Solutions and Challenges for Addressing Misinformation

Abstract

Research on mitigating the effects of misinformation has contributed to the development of multiple feasible interventions designed to reduce belief in, and sharing of, falsehoods. The authors review these interventions and discuss challenges and open questions for future research. First, they provide an overview of content-neutral and content-based interventions. Next, they discuss two practical challenges to deploying and assessing these interventions in the field: scalability and pushback against content moderation efforts due to perceived political bias. Finally, they highlight several open theoretical questions and common pitfalls of research on misinformation. In particular, they argue for critical evaluation of how interventions may be effective across different types of misinformative content, different key subpopulations, and different media and environmental contexts.

Fostering well-informed individuals and functional information ecosystems is an important goal for academics and policy makers. Misinformation and related forms of harmfully misleading content threaten this goal. Indeed, for the past two years, experts have identified misinformation as one of the top global risks (World Economic Forum 2024, 2025). Social scientists across disciplines have accordingly produced a large body of work investigating the belief in, and spread of, falsehoods and “fake news” (Lazer et al. 2018), which largely proliferated from concerns about misinformation in the wake of Brexit and the 2016 U.S. presidential election.

The potential harms of misinformation remain of great concern. Misperceptions of vital political importance, such as election fraud in the 2020 U.S. presidential election, are widespread (Pennycook and Rand 2021b; PRRI 2024), and researchers have posited that false claims about such political matters pose threats to democratic norms and institutions (Ecker et al. 2024). False and misleading claims about essential public health measures, such as the COVID-19 vaccine, have been demonstrated to reduce individuals’ intentions to receive vaccinations for themselves and their children (Allen, Watts, and Rand 2024; Loomba et al. 2021). Deceptive and false claims in the marketplace can also cost consumers by harmfully misleading them about products and contributing to deteriorating consumer trust (Di Domenico and Ding 2023; Fong, Guo, and Rao 2024; Rao 2022). Furthermore, misinformation exposure and risks are not uniform, but rather may be concentrated where most potentially harmful. Exposure to low-quality news sources is largely concentrated among a small minority of the population in the U.S. information context, primarily on the political right (González-Bailón et al. 2023; Grinberg et al. 2019; Guess, Nagler, and Tucker 2019; Guess, Nyhan, and Reifler 2018), where the risk of translation from misinformation to real-world harm could be greatest (Budak et al. 2024). And threats of tangible harm stemming from falsehoods may be even greater in the Global South, where unsubstantiated rumors about minority groups have been linked to support for vigilante violence (Badrinathan, Chauchard, and Siddiqui 2025).

In the current work, we review potential solutions for mitigating misinformation and harmfully misleading content. First, we discuss a wide range of interventions that have been proposed and tested to reduce the belief in and sharing of misinformation. We broadly categorize these interventions as content-neutral and content-based approaches. We detail strategies and designs for enacting these interventions at scale, as well as contemporary challenges facing content moderation implementation and investigation. We then identify several open questions and common pitfalls in research on misinformation and content moderation, offering potential avenues for future research. Finally, we conclude by summarizing the existing evidence on misinformation interventions and how marketers and public policy researchers may help advance efforts to promote a well-informed public.

Interventions for Mitigating Misinformation

Building on theories of why people believe and share misinformation (for reviews, see Ecker et al. [2022], Pennycook and Rand [2021a], and Van Bavel et al. [2024]), researchers and practitioners have developed (for a review, see Kozyreva et al. [2024]) and comparatively tested (Fazio et al. 2024; Gollwitzer et al. 2024) numerous interventions designed to target false and misleading claims. Here, we briefly review these approaches and discuss subsequent challenges of scalability and backlash to content moderation efforts.

Content-Neutral Interventions

One category of antimisinformation strategy is content-neutral intervention. We define this style of intervention as approaches that do not rely on identification of false or misleading content per se but rather promote or encourage faculties that may improve individuals’ ability to discern truths from falsehoods.

A broad type of such content-neutral interventions are media literacy strategies. Media literacy strategies may be deployed at a variety of different levels of depth and intensity. At the most intensive level, education researchers have developed school-based interventions to foster students’ ability to discern the veracity and credibility of information and sources. Some of these approaches include verification strategies (Caulfield and Wineburg 2023), lateral reading (the process of evaluating a source's credibility by searching for information about it on other reliable sites; Wineburg and McGrew 2019), and critical ignoring (strategies for avoiding low-quality and potentially harmful content; Kozyreva, Wineburg, et al. 2023). Educational strategies like lateral reading have been found to improve sharing and accuracy discernment among even young students (Barzilai and Stadtler 2025).

