Conservative (CDU), respondent prefers low salience over corrections of false information. Opinion on science (negative, rather negative, rather positive, positive). CRT: higher values represent higher cognitive reflection test (CRT) capabilities. Anti-expert sentiment score: higher values indicate greater anti-expert sentiment (index consisting of three items; see online appendix). Statistical significance: *** p < 0.001

It is worth exploring in greater depth the findings regarding education and cognitive reflection on one hand and anti-expert positions on the other. As Table 4 shows, most respondents are confident in their abilities. 70 percent of those in the low education group and the lower tercile of anti-expert views feel confident in their ability, and this rises to 77 and 79 percent in the high education and high anti-expert groups, respectively. Confidence is similarly high across levels of CRT as well. When considering actual self-reports of behavior though – whether they actually engage in corrective commenting – differences emerge. Only about 22 percent of the highly educated post comments often or most of the time, about the same as the average respondent, and among those scoring highest in CRT, only about 16 percent write corrective corrections often or most of the time. Meanwhile we see significant variation across the anti-expert measure. A considerable number – 42 percent – of those most opposed to experts say they post corrections often or most of the time, compared to just about 16 percent of those with the most positive views toward experts.

This sizable imbalance suggests anti-expert positions are overrepresented in comments. This may also suggest that the content produced by the most prolific commenters is less accurate on average, particularly if they consistently contradict expert consensus (Light et al., Reference Light, Fernbach, Rabb, Geana and Sloman2022), though we cannot assess this. This result should be understood in the context of what others have found about those with anti-expert views instead: They are not only less knowledgeable, but also more vocal, and more likely to spread dubious news online (Light et al., Reference Light, Fernbach, Rabb, Geana and Sloman2022; Motta et al., Reference Motta, Callaghan and Sylvester2018; Lyons et al., Reference Lyons, Montgomery and Guess2021), which may lead to misperceptions of the prevalence of their extreme views over time (McKeever et al., Reference McKeever, McKeever, Holton and Li2016).

One takeaway from these analyses may be to target those who are likely to be better equipped to correct misinformation but who otherwise refrain from doing so. For instance, respondents high in cognitive reflection reported engaging in corrections less often. This makes sense because cognitive reflection is a measure designed to gauge one’s tendency to override an incorrect intuition, or “gut” response – a tendency quite opposed to confidently spreading a wrong message unchecked. But as others have consistently found, cognitive reflection is associated with better discernment between true and false news (Pennycook & Rand, Reference Pennycook and Rand2019b, Reference Pennycook and Rand2020) and sharing better quality news in real-world social media settings (Mosleh et al., Reference Mosleh, Pennycook, Arechar and Rand2021). Improving confidence or otherwise incentivizing the cognitively reflective to engage in corrective action is one potential route to improving the quality of online discussions.

5.5 Conclusion

In sum, we find that many social media users say they avoid engaging in corrective actions. Aside from interpersonal costs, one reason appears to be that they assign responsibility to platforms to police such content. They also believe that user comments can further muddy the waters of what’s true and what’s false. Still, they see corrections as valuable, and preferable to inaction even at the cost of increasing visibility to false content. When examining who is more likely to engage in corrective action, we find that those who have negative views of experts are more likely to feel that they can effectively post corrections, and self-report that they do so more often than those who view experts favorably. Taken together with our finding about the effects of miscorrections, this paints a troubling picture of the potential dynamics of online comments.

6.1 Writing Effective Corrections

Identifying false news on social media can be challenging (and can be more challenging for some). The false news that we used in our fieldwork was considered true by at least one out of five respondents. When false claims align with one’s political predispositions, then respondents in our studies were more likely to perceive these headlines as accurate. Similarly (and consistent with prior research), individuals with less developed cognitive reflection skills are less able to identify false news, as are those with far-right or anti-expert positions. Based on the fieldwork from three countries, however, we find that social corrections can help people to identify false news in spite of these individual differences. Our results allow us to identify a set of best practices for how users could approach corrections.

