Skip to main content
PNAS Nexus logoLink to PNAS Nexus
. 2026 Apr 28;5(6):pgag142. doi: 10.1093/pnasnexus/pgag142

Online communitarian appeals increase opposition to violent extremism

Jason Lyall 1,✉, Brendan Nyhan 2, Elsa Voytas 3
Editor: Paul Van Lange
PMCID: PMC13229404  PMID: 42238351

Abstract

Governments and nongovernmental organizations have implemented hundreds of countering violent extremism (CVE) programs designed to combat radicalization and reduce support for violent extremist organizations (VEOs). However, the effectiveness of these programs is rarely evaluated rigorously, providing little guidance to scholars or practitioners. We therefore pair a large-scale Facebook study (39.8 million impressions) with an online survey experiment (N=7,707 respondents) to compare the effectiveness of three brief videos designed with local partners to reduce support for violent extremism and VEOs in Bangladesh. We find that a message of communal harmony from religious authorities received greater Facebook engagement and more frequent supportive emojis than messages stating that all religions oppose terrorism or warning about the costs of terrorism. We test the robustness of these messages by randomly exposing respondents to brief footage of an antigovernment protest organized by a local VEO. Communal harmony was the only message that increased net opposition to extremism after this counter-programming was introduced. An exploratory analysis further reveals that the messages do not affect intentions to attend VEO events or share VEO messages online. Instead, they increase intentions to oppose VEOs by sharing messages or attending rallies by mainstream or antiextremist groups. These results highlight the value of communitarian appeals by religious authorities as well as the need for more experimental, multiplatform evaluations of CVE messaging that separately measure VEO support and opposition.

Keywords: violent extremism, social media, online experiments


Significance statement.

In recent decades, governments, nongovernmental organizations, and aid agencies have crafted hundreds of countering violent extremism (CVE) programs to meet the threat from violent extremist organizations. Yet few have been rigorously evaluated, especially in non-Western contexts, and little consensus exists around which messages work best. Using Facebook and survey experiments, we test the effectiveness of brief videos featuring religious authorities that were carefully designed with local partners to reduce support for extremism. Our results indicate that messages of communal harmony in Bangladesh generate more Facebook engagement than alternative messages and most effectively promote opposition to extremism in the face of extremist counter-messaging. These findings offer lessons for designing CVE programs and evaluating their effectiveness.

Introduction

Over the past decade, a rising tide of violent extremism has killed and injured tens of thousands of civilians. Violent extremist organizations (VEOs) such as Boko Haram, Al Qaeda, Islamic State, Al-Nusra Front, and Al-Shabab now seek to destabilize and topple fragile governments while marshaling the power of the Internet to recruit and mobilize new followers. For their part, governments, nongovernmental organizations, and aid agencies have implemented hundreds of countering violent extremism (CVE) campaigns in more than 40 countries to reduce support for VEOs (for reviews, see Refs. (1, 2)).a

There is a pressing need, then, to determine whether the messages promoted in CVE campaigns actually reduce support for VEOs. Despite decades of research, however, few CVE studies leverage random assignment to identify causal effects (for reviews, see Refs. (3, 4)). One review of 463 CVE studies from 2005 to 2020, for example, found only eight with randomized treatment and control messages (5). Moreover, evaluations of online CVE messaging tend to focus on engagement rates, which often diverge from effects on attitudes (6–8). In general, considerable confusion remains over how to measure support for violent extremism, leaving CVE efforts with unclear metrics for success (1, 3). Finally, most studies focus on minority populations in Western countries, neglecting threats elsewhere. For instance, 30% of CVE evaluations have been conducted in Europe compared to only 4% in South Asia despite the latter having more VEOs and associated civilian casualties (5).