Researchers have also developed “inoculation”-style educational interventions, whereby individuals are instructed on how to detect specific manipulation tactics (e.g., emotional language, false dichotomies, scapegoating; Roozenbeek et al. 2022). Training people to identify manipulation tactics and features via educational games and videos (Basol, Roozenbeek, and Van der Linden 2020; Roozenbeek et al. 2022; Roozenbeek and Van der Linden 2019) has been shown to improve individuals’ propensities to accurately identify such tactics and can boost discernment between misinformation and true news, particularly when combined with instructive feedback (Leder et al. 2024).

Short but scalable lists of media literacy tips (e.g., “be skeptical of headlines,” “look closely at the URL”; Guess et al. 2020) have also been found to effectively improve discernment between mainstream and false news headlines (Gollwitzer et al. 2024; Guess et al. 2020). Media literacy tips can also be modified to promote trust in true news, as well as skepticism of falsehoods, while still maintaining efficacy in improving discernment (Altay, De Angelis, and Hoes 2024), and such an approach may also alleviate concerns of general warnings promoting overall skepticism of even true content (Clayton et al. 2020).

The aforementioned content-neutral interventions each involve teaching or boosting (Herzog and Hertwig 2025) individuals’ abilities to discern content veracity. Researchers have additionally developed even lighter-touch interventions that help promote existing faculties by prompting users to more carefully and deliberatively consider accuracy when deciding how to engage with online content. Drawing from evidence that accuracy perceptions are typically unaligned with sharing intentions, despite individuals stating a strong preference for sharing accurate content (Pennycook, Epstein, et al. 2021), researchers have found that prompting individuals to consider accuracy before making sharing decisions can improve sharing quality (Pennycook, Epstein, et al. 2021; Pennycook, McPhetres, et al. 2020; Pennycook and Rand 2022). Such “accuracy prompts” can be delivered in different formats, such as having people evaluate the veracity of one or several news headlines, prompting them to reflect on the importance of accuracy, or reminding them of social norms around sharing accurate content (Epstein et al. 2021). Accuracy prompts have been demonstrated to reduce misinformation sharing when delivered via digital advertisements at scale on platforms like Facebook and Twitter (Lin et al. 2024).

Such interventions can also be delivered in strategic ways. For example, platforms or users may circulate literacy tips or accuracy prompts during times when the potential harm or information quality of newsfeeds is expected to be highly variable or at heightened risk, such as during elections or following natural disasters. Timely deployment of such interventions may be useful not only for prompting users to consider accuracy when information quality may be particularly low but also for mitigating the risk of users becoming desensitized to such messaging over time. Relatedly, content-neutral interventions may be targeted at users more likely to encounter, believe, or share low-quality or false content, as determined via previous work on exposure to and sharing of low-quality posts (Baribi-Bartov, Swire-Thompson, and Grinberg 2024; Grinberg et al. 2019; Guess, Nagler, and Tucker 2019) or work on identifying linguistic or psychological correlates of false news sharing and poor sharing discriminability (Jun and Johar 2022; Schoenmueller, Blanchard, and Johar 2025).

Content-neutral interventions have several key benefits. First, they are effective on average: A recent megastudy tested media literacy tips, inoculation, and accuracy prompts and found all three to effectively promote veracity discernment (Fazio et al. 2024). Second, they are relatively scalable: Light-touch interventions like media literacy tips and accuracy prompts can be easy to deploy on social media platforms (e.g., via banners or announcements, as done with other forms of content moderation [Matias 2019] or in-feed via prompts or ads [Lin et al. 2024]) and avoid scalability issues inherent to approaches that require identification of specific falsehoods. Similarly, strategies like inoculation can help individuals identify potential broader categories of deception, assuming the manipulation strategies identified are diagnostic of false or misleading content. Third, such approaches may be relatively noncontroversial: Since these interventions do not target specific types of content or users, they may avoid pitfalls of individuals disagreeing with specific refutations or classification of claims.

That said, content-neutral interventions do have notable shortcomings. They are highly dependent on users engaging with or attending to the intervention, which may be particularly challenging for more intensive trainings delivered via educational modules or online interactive games. Conversely, lighter-touch interventions are likelier to have smaller effect sizes; for instance, accuracy prompts delivered at scale on Facebook were found to have an effect size of a 2.6% reduction in the probability of sharing misinformation in the hour after delivery (Lin et al. 2024). While this effect size is sizable for that of an online digital ad, it may be important to use heavier-handed interventions that can have greater effects on decreasing misinformation sharing. Finally, content-neutral interventions definitionally cannot focus on and target specific harmful or pernicious false claims or narratives, which may be problematic if such claims are relatively highly plausible and thus require targeted information or action to dispute.

Content-Based Interventions

A second category of antimisinformation strategy that alleviates the shortcomings of content-neutral approaches in exchange for lesser scalability is content-based intervention. We define this type of intervention as approaches contingent on the identification of specific falsehoods or misleading claims, which in turn act on such content or content purveyors.