It is first worth reflecting on the audience for corrections: The target audience for a correction is other users who see the original post. Thus, corrective comments can be useful even if they do not convince the original poster that their post contains false news. Posts can be seen by thousands of people – helping this audience navigate what is true and false in their social media feed is perhaps the most useful aspect of social corrections. This use case is different from debating with a potential stranger about the accuracy of content they have shared online. The person sharing inaccurate content might have truly overlooked that the information they shared is inaccurate, or they might have shared it in spite of the fact that they know it is flawed (Munyaka et al., Reference Munyaka, Hargittai and Redmiles2022; Littrell et al., Reference Littrell, Klofstad and Diekman2023). In these situations, corrections might positively affect the account sharing misinformation. However, it might also be the case that challenging the accuracy of a claim has no effect on the person who posted or shared it in the first instance. Being unable to change the original poster’s mind does not mean that corrective comments are in vain; rather, as our research and the results from others have shown, comments can help other people to identify false information. That is, social corrections are not primarily thought to change worldviews or established false beliefs of the posters, but instead are mainly to help people decide more carefully on the accuracy of information that shows up in their feed. This more nuanced view of the audience for corrections allows us to appropriately think about how to maximize the effectiveness of corrections.

We wanted to explore different ways that corrections could be made more powerful or send a stronger signal. One specific component we examined was whether corrective posts that linked to a fact-check would have a stronger effect. We do not find comments with a fact checking link to be more effective than simple corrections. This finding somewhat lowers the bar of effort for those who want to write corrective comments. Users considering writing a corrective comment do not need to search for a fact-check to include in their comment to bolster the authority and veracity of their position. Yet, we do not take the fact that inclusion of a fact-checking link has minimal effect on increasing the effectiveness of a correction to mean that fact-checking as a journalistic enterprise is pointless. Providing such links improves the quality of the information in the comment for those who might encounter it, as they can more easily check the corrective claim. (Our experimental design does not provide users with the opportunity to click through to links when they are included in comments.) Providing a link is likely associated with increased deliberation exerted by the commenter and the odds that they themselves have avoided posting a miscorrection. As human beings, we are all subject to the same biases: Those who comment are also more likely to reject information that challenges their worldview, and they are more likely to post a correction in line with their worldview but maybe not in line with the facts. Consulting a fact-checking website can serve as an important buffer against this bias. The same strategies that help people identify false news in the first place (Wineburg & McGrew, Reference Wineburg and McGrew2017; Guess et al., Reference Guess, Lerner and Lyons2020) should help people to make sure that they are not contributing to spreading false news.

Our results also reveal that corrections with a higher number of likes are not more effective in reducing perceived accuracy than comments with a few likes. When we showed respondents more corrective comments rather than just one, we did not observe stronger effects, either. However, this does not imply that likes (or additional comments) are irrelevant. We designed our treatments so that respondents would see the corrective comments, for example, by showing them right below a post. While the number of likes did not matter in how people responded to our treatments, the number of likes will play a factor in whether a comment is displayed in real world social media feeds. When corrective comments get more likes, the odds that this comment will be a featured one that appears automatically below a post is increased. In a competitive landscape in the attention economy, more likes may change the probability that a post is shown to users, making those comments more effective in real world settings, even if the number of likes has no direct effect on the persuasiveness of the comment (conditional on exposure). Liking a corrective comment by another user can help to ensure that it is visible among the top comments of a post. (Indeed, this is another way that fact-checking links may be helpful – they look more persuasive and garner additional likes as a result, increasing the chances the comment is easily seen in someone’s feed.)

In a similar vein, additional comments could make corrective comments salient. As people reply to a comment to show their agreement, this may upweight the original comment. Our experimental manipulations suggest that once corrective comments are salient, additional likes or comments do not increase the corrective signal strength. But, additional reactions or comments might change the likelihood that the corrective comment is shown to other users (even if the corrective strength stays constant). These processes can be dynamic: A corrective comment that is salient in one moment may not be the top comment a few minutes, hours, or days later. Organic responses to content – especially if something goes even semi-viral – could quickly change the visibility of corrective comments.