In this study, we evaluate the effectiveness of online CVE messaging at reducing support for violent extremism and VEOs by testing how different messages affected online engagement and attitudes toward violent extremism in Bangladesh. Drawing on three purpose-built videos designed with local partners to reduce support for extremism while minimizing risks to participants, we pair a large-scale Facebook study (39.8 million impressions) comparing message performance with an online survey experiment (N=7,707 respondents) that randomizes exposure to one of these videos (or a placebo). Each video featured religious authorities narrating one of three messages designed to resonate with Bangladeshis: (i) a focus on how all religions promote peace and renounce violence (“peaceful religions”); (ii) how terrorism destroys lives and property (“costs of terrorism”); or (iii) an appeal to cherish and protect Bangladesh’s multifaith harmony by standing “shoulder to shoulder” against terrorism (“communal harmony”).

Taken together, our empirical strategy combines multiple online engagement metrics on Facebook with a rich set of survey-based attitudinal measures, including multi-item measures of net opposition to violent extremism, attitudes toward religious outgroups, and prosocial emotions of hope and pride that might counter the negative emotions, including anger and perceived unfairness, that have been linked to radicalization (9). We also provide new insight into attitudinal effects by disaggregating our measure of net opposition to violent extremism into pro- and antiextremist behavioral intentions, permitting a more fine-grained assessment of how these messages affect attitudes toward VEOs.

Contrary to prior research and our own preregistered hypotheses, we find that a message of communal harmony, rather than appeals to peaceful religious coexistence, performed best at generating engagement on social media and in increasing opposition to extremism. The communal harmony message generated the highest level of engagement on Facebook, including the highest rate of positive emojis. In the survey experiment, the communal harmony messaging also generated the largest increase in the positive emotions of hope and pride as well as opposition to extremism. Notably, its effects on opposition to extremism were robust even when a random subset of survey participants was exposed to a counter-messaging video from a violent protest led by Hefazat-e-Islam, the largest domestic Islamist movement in Bangladesh. By contrast, the effects of the peaceful religions and costs of terrorism messages were sharply reduced by exposure to the protest video. Our results underscore the need to evaluate the effectiveness of CVE messaging in experiments conducted across multiple platforms and outcome measures and to explicitly incorporate VEO counter-messages into these evaluations.

Context: countering extremism in Bangladesh

We evaluate the effectiveness of online CVE messages at reducing support for VEOs in Bangladesh. Despite being the eighth-most populous country in the world (174 million in 2024) and the home to the fourth-largest Muslim population, Bangladesh has largely been ignored in studies of violent extremism (10). This omission is a mistake. Its mix of political fragility, surging inflation, and democratic backsliding has made the country a recruiting ground for no fewer than eight major homegrown and international VEOs as well as numerous smaller factions (11). These groups, which compete for support among young male recruits from the educated middle class, view Bangladesh’s democratic institutions as un-Islamic and corrupt and share similar ideological precepts, including hostility to minorities (including ethnic, sexual, and religious groups), women, India, and the West. After a 2016 terrorist attack killed 22 civilians, the governing Awami League launched an aggressive counter-terrorism campaign of increased police action and limits on civil liberties.

Despite these efforts, Bangladesh recorded 1,477 terrorist incidents, including 296 lethal attacks in which 188 civilians and 250 extremists were killed, from January 2016 to September 2025 (12). Surveys also recorded fairly broad public support for these organizations. In 2017, nearly one-third of Bangladeshis expressed support for the goals of Islamic State-Bangladesh (ISIS-B) and 42% acknowledged supporting one of three named VEOs (13). The perception that Bangladesh is becoming politically unstable is also gaining traction. In September 2019, 25% of survey respondents viewed stability as “poor or very poor,” while 6% believed it would worsen “next year.” By April 2023, those numbers had climbed to 49 and 27%, respectively (14). These views proved prescient: in July 2024, massive street protests erupted against the Awami League, forcing the resignation of Prime Minister Sheikh Hasina after nearly 1,000 demonstrators were killed by security forces. A caretaker government now faces surging anti-Hindu and antiregime violence, much of it orchestrated by VEOs.