One type of content-based intervention is debunking—providing corrective information to individuals who have seen or shared false information. There is extensive literature on the efficacy of refutations and corrections, and meta-analyses suggest that corrections have generally robust, directionally positive effects on belief (Chan et al. 2017; Walter and Murphy 2018; for a review, see Prike and Ecker [2023]). And promisingly, “backfire” effects, whereby corrections reduce belief accuracy, are rare (Porter and Wood 2024; Wood and Porter 2019). Corrections have also been demonstrated to be effective for important topics such as vaccine misinformation (Porter, Velez, and Wood 2022) and are generally robust to stylistic differences in how rebuttals are framed or written (Bode, Vraga, and Tully 2020; Martel, Mosleh, and Rand 2021; Pillai, Brown-Schmidt, and Fazio 2023; Swire-Thompson et al. 2021).

Another type of content-based intervention, and one of the most historically widely used antimisinformation strategies by online platforms (Instagram 2019; Meta 2020b), is providing warning labels concurrent with content exposure. Typically stylized as an overlay or tag informing viewers that the underlying content is false or misleading, such warning labels have also been demonstrated to be widely effective, on average, at reducing belief in, and sharing of, labeled content (Clayton et al. 2020; Martel and Rand 2024; Mena 2020; Pennycook, Bear, et al. 2020; Porter and Wood 2022; for a review, see Martel and Rand [2023]) and are theoretically related to research on disclosures and consumer protection (Mende et al. 2024). Relatedly, credibility indicators at the information source level have also been proposed (Zhang et al. 2018) and have been found to reduce sharing intentions for false news posts (Celadin et al. 2023). However, in field studies they have been found to have limited average effects on consumption of low-quality news (though may help promote news diet quality among those exposed to the most misinformation; see Aslett et al. 2022).

Refutations can also be provided prior to misinformation exposure—a strategy often referred to as “prebunking.” This approach involves providing individuals with forewarnings about specific false claims they are likely to encounter or that may be particularly harmful (Lewandowsky and Van der Linden 2021; Traberg et al. 2023, Section 8.2.1) and has been used by social media platforms like Twitter (now X) to forewarn users about likely voting misinformation (Twitter Support 2020). Although some evidence suggests that prebunks are less effective than debunks or concurrent warning labels (Brashier et al. 2021), providing individuals with preemptive credible information about specific falsehoods can help decrease misinformed beliefs even about important topics like election fraud (Carey et al. 2025).

Debunks, warning labels, and prebunks have been shown to be effective on average, and in a recent megastudy they were each demonstrated to promote both accuracy and sharing discernment between true and false or misleading news headlines (Fazio et al. 2024). That said, platforms do have the ability to take stronger action against content identified as harmfully misleading. These approaches also have the benefit of operating at the system, rather than the individual, level (Chater and Loewenstein 2023). One such type of action is adding “friction” to viewing or sharing content. Such friction can come in the form of interstitial warning labels (e.g., requiring users to click through an additional overlay to view underlying content; e.g., Sharevski et al. 2022) or by requiring users to click an additional confirmation before being able to share or repost content (Ershov and Morales 2024).

An additional step beyond friction is downranking misinformation in newsfeeds. Downranking other types of content, such as politically polarizing posts, has been found to decrease exposure to such content and have downstream effects on antidemocratic attitudes and partisan animosity (Bernstein et al. 2023; Piccardi et al. 2024). Downranking of identified misinformation and groups and websites that repeatedly share misinformation on Facebook has also been shown to decrease engagement with such content (Lin et al. 2024; Vincent, Théro, and Shabayek 2022).

Platforms can also choose to entirely remove, or deplatform, certain types of content or users. For instance, Twitter's deplatforming of repeated sharers of misinformation following the January 6 attack on the U.S. Capitol substantially reduced the reach of misinformation on the social media site (McCabe et al. 2024). Deplatforming norm-violating users—particularly prominent influencers and political elites—has also been demonstrated to have positive effects on information quality and toxic and hateful language (Jhaver et al. 2021; Müller and Schwarz 2023; Rauchfleisch and Kaiser 2021; Ribeiro et al. 2024). Removal and deplatforming are inherently effective at reducing targeted content or user influence on the focal platform, though these approaches may be subject to greater scrutiny by users and require platforms to maintain and enforce clear guidelines about community standards. Further, such policies may result in important externalities—for instance, the deplatforming of the right-wing social media platform Parler in January 2021 resulted in some users migrating to Telegram and being exposed to a greater amount of misinformation and partisan content; however, a majority of Parler users did not undertake such platform migration (Agarwal, Ananthakrishnan, and Tucker 2022; Horta Ribeiro et al. 2023).