One concern that people might have about posting corrections is that doing so contributes to the spread of misinformation as this engagement feeds into the social media algorithm thereby pushing the false content into more people’s feeds. While we cannot answer questions about the algorithmic mechanics of how engagement changes the likelihood of “treating” new people with false content in their feeds, we can address questions about how people feel about this potential trade-off when engaging in corrective activity. Based on the results from Section 5, three out of four people prefer inaccurate content to be corrected, even if it makes misinformation more salient. Thus, more people are likely to appreciate corrective comments than to see it negatively. If the visibility of a comment – or any engagement with a false news post – is an obstacle for someone who is eager to get involved, they could also flag a post to the social media platform directly. Most platforms offer this option to flag content, though it requires more effort on some platforms than others (and default options for labelling problematic content do not always include an option for “false”). Using this approach also comes with the advantage of not contributing to the salience of a piece of false news. The downside is, however, that the consequence of this reporting is not clear either. While platforms claim to be interested in combating false information, traffic is essential for their business model. Social media platform policies do not necessarily prohibit false content, and there are often issue-by-issue policies. Moreover, the threshold and time frame for deletions as a result of content flagging is unknown. Finally, content-flagging by users is not without peril as users could flag content erroneously, and how platforms deal with these is not clear cut.

Commenting is a simple behavior, and one that shows promise as a strategy to address false claims in online social media spaces. If people were to inculcate the habit of writing social corrections as part of good social media hygiene, it would have the potential to improve online discourse. Just as there is evidence that casting a ballot in one election can increase the odds one turns out to the polls in the next election (e.g., Gerber et al., Reference Gerber, Green and Shachar2003), getting the bulk of social media users who never leave corrective comments to try it on an (initially) limited basis could result in significant aggregate improvements to the information environment long term.

6.2 Outlook: AI and Social Correction

Artificial intelligence (AI) is powerful and can be abused to disinform others on social media. However, these technologies may also enhance our corrective capacities. In this final section, we explore these issues as a future frontier for social corrections.

Artificial intelligence – especially generative AI like large language models (LLMs) is receiving considerable research and public interest (Vicari & Komendatova, Reference Vicari and Komendatova2023). However, there is also potential for this technology to be abused to create swathes of misinformation. For instance, when asked the right questions, language-learning models like ChatGPT are able to mass-produce false content. It is even possible for such misinformation to spread accidentally. The responses by ChatGPT are not researched, and when asked for sources the tool may freely invent them and possibly fool the reader (De Angelis et al., Reference De Angelis, Baglivo and Arzilli2023; Haupt & Marks, Reference 76Haupt and Marks2023). Thus, individuals who do not know this or who misuse the AI can contribute to the spread of misinformation.

Furthermore, deepfakes or fake images pose an issue concerning misinformation. The former are videos of people that have been modified by machine-learning tools (Paris & Donovan, Reference Paris and Donovan2019). A famous example of a fake image that received a lot of media attention would be the picture of Pope Francis in a white puffer jacket (Huang, Reference Huang2023). Such false material can be particularly problematic since it is often difficult to recognize as AI-generated (Partadiredja et al., Reference Partadiredja, Serrano and Ljubenkov2020). To combat this problem, social media companies like Meta have implemented AI tools themselves to analyze images and point out potentially fake ones to professional fact-checkers (Woodford, Reference Woodford2018). Nevertheless, they are unlikely to catch all AI-generated content. As Groh et al. (Reference Groh, Epstein, Firestone and Picard2022) find, artificial intelligence tools are not yet able to correctly identify all deepfakes they are presented with. Hence, deepfakes and other fake AI images may continue to be used to promote false news.

Relatedly, large numbers of bots on social media platforms may help spread misinformation. They often serve malicious purposes, like trying to sway public opinion through the dissemination of false information (Lazer et al., Reference Lazer, Baum and Benkler2018; Bastos & Mercea, Reference Bastos and Mercea2019). Although social bots have existed since at least 2008, their popularity has steadily increased due to their cost-effectiveness. According to a 2018 study by the Pew Research Centre, about 66 percent of all content on Twitter at that time stemmed from bots (Wojcik et al., Reference Wojcik, Messing and Smith2018). Some bots are even equipped with AI technology, making them more able to respond to different content like a human would (Hajli et al., Reference Hajli, Saeed, Tajvidi and Shirazi2022), which may make distinguishing between real and fake accounts even more challenging.