We focus on two of the most important VEOs in Bangladesh: ISIS-B and Hefazat-e-Islam. ISIS-B is the largest and deadliest transnational VEO in Bangladesh. Active since 2015, ISIS-B seeks to overthrow Bangladeshi democracy and impose an Islamic state, with eventual designs on building a regional caliphate. It embraces violent jihad and has claimed credit for high-profile attacks against police, non-Muslims, and Western targets in Bangladesh, including aid agencies. It also retains close ties with several important, if small, domestic VEOs (15). ISIS-B has an extensive social media and online presence, with sophisticated propaganda and recruitment efforts, including slickly produced videos that appear on Facebook, YouTube, and Telegram. These appeals typically target educated young men, especially university students, and combine criticism of the government with calls for violent action against it. At least 2,700 militants affiliated with ISIS-B and its allies have been arrested since 2016 (12).

Hefazat-e-Islam is, by contrast, the largest homegrown Islamist organization in Bangladesh, with a membership estimated in the tens of thousands. Created in 2010, Hefazat burst onto the scene in 2013 with a massive rally in Dhaka to support its “13 demands” agenda, including compulsory Islamic education, ending gender mixing in public spaces, and capital punishment for blasphemy (15). Drawing from its extensive system of madrassas, Hefazat has increasingly organized large-scale street demonstrations, including against a visit to Bangladesh by Indian Prime Minister Narendra Modi in 2021, that have devolved into armed clashes with police. Hefazat supporters have also engaged in mob violence against religious and sexual minorities (16). Though formally rejecting violence, Hefazat maintains ties to VEOs and has created the conditions for radicalization through confrontational rallies and far-reaching social media campaigns on Facebook and YouTube. The governing Awami League, which initially tried to placate Hefazat by acceding to several of its demands, reversed course in 2021 and jailed dozens of Hefazat leaders and supporters following the violent anti-Modi protests (16).

Theoretical framework: the effects of CVE message content

We build on existing research to develop our argument and two alternative explanations about how the content of CVE messaging might affect online engagement, prosocial emotions such as pride and hope, attitudes toward religious outgroups, and willingness to take pro- or antiextremist actions.

We argue that religious messaging can lower support for VEOs by appealing to viewers’ morality. When backed by a leading religious figure, religious messages can cast VEOs as sinful or hostile to one’s faith, lowering support for organizations that cite religion as inspiration for violent extremism. Religious appeals can also bolster prosocial emotions such as hope while providing an authoritative basis for messages of peaceful coexistence and tolerance. These appeals also challenge extremist organizations on their own terms, using religious themes and symbols to reject violence. Such efforts can delegitimize VEOs by drawing attention to the disconnect between shared religious norms that extoll the virtues of peace and the violent behavior of these organizations. Religious messages can therefore leverage shared moral norms to discredit VEO conduct, including its treatment of outgroups.

In addition, prior studies have found that religious messages from religious authorities have reduced the spread of sectarian hate speech online (17); favorably shifted attitudes and behavioral intentions toward former Boko Haram fighters in Nigeria (18); increased charitable giving in Afghanistan (19); and increased support for women holding political office in Egypt (20). As a consequence, we expect that religious messages are especially potent for reducing support for VEOs when endorsed by religious authorities.

We therefore propose two preregistered hypotheses. First, we anticipate that a message about peaceful religions will generate the highest level of Facebook engagement among our three campaigns. Second, we expect that this religious appeal will increase (i) prosocial emotions of hope and pride, (ii) positive attitudes toward religious outgroups, and (iii) opposition to extremism and related VEOs.

We note that these hypotheses are falsifiable: it is possible that these messages have no, or even negative, effects on respondents. While most studies conclude that religious programming increases prosocial attitudes and behavior, the evidence base remains thin. A handful of studies have reached the opposite conclusion. Experimental evidence from Burkina Faso, for example, found that religious messaging failed to influence attitudes and behavior toward religious extremism. Worse, exposure to Islamic teachings actually reduced tolerance and interethnic trust among participants from the dominant ethnic group (21).