Scaling Content Moderation

Content-neutral interventions may be delivered at scale via educational initiatives, online games or trainings, or platform-based reminders or banners. The latter approach has the important benefit of scalability: Platforms can relatively easily deliver content-neutral educational tips or reminders about accuracy and misinformation to users. However, when it comes to addressing specific harmfully misleading claims, it is likely optimal to use content-based interventions that can more forcefully refute or mitigate exposure to specific content. Unfortunately, content-based interventions have a notable drawback: They necessitate a scalable process for identifying false and misleading claims.

One prominent approach for identifying falsehoods, used at various times by major social media platforms like Facebook, Instagram, and TikTok (Bettadapur 2020; Meta n.d. b), is partnership with professional fact-checkers. Typically, this process has involved platforms using internal algorithms to identify posts that may be false or misleading. These identified posts would then be sent to third-party fact-checkers, who would choose which claims to thoroughly fact-check and report back to the platforms with their final evaluations. Posts identified as false or misleading by fact-checkers would then be acted on by the platform—typically via labeling and demotion, as was Meta's prior longstanding policy (Meta n.d. b). However, a shortcoming of this process is its scalability, as there are a limited number of professional fact-checkers and experts. In turn, a process relying primarily on expert fact-checkers for identifying falsehoods cannot keep up with the overall amount of content produced on major online platforms.

Several approaches have been proposed for scaling up the identification of false and misleading claims. One is leveraging social media users and the “wisdom of crowds” to enable layperson flagging and identification of misinformation. Recent research has shown that layperson crowds can accurately identify misinformation, exhibiting high agreement with fact-checker evaluations (Allen et al. 2021; Arechar et al. 2023; Resnick et al. 2023; for a review, see Martel, Allen, et al. [2024]). Complementing expert evaluations with novice evaluations may be one way to incorporate crowd ratings to identify misinformation, an approach similar to Facebook's Community Review (Meta n.d. a). To further increase scalability, platforms could also enable users to voluntarily and endogenously flag and fact-check content they encounter in their own social media feeds. Data from X's Community Notes, designed to do just this, shows that users leverage the platform largely to flag counterpartisan users (Allen, Martel, and Rand 2022; Yasseri and Menczer 2023), but that these classifications again are generally consistent with those by professional fact-checkers (Allen, Martel, and Rand 2022). Promisingly, recent evidence even suggests that “additional context” added by users via Community Notes can be effective: Contextual labels evaluated as cross-ideologically helpful (Ovadya and Thorburn 2022; Wojcik et al. 2022) have been shown to reduce reposting of, and increase deletion of, posts that have received Community Notes (Renault, Amariles, and Troussel 2024; Slaughter et al. 2025). That said, questions still remain regarding the ultimate scalability and impact of Community Notes, given that the majority of notes are not shown to all users, since they fail to reach consensus helpfulness by note evaluators (Bak-Coleman 2023; Fan, Dottle, and Equality 2022). Other platforms, such as YouTube, TikTok, Facebook, and Instagram, are trialing Community Notes–style crowd-based content moderation systems. Continued research on these platforms and their crowd-based content moderation initiatives is important for understanding their impact and scalability.

Another approach for scaling up identification and addressing of falsehoods is automatic detection via machine learning. Computer scientists have developed various models for identifying potentially false and misleading posts (for a review, see Islam et al. [2020]), and tech companies such as Meta have their own internal algorithms for identifying likely falsehoods and matching claims across posts (Meta 2020a.). More recently, researchers have even proposed processes for integrating large language models with (1) scalable fact-checking (Zhou et al. 2024), (2) debunking conspiracy theories (Costello, Pennycook, and Rand 2024), and (3) synthesizing fact-checking context on social media posts (i.e., aggregating across Community Notes contributions; De et al. 2024).

Fact-checked, crowdsourced, and automated detection and classification of misleading content are, of course, not mutually exclusive. Although such approaches have typically been studied in isolation or in comparison to one another (Drolsbach, Solovev, and Pröllochs 2024; Epstein, Foppiani, et al. 2022; Martel, Berinsky, et al. 2025; Pan et al. 2022; Yaqub et al. 2020), future research should propose and examine how to best combine these procedures for identifying and acting on harmfully misleading content at scale.