Even though the development of AI has caused several difficulties concerning misinformation, it may also help to address them. Some literature suggests that AI could be used to detect bots, deepfakes, and other misleading content. Shabani and Sohkn (Reference Shabani and Sokhn2018), for example, advocate for the use of an AI that recognizes its weaknesses when detecting misinformation and asks for human input when necessary. However, there may also be other approaches, as suggested by Sager et al. (Reference Sager, Kashyap, Tamminga, Ravoori, Callison-Burch and Lipoff2021), for instance. In an experiment they used bots with natural language processing to post corrective comments on a Reddit forum on dermatology, with test accuracies from 95 to 100 percent. According to another idea by Lu et al. (Reference Lu, Li, Wang and Yin2022), an AI may come up with credibility ratings for articles. This approach has been found to influence readers and, hence, may be another viable option to combat false news content – given that the AI is sufficiently accurate.

However, these solutions are likely not as straightforward as they appear. For instance, most of these approaches require a certain level of trust in AI, which may not always be present. Although several factors can determine such attitudes (Molina & Sundar, Reference Molina and Sundar2022), Wojcieszak et al. (Reference Wojcieszak, Thakur, Ferreira Gonçalves, Casas, Menchen-Trevino and Boon2021) find that individuals generally have more trust in content moderation by humans than artificial intelligence. Moreover, an AI tool may be biased due to the data it was given to train it (Demartini, Mizzaro & Spina, Reference Demartini, Mizzaro and Spina2020). As Groh et al. (Reference Groh, Epstein, Firestone and Picard2022) explain, this is one of the reasons why humans are more suited to moderate content. Lastly, the delicacy of freedom of speech may require a human to be responsible for determining what false or misleading content can be blocked. While some have argued for a more stringent filtering of content with AI before it is posted, this raises additional concerns about censorship (Dias Oliva, Reference Dias Oliva2020; Llansó, Reference Llansó2020). Hence, many measures using AI are problematic due to the possible bias of AI and the implications on freedom of speech.

AI is already being employed by companies like Meta to detect misinformation and flag it for fact-checkers. Such uses may aid the fight against misinformation as it is a more time-efficient way to look for misleading content (Woodford, Reference Woodford2018; Gillespie, Reference Gillespie2020). However, extending the use of AI to delete content has many negative implications on matters like freedom of speech and cannot be recommended (Dias Oliva, Reference Dias Oliva2020; Gillespie, Reference Gillespie2020; Llansó, Reference Llansó2020). Therefore, AI may only be a tool to tackle misinformation when it is used with care and to aid fact-checkers rather than to take over their jobs.

We show that user-driven approaches to debunking misinformation can be effective, but we also reveal important limitations of such efforts. Our research was conducted in an environment where fact-checking organizations played a significant role, even though corrections that included links to the fact-checkers’ websites did not significantly increase their effectiveness. Nevertheless, the availability of professional fact-checkers’ expertise remains crucial for users to verify the accuracy of corrections, thereby reducing the risk of miscorrections. Fact-checking organizations provide specialized knowledge that users do not necessarily possess, and they make high-quality debunks readily accessible. In this way, professional fact-checkers are an integral component of effective user-driven corrections, rather than an alternative. User-driven systems face critical limitations, particularly when deployed as substitutes for professional fact-checking rather than as complementary tools. Community Note systems usually flag content based on user consensus, which is determined algorithmically. Without the input of fact-checking organizations, the burden of fact-checking content is shifted to unpaid users while the availability of reliable debunking resources is simultaneously decreased. This risks undermining the overall quality of information on social media platforms. Overall, community-based approaches are likely most effective when combined with the expertise of professional fact-checkers.

In sum, we have offered some final arguments about the value of social corrections and potential applications going forward. We stress that social correction in aggregate may outweigh the insidious side effects of miscorrections, particular if the best commenters are incentivized or if comments are filtered in some way. In addition to some practice advice for commenters, we have also highlighted the perils and promise that an AI-infused future holds for the act of social correction.