We test our claims about the efficacy of religious messaging against two alternative explanations. First, it is plausible that communitarian appeals can build bridges between groups by emphasizing shared membership in the same political community. These messaging campaigns aim to activate a superordinate identity that emphasizes the importance of the broader community and each individual’s responsibility to uphold its values and norms. In this framework, VEOs are cast as threatening shared values, a position reinforced when authorities from different communities endorse the message. Communitarian messaging also emphasizes the harm that violence can do to the solidarity and cohesion of a civic community while stressing an individual’s duty to defend communal interests against threats. Though evidence is sparse, narrative-driven campaigns have used communal appeals to reduce support for extremism in Burkina Faso (22) and to encourage reporting corruption in Nigeria (23).

Second, negative messages that highlight the physical damage and emotional harm of terrorism can reduce support for VEOs. The reminder of prior losses, along with the prospect of future harm, might shift attitudes away from VEOs and motivate action against violent extremism. The logic here is simple: individuals are unlikely to support VEOs if confronted with the costs associated with extremist violence, whether measured in terms of lives lost, property damage, or more general societal disorder. While a few ideologues might be attracted to violent imagery, most viewers are likely to be alarmed by this content (24). Similarly, destructive protests have been found to reduce support for the organizations that carry them out as moderate followers become disgusted with the harm inflicted in their name (25).

Each of these alternative explanations has potential drawbacks. It is not clear, for example, whether communal appeals are sufficient to motivate marginalized individuals, or even those in privileged positions, to undertake costly action to defend the political community. Negative messages that highlight the graphic nature of extremist violence might also prove ineffective. Some viewers might feel uncomfortable with such content and simply bypass it. These messages might also exacerbate intergroup tensions in diverse settings if the VEO is perceived as representing a particular group’s interests. As a result, some CVE handbooks now caution against relying too heavily on negative messaging (8, 26, 27).

The effects of VEO counter-messaging

To date, nearly all CVE evaluations ignore how individuals are enmeshed in complicated information environments that often expose them to VEO counter-messaging. We therefore consider how VEO messages might amplify or blunt online CVE programming. For example, it is plausible that VEO messages, especially if accompanied by violent imagery (24), increase the antiextremist effects of CVE messaging by reminding the public of the threat they pose to the broader community (18). Alternatively, VEO messages might undercut CVE campaigns by increasing grievances against governing authorities or stoking resentment against targeted groups (21), producing a null effect or even a proextremist attitudinal shift.

Empirical strategy

We designed two studies in cooperation with the International Republican Institute (IRI), an American nonprofit organization that specializes in global democracy promotion, to evaluate its ongoing CVE campaign in Bangladesh while minimizing risk to participants (see the Ethics section below). We first tested the effects of three different video messages created by local partners—communal harmony, religious coexistence, and the costs of terrorism—on Facebook engagement in two waves during and after Ramadan. We then conducted an online survey experiment measuring how exposure to one of these three messages influenced political attitudes (relative to a placebo), including views toward religious outgroups, prosocial emotions of hope and pride, and opposition to violent extremism. This combined evaluation approach allows us to merge large-scale engagement data from a real-world social media platform with a direct experimental test of message effects on several relevant political attitudes. As we show below, our online survey sample closely resembles the demographics of Facebook users in a nationally representative survey, allowing us to combine insights from sets of results in a sample that includes large numbers of the demographic that is most vulnerable to political extremism in Bangladesh (young, educated Muslim men).

Developing the messages

Facebook content has been tested in studies of topics ranging from partisanship and climate change (28) and ideological polarization (29) to intergroup relations (30) and compliance with COVID-19 restrictions (31). Some CVE campaigns use Facebook videos as well, though typically without testing multiple ads (32).