Relatedly, the content-neutral and content-based antimisinformation strategies reviewed above also need not be applied individually. Indeed, researchers have proposed the implementation of multiple overlapping interventions to best address misinformation thoroughly and at scale (dubbed a “Swiss cheese model”; Bode and Vraga 2021). While the interventions reviewed here have all been demonstrated to be effective in isolation (Fazio et al. 2024), some recent work provides burgeoning evidence as to the improved efficacy of combining approaches. For instance, some manipulation technique inoculation trainings can help improve individuals’ ability to identify emotional manipulation but not improve their ability to discern veracity; however, when combining inoculation with an accuracy prompt intervention, the combined intervention is able to achieve both deception technique recognition and improved truth discernment (Pennycook et al. 2024). Platforms deploying a range of interventions depending on the type and severity of content may also be in line with what users want. When asked what content moderation implementation individuals wanted on various news articles, users preferred adding informational warning labels on the most articles, followed by reducing distribution, and lastly by removal (Atreja, Hemphill, and Resnick 2023). Given heterogeneous preferences in how to act on different types of potentially misleading content (Atreja, Hemphill, and Resnick 2023; Kozyreva, Herzog, et al. 2023), future work may examine how to develop policies for retributive content moderation, such that the interventions applied on certain posts are proportional to the harm or falsity transgressions of that post. Finally, field data from Facebook suggest that during the COVID-19 pandemic, the platform's policy of labeling and downranking posts identified as false by third-party fact-checkers severely reduced sharing of that content (Lin et al. 2024). These findings again support the idea that combining approaches—in this case, labeling and downranking—can be highly effective for mitigating the impact of misinformation once it is identified, again highlighting the importance of developing fast and scalable identification of potentially misleading content.

Contemporary Challenges for Content Moderation

Concerns about the scalability and breadth of content moderation approaches stem from criticism of social media platforms for not having expansive or effective-enough policies to address misinformation, an issue that researchers and the media highlighted during the COVID-19 pandemic (Broniatowski et al. 2023; Donovan 2020). However, content moderation decisions made regarding the pandemic and information relating to the 2020 U.S. presidential election (namely, how to moderate posts about the COVID-19 “lab leak theory” and the Hunter Biden laptop controversy; Chan 2020; Hern 2021) also bolstered attitudes against content moderation practices, particularly among right-wing political elites. Republican lawmakers attempted to pass laws prohibiting social media platforms from moderating posts from political candidates (Allyn and Totenberg 2024), and Republicans in Congress have even investigated misinformation researchers as being part of a conspiracy to censor conservatives (Zakrzewski and Nix 2024). On the first day of his second presidential term, Donald Trump signed an executive order titled “Restoring Freedom of Speech and Ending Federal Censorship,” which targeted social media companies’ abilities to moderate speech and combat misinformation on their platforms (The White House 2025). Under the Trump Administration, the National Institutes of Health has directed its staff to identify grants related to “fighting misinformation”—a step that has typically preceded the termination of research funding (Oza 2025). Most recently, the National Science Foundation has declared that research “with the goal of combating ‘misinformation,’ ‘disinformation,’ and ‘malinformation’” is no longer aligned with its priorities and has terminated existing grants related to these research directives (Mervis 2025; Panchanathan 2025).

In turn, social media platforms themselves have rolled back their policies for addressing harmfully misleading content over the last several years (Kern 2022; Paul 2023; The YouTube Team 2023). Most recently, Mark Zuckerberg announced an end to Meta's partnership with third-party fact-checkers in the United States, despite acknowledging that this change would lead to more “bad stuff” on its platforms (Kaplan 2025).

These policy changes present new hurdles for researchers investigating how to mitigate harmfully misleading content. Here, we focus on three specific challenges for researchers in marketing and public policy: (1) leveraging public opinion and advertiser pressures for improving content moderation, (2) assessing claims of political bias given imbalanced enforcement outcomes, and (3) conducting on-platform research given limited access and transparency.

Do the attitudes of right-wing political elites and social media executives on misinformation mitigation align with those of social media platform users? An abundance of evidence suggests no. Polling of Americans suggests that the public overwhelmingly believes that social media companies should try to reduce the spread of harmful misinformation on their platforms—including the majority of Republicans (Martel, Berinsky, et al. 2024). Furthermore, the majority of Americans, including Republicans, support platforms applying warning labels from independent fact-checkers (Rand and Martel 2025) and want fact-checkers employed by platforms (Horne and Craig 2025). Related work also suggests that individuals prefer experts like fact-checkers to layperson crowds for evaluating potential misinformation, but that large, well-informed crowds can also approximate experts in perceived legitimacy as content moderators (Martel, Berinsky, et al. 2025).

Given this general support for misinformation mitigation, one avenue for future important investigation is determining how to better align platform policies with user preferences. One approach, particularly from a marketing perspective, is to examine the impact of indirect misinformation on consumers and advertisers. Researchers (Di Domenico and Ding 2023) have argued that low-quality information systems can harm overall consumer trust, perhaps making consumers less likely to positively evaluate on-platform content and harm trust relationships with brands and advertisers in this low-quality information ecosystem. Relatedly, recent work has shown that consumers actively punish companies that advertise on websites publishing misinformation (Ahmad et al. 2024). These findings are in line with public opinion supporting the need for better, not lesser, content moderation of misinformation, and highlight the importance and need for future work investigating the financial and corporate costs of advertising on platforms with low-quality and harmfully misleading content.