Our messages were created through a multistage process. IRI first conducted national surveys (2017–2019) and focus groups (2017–2018) to collect data on the drivers of support for extremism in Bangladesh. Based on these findings, IRI’s office in Dhaka developed three broad types of CVE messages: interfaith appeals to peace (a religious message based on rejection of violence by each of the major faiths in the country); calls for communal harmony using communitarian reasoning; and negative messages emphasizing the destructiveness of terrorist attacks. Various combinations of potential endorsers of these messages were considered, including religious leaders, women, famous athletes, students, and victims of terrorist violence. Additional focus groups were then convened in rural and urban areas in 2022 to pilot 10 combinations of message and messenger. Consistent with other studies (18, 19), IRI concluded that religious leaders commanded the greatest trust and deference.

A leading Bangladeshi media firm, Beatnik Digital, was then tasked with creating ad scripts and producing three short (37–38 s) Facebook videos optimized for mobile phone viewing. These videos all use the same actors portraying the same religious figures (a Muslim imam, a Catholic priest, a Hindu priest, and a Buddhist monk). Each combines live action and a brief animated segment of similar duration (9–13 s). Each video opens with all four actors standing side-by-side and stating the message’s central theme; each concludes with a written restatement of that message. Following the advice of our local partner, Beatnik, we allowed for some naturalistic differences between the videos. For example, only the communitarian video concludes with all four actors on the screen. Moreover, the religious video does have a very brief animated scene of militants firing their weapons (1 s long). It is possible that this imagery might depress estimates of the efficacy of our religious messaging. Similarly, the animation in the costs of terrorism video spends the most time on terrorist violence or its threat (about 11 s, compared to about 5 s for the other two messages). In short, while we worked to produce naturalistic content of similar structure and duration, we cannot rule out that remaining differences, however small, might have influenced our estimates of the efficacy of each message, including for the communitarian video.

Table 1 reproduces each video’s thumbnail text and script.

Table 1.

Study stimuli (translated from Bengali to English). Treatment videos: communal harmony (https://youtu.be/bAEXcXUNymA), peaceful religions (https://youtu.be/Ls_JXxj2xNc), and costs of terrorism (https://youtu.be/2NSy0JM242c). English subtitles in the videos seen by participants vary slightly from the translations provided below.

Ad name Ad text Video script
Communal harmony Inline graphic We Stand Together Against Terrorism
There is no place for terrorism in our harmonious Bangladesh.
Greed, violence, and arrogance destroy people. Our Bangladesh is a shining example of communal harmony. All religions have been here since time immemorial. People of color are living in harmony. We have made the country independent by fighting shoulder to shoulder in 1971 as well. But some evil forces are trying to establish terrorism here in the name of religion. But we have to remember, there is no place for terrorism in any peaceful country and religion. So, let’s all swear—we are all together, against terrorism.
Peaceful religions Inline graphic Terrorism is Sinful
Whatever religion one belongs to, terrorism is sinful
People can keep the country and society free from terrorism and militancy if they want. Every religion speaks of peace and humanity. Whichever religion one belongs from, it is wrong to kill someone, to cause loss of life or property, or to coerce. This is sinful! To be a real human being, society and life must be kept free from wickedness, terrorism, and militancy. Man is the best creature of creation. Therefore, without capitalizing on religion and engaging in injustice, let’s build a peaceful, conflict-free, country together.
Costs of terrorism Inline graphic Terrorism is a crime
The crime of terrorism threatens lives, property, and safety
Let us stand up against terrorism. When a terrorist attack occurs, many people are harmed at the same time. Some of the victims die, while others are left disabled, spending their entire lives haunted by horrific memories. Likewise, among the attackers, some are killed and some are punished, and their loved ones suffer the consequences. We must not forget: destruction is not the way; peace is the path of light. To bring people onto the path of light, we must hold on to patience and forbearance. It is through the virtue of patience that a person can refrain from wrongdoing and keep themselves happy and secure.

Experimental design and outcomes

Facebook study

Our Facebook study was conducted in two waves, one during Ramadan (2022 April 23–30) and one after (2022 May 25–June 7). Tables S1 and S2 detail our schedule and spend rate; Section S2 provides implementation details.