A second challenge from a public policy perspective is addressing concerns of political bias in content moderation and misinformation enforcement decisions. Right-wing pushback of misinformation mitigation efforts largely stems from claims of political bias against conservatives (Bond 2020), a concern also voiced by Republican public opinion in general (Vogels, Perrin, and Anderson 2020). However, multiple investigations have found that exposure to, and sharing of, low-quality and misleading content online is disproportionately by conservatives (González-Bailón et al. 2023; Grinberg et al. 2019; Guess, Nagler, and Tucker 2019; Guess, Nyhan, and Reifler 2018). Accordingly, recent work has found that Twitter users posting Trump hashtags were more likely to be suspended than their Biden-supporting counterparts. However, because these Trump-supporting users were also sharing content from lower-quality news outlets, they would still be suspended at much higher rates due to the quality of their shared content, even under entirely political neutral enforcement policies (Mosleh et al. 2024). Such a pattern of enforcement is also the case when the “ground truth” of misleading content is decided via cross-ideological agreement from layperson crowds: On X's Community Notes, more Republican posts (vs. Democratic posts) were flagged by users as misleading (Renault, Mosleh, and Rand 2025). Political imbalance in moderation thus appears to be the result of underlying asymmetry in sharing quality. Right-wing posts and users are acted on more by moderation systems because they are more likely to share low-quality or misleading content, even when the determination of such content is made by politically balanced crowds, rather than experts or fact-checkers (see Jackson 2025). A challenge for policy researchers is how to better persuade the public, as well as political elites, against a false equivalence between political bias and moderation imbalance.

A third challenge for researchers is how to investigate the policies and impacts of platform content moderation efforts given declining access to and transparency of on-platform data. In recent years, Elon Musk has ended access to the Academic Twitter API, making research about X far more costly for researchers (Fung 2023). Similarly, Meta shut down CrowdTangle, making it more difficult for researchers to monitor public posts on Facebook and Instagram (Ortutay 2024). Fortunately, some moderation initiatives still have relatively high levels of public transparency—for instance, X's Community Notes publishes its “bridging algorithm” for identifying helpful notes, along with a full database of user notes and helpfulness ratings (X n.d.). As companies like Meta look to emulate the Community Notes model (Meta n.d. c), it is integral for such platforms to make available their procedures and data for content moderation practices if they want external researchers to be able to evaluate such approaches. In lieu of such direct transparency or direct data sharing agreements or collaborations (e.g., Meta’s initiative to study the 2020 U.S. elections; Clegg 2020), researchers should also advocate for the continued ability to externally assess platforms via methods such as scraping of public content and audit studies (Matias, Hounsel, and Feamster 2022; for policy implications of algorithmic transparency for assessing the effects of ranking algorithms, see Eckles [2022]).

In sum, prior work on misinformation has provided an extensive array of interventions that can improve belief discernment and sharing quality. However, given the current political climate and despite cross-partisan public demand, policy makers and platforms are reticent to apply or assess many of these approaches on platform. Even the most effective misinformation policies provide no benefit if they are not adopted. Therefore, investigating how to demonstrate the importance of content moderation policies—to the public, to advertisers and platforms, and via assessing existing and potential policy decisions and systems—is integral to maintaining and building on current practices for reducing the impacts of harmfully misleading content.

Open Questions and Pitfalls for Misinformation Research

Scalability and platform adoption and assessment are two major challenges facing the application of well-researched misinformation mitigation interventions. However, misinformation research as a field also faces numerous open questions regarding how to develop novel and impactful interventions, as well as how to better understand the scope and harms of misinformation to begin with (Budak et al. 2024; Kupferschmidt 2024). Here, we briefly review several of those open questions and provide suggestions for avoiding common pitfalls in misinformation research.

Defining, Identifying, and Sampling Misinformation and Misleading Content

Misinformation research has typically gone about identifying false and low-quality content in one of several ways. One approach, particularly used in field research, is to define content at the domain level—for instance, identifying low-quality domains (Lin et al. 2023) and examining exposure to, and sharing of, content from these low-quality sites. A second approach is to define misinformation as content that has been evaluated as false by professional fact-checkers. This practice is common in both lab-based experiments (Pennycook, Binnendyk, et al. 2021) and different forms of field research (e.g., Martel, Mosleh, et al. 2025; Vosoughi, Roy, and Aral 2018). Misinformation researchers have also proposed using sets of constructed, rather than naturalistic, true and false news stimuli (Maertens et al. 2023). However, each of these approaches omits categories of false or misleading content that may be incredibly important to study. Defining misinformation via low-quality domains ignores misleading content published by high-quality, mainstream outlets (and also can miscategorize accurate content published by low-quality sites). Further, identifying misinformation only through fact-checked false news items omits content that fact-checkers have not yet evaluated as well as content that may not be technically false but nonetheless could be harmfully misleading.