Our sample was drawn from Facebook users in Bangladesh aged 18 years or older. We use two different approaches to conduct our evaluation: (i) a cluster-based approach that closely approximates an experiment by randomly assigning videos across specified partitions of the sample; and (ii) a split A/B testing approach in which Facebook calculates relative engagement rates for each video. Our results are consistent across methods, increasing confidence that observed differences across messages are not an artifact of one approach or the other.

We partitioned our sample into clusters by age (18–65+ in 3-year groups), sex (male/female), marital status (married or not married), education (at least some high school, at least some college, or neither), and proximity to major cities (within 25 miles of Dhaka, within 25 miles of Chittagong, within 25 miles of Khulna, within 25 miles of Rajshahi, within 25 miles of Sylhet, or none of the above). This procedure generated 1,224 clusters.

Next, we divided these clusters between Facebook’s split test (or A/B testing) function and a randomized cluster design. To do so, we randomly assigned the first age group (18–20) to the “split” condition. We then alternated assignment by age group (eg 21–23 receives the cluster design, 24–26 receives the Facebook “split” condition, etc.). Since we have an odd number of age groups and the first age group is assigned to the “split” condition, nine age groups were assigned to the “split” condition and eight to the “cluster” condition.b

For those assigned to the Facebook “split” condition, we utilize Facebook’s split test function. Specifically, the nine age groups assigned to the “split” condition each receive three ads for A/B testing, resulting in a total of 27 ads. In that case, we specify our three treatment conditions and Facebook runs each treatment simultaneously among the specified population for the designated amount of time. Facebook calculates engagement rates for each ad to declare a final “winner”—the ad which generates the most engagement at the lowest cost.

For those assigned to the cluster condition, we follow the procedures used in Refs. (33, 34), and (35) to implement our cluster randomization, manually crossing the eight age groups assigned to the “cluster” condition with possible values for sex, education level, marital status, and geographic area to create a total of 576 clusters.

The clusters were then blocked by the midpoint of the estimated cluster size provided by Facebook (eg 20,000 for a cluster of 10,000–30,000) and randomly assigned to receive a different video. Because Facebook limits pages from running more than 250 ads at a time, we conducted our experiment in three phases, randomly assigning each cluster to a phase that corresponded to the order in which the videos ran. We used the treatment condition assigned to each cluster as a blocking variable to ensure that each phase consisted of a balanced set of treatment videos. In the split test and cluster design, individuals were assigned only one video to watch.

We use multiple Facebook engagement measures to test the relative effectiveness of our three CVE videos when they appeared in user Feeds on Facebook. These measures include an omnibus measure of post engagement (post shares, reactions, saves, comments, likes, and interactions; >3 s video plays; photo views; and link clicks) as well as separate measures of link clicks (a measure of active information-seeking), post comments (a more costly form of engagement), post shares (actively spreading information), and page likes (expressing approval of the source and requesting more information from it). We standardize these by dividing each outcome by impressions (converting them into engagement rates) and multiply the resulting values by 10,000 to reduce leading zero clutter and facilitate understanding (34).

Survey experiment

Our online survey was piloted on 2022 August 19–26 and fully launched from 2022 August 27 to September 19. Our sample was recruited by TGM Research via Facebook ads and its own panel of Bangladeshi respondents. Respondents completed the survey on Qualtrics. To be eligible, participants had to pass two pretreatment attention checks: (i) verifying that they could view an embedded video (eg on cellphones) by answering a factual question about its content; and (ii) correctly answering a basic factual question to confirm that they were reading questions carefully. We dropped all respondents who failed either attention check. Our sample consists of 7,707 respondents who passed pretreatment attention checks and reached the first experimental manipulation. As a result, attention rates to the experimental stimuli were very high: 77.8% of respondents assigned to watch a CVE video correctly answered a factual question about it as did 99.3% assigned to watch the protest video (we did not drop respondents who failed either question to avoid post-treatment bias).

Given our interest in CVE campaigns, our sample skews toward individuals most likely to support or join extremist organizations in Bangladesh (see Table S4). The sample is young (median 27 years, the same as the Bangladeshi median) and overwhelmingly Muslim (89%, compared with 91% for the country). It is also disproportionately male (76%), educated (37% have a college degree), married (53%), and urban (41% live within 40 km of Dhaka, the capital city). Nearly all (85%) respondents report using Facebook a few days a week or more.