For instance, recent work examining the impacts of content on Facebook skeptical of the COVID-19 vaccine found that posts evaluated as false by professional fact-checkers had large effects on decreasing vaccination intentions for those exposed to such posts. However, these flagged posts were viewed far less by Facebook users compared with vaccine-skeptical posts that had not been evaluated as false. Thus, when accounting for views, the unlabeled posts had far greater aggregate effects on decreasing vaccination intentions (Allen, Watts, and Rand 2024). These findings demonstrate the importance of investigating the misleadingness of content beyond simple “fake news”–style posts and focusing more on highly viewed content and information about important topics that may harmfully shape beliefs and actions (Allen and Rand 2024).

In practice, this may take the form of greater investigation of mainstream media amplification of falsehoods and legitimization of misleading content (Goel et al. 2024; Shi, Liu, and Srinivasan 2022) as well as closer assessment of how choices by the media may enable readers to embrace incorrect beliefs (e.g., through cherry-picking; Li, Hsee, and Wang 2024). Another incredibly important domain for further assessment is focusing on political elites and primary purveyors of misinformation and misleading claims. For example, previous work has investigated the propensity of (primarily right-wing) political elites to share posts from low-quality domains (Lasser et al. 2022) and has constructed measures for assessing exposure to misinformation from political elites based on professional fact-checking of public figures (Mosleh and Rand 2022). Future work should extend such research by also examining how political elites may spread false or misleading claims without linking to low-quality domains or relying on only claims that have been fact-checked—perhaps by evaluating elite claims via a combination of layperson crowd evaluations and large language model assessment. For marketers in particular, one place to start may be in online political ads, which have largely been unmoderated by social media platforms (Isaac and Kang 2020; Klepper 2024). Misinformation investigation focused on messaging from political elites is particularly important given the reach and persuasiveness of elite messaging (Lenz 2009; Zaller 1992). Ascertaining which types of content-neutral and content-based interventions may be effective against misinformation from political elites is important future work. Relatedly, more research is needed on how misleading claims may be spread by online influencers such as internet personalities and podcasters (Kelly 2024)—and, relatedly, how to alert consumers to misleading claims in longer-form, multimedia content in a scalable and effective way (Pathiyan Cherumanal, Gadiraju, and Spina 2024).

Average Versus Heterogeneous Treatment Effects

Relatedly, descriptive work on misinformation has shown that exposure to blatantly false or low-quality content is highly concentrated at the extreme tails of the population (Budak et al. 2024; González-Bailón et al. 2023) such that, at least in the U.S. context, exposure to outright misinformation is quite rare (Allen et al. 2020). Similarly, sharing of misinformation is also highly concentrated at the extremes; for example, it has been estimated that on Twitter, .3% users account for 80% of tweets linking to low-quality domains (Baribi-Bartov, Swire-Thompson, and Grinberg 2024).

Such findings have led to calls for work that specifically tests the efficacy of interventions for individuals most likely to be exposed to, or to share, misinformation (Brashier 2024; Budak et al. 2024). Recent work has provided some evidence for the broad effectiveness of common misinformation interventions; for example, fact-checker warning labels decrease belief in, and sharing of, labeled false posts even for those who express distrust of fact-checkers (Martel and Rand 2024), and accuracy prompts are effective at improving sharing quality even for strong Republicans (Martel, Rathje, et al. 2024). However, further research should continue to examine not only the average treatment effects of misinformation interventions but also heterogeneous treatment effects among subpopulations who are particularly prone to seeing or sharing misinformation.

Measuring and Assessing Intervention Efficacy

Beyond examining potentially heterogenous effects of treatments on different types of misleading stimuli and different categories of individuals, researchers investigating the mitigation of misleading content must also be careful about how they go about measuring and assessing intervention efficacy.

One common pitfall of misinformation research is only examining the effect of interventions on misinformation, rather than examining the effect on how individuals discern between false (or misleading) and true content (Guay et al. 2023). Looking at discernment as a dependent variable is critical for evaluating whether interventions are actually improving belief or sharing quality, or if instead they are merely causing general skepticism (i.e., a similar reduction of belief in, or sharing of, both accurate and inaccurate content). In addition to this, future policy research should also assess the relative costs of differential errors in content moderation. For instance, content-based interventions can commit different types of errors, such as erroneous application to true content (false positives; see, e.g., Freeze et al. 2021) or erroneous misapplication to false or misleading content (false negatives; see, e.g., Pennycook, Bear, et al. 2020). Overall content moderation policies should examine the differential harms of such errors when evaluating the overall impact of intervention.

The selection of dependent variables of interest is also a critical aspect of misinformation research and evaluation of interventions. Typically, interventions have focused on misinformation belief or sharing (Fazio et al. 2024). When examining these outcomes, it is important to assess them separately between-subjects rather than use joint elicitation within-subjects, since asking about accuracy can impact sharing intentions and vice versa (Epstein et al. 2023). Further research should also examine how misinformation interventions impact broader beliefs and downstream behavior. For instance, recent work has shown that while refutations are effective at disputing the focal targeted belief, they have minimal impact on more general beliefs and attitudes and can be short lasting in efficacy (Carey et al. 2022, 2024; Porter and Wood 2024). Other research has even found evidence for potential downstream backfire effects on behavior—for example, Mosleh et al. (2021) found that individuals corrected on Twitter were more likely to subsequently retweet lower-quality, more partisan content. In particular, effects on user attention and endogenous information search may be important domains for future work (e.g., see Aslett et al. 2024; Epstein, Lin, et al. 2022; Graham and Porter 2025).

Generalizing Across Topics, Media, and Platforms

Future research should also continue to assess the generalizability of feasible antimisinformation interventions across various levels of application. At the topic level, much research has focused on political misinformation. Other domains, such as public health, science, and climate change, are also important to investigate (Swire-Thompson and Lazer 2020; Treen, Williams, and O’Neill 2020; West and Bergstrom 2021). In addition, misinformation interventions should be applied and tested in the domain of misleading advertising or other forms of commercial speech; for instance, recent reporting has documented how deceptive political fundraising has cost consumers millions of dollars (Ellis et al. 2025).

Interventions developed by misinformation researchers have also largely been focused on addressing news and text-post-style content on platforms like Facebook and X. However, there is a growing need to extend such work to different media (e.g., audio content like podcasts, visual content like memes, video content such as on YouTube and TikTok). While approaches like warning labels and Community Notes have begun to be tested in these settings (Ling, Gummadi, and Zannettou 2023; The YouTube Team 2024), more work is needed on how to best apply interventions designed for scrollable newsfeeds to cross-medium platforms. Relatedly, social media has become increasingly fragmented, requiring scholars to examine how engagement with low-quality news and misleading content may vary across platforms more generally. Some research has begun to investigate cross-platform descriptive differences in low-quality and partisan news exposure (Mosleh, Allen, and Rand 2024), but future research is needed to examine how interventions may generalize across social media platforms.

Cross-Linguistic and Cross-National Interventions

Finally, the majority of research published on misinformation has focused on English-language misinformation in the Global North. Similar patterns are reflected in the actual implementation of misinformation-mitigating policies—for example, an assessment of Facebook's misinformation labeling practices during the COVID-19 pandemic found that Spanish, Italian, and Portuguese COVID-19 misinformation was far less likely to be labeled as false than English-language false content (Avaaz 2020). Advances in automated detection of potential misinformation via large language models also necessitates that researchers assess how well such detection applies to non-English content. The efficacy of antimisinformation interventions should also be continually examined in cross-national contexts, particularly in the Global South. Some research has found that interventions like digital literacy tips, accuracy prompts, debunking, crowdsourced fact-checking, and classroom-based misinformation identification education have cross-national efficacy (Amar et al. 2025; Arechar et al. 2023; Badrinathan, Chauchard, and Siddiqui 2025; Guess et al. 2020; Porter, Velez, and Wood 2023). Researchers should heed calls for future work to learn how lessons from misinformation research in the Global North can best be applied to the Global South (Blair et al. 2024).

Conclusion

The interdisciplinary study of misinformation has helped develop and test various content-neutral and content-based interventions. Promisingly, growing evidence suggests that many of these interventions are effective, on average, at reducing belief in, and sharing of, falsehoods. However, important practical and theoretical challenges remain. Successful implementation of misinformation-mitigating efforts relies on approaches that are scalable. Further work on how to combine interventions and leverage expert, crowd, and machine-learning-based procedures for identifying and classifying misleading content is integral for overcoming this challenge. Furthermore, content moderation efforts have received significant pushback over claims of bias and censorship, largely by right-wing political elites. However, this is out of line with public opinion and consumer demand for moderation. Thus, examining how to demonstrate the importance and efficacy of moderation interventions from a marketing and public policy perspective is necessary for translating effective misinformation interventions from the lab to the field and for continuing to assess policy impacts on-platform. Finally, we review open theoretical questions and common pitfalls for researchers investigating misinformation: how to define the topic and impacts of interest, how to measure efficacy and for whom, and how to examine the generalizability of interventions across topics, media, and contexts.