As Table S5 shows, the age, gender, and educational composition of the survey experiment sample resembles Facebook users in the nationally representative survey of Bangladesh conducted by IRI in 2020, the best available data on users of the platform there (63% aged 18–30 in both samples, 76% male in our sample vs. 75% among Facebook users in the IRI survey, 37% college graduates vs. 19% in the IRI survey). The results from the survey experiment are thus comparable to the Facebook study and should generalize to the population that would see the videos on the platform under real-world conditions.

We first measured respondent demographics and baseline attitudes. We then randomly assigned each with equal probability to watch one of our three CVE videos or a placebo video about climate change from a local Bangladeshi television station. We aimed for a naturalistic placebo rather than a pure control since it better reflected the content found on a typical user’s Facebook feed. The video was the same length as our treatment videos.

Next, we asked questions about their emotional state before turning to questions about religious outgroups and intergroup trust. We measured opposition to extremism by asking about the likelihood they would engage in nine pro- or antiextremist actions: four proextremist actions (attending a public Hefazat or pro-ISIS event and sharing a pro-Hefazat or a pro-ISIS post on Facebook) and five antiextremist actions (calling a hotline to report extremists and attending a public event or sharing a social media post of the Awami League or an antiextremist youth organization). The English version of our instrument is in the Supplementary information.c

To test the effectiveness of our CVE videos against counter-messaging, we randomly assigned respondents with equal probability to view a brief 20 s video covering violent protests against Modi’s visit in 2021 or a placebo video. The counter-message video showed a protester at a Hefazat-organized demonstration denouncing the government, India, and Modi in the name of Bangladeshi Muslims—common features of VEO propaganda in Bangladesh (Table S3). Our placebo video was a brief 20 s news report on artificial intelligence that mimicked content found on an average user’s Facebook feed. This content represents a more naturalistic placebo than a pure control (ie no video). After the videos, we asked respondents about attitudes toward religious in- and outgroups and their likelihood of engaging in pro- and antiextremist behaviors.

Participants in the placebo condition fell below the midpoint of the scale used to measure the prosocial emotions of hope and pride (3.20 on a 1–7 scale). They also landed somewhat above the midpoint of the scales used to measure attitudes toward religious outgroups and opposition to extremism (means of 3.02 on a 1–4 scale and 3.30 on a 1–5 scale, respectively; see Table S7).

Ethics

To ensure that the research was conducted in an ethically appropriate manner, our study was reviewed and approved before fielding by the Dartmouth Committee for the Protection of Human Subjects (CPHS). We addressed the challenge of conducting ethical research on violent extremism in several ways. First, all videos were carefully constructed to oppose violent extremism. We also observed reactions to our messages in focus groups and among a small subset of Facebook users to check for adverse reactions before scaling up. Facebook responses, as measured by emojis, were overwhelmingly positive in nature. Second, we obtained informed consent from all survey respondents. Participant interactions with the Facebook ads were determined to be minimal risk as defined in the Common Rule 45 CFR 46.102(j); we received a waiver of consent for measuring those interactions. Third, we deliberately chose not to use actual messages from extremist groups in the counter-messaging condition to avoid causing unintended harm to participants. Instead, we selected a brief clip of a widely televised protest in which extremist slogans were chanted to serve as a plausible stand-in, meeting the minimal risk standard defined in the Common Rule (36). Finally, all survey participants who completed the survey were debriefed that “the purpose of the study was to learn about the effects of videos intended to promote nonviolence in Bangladesh” and asked to read the script of one of the CVE videos they had not previously seen, which we introduced by stating “please read the text below about the importance of rejecting violence.”

Analysis

For our Facebook study, we estimate weighted least squares (WLS) models for data from the clustered condition weighted by impressions per cluster (34).d Each cluster model includes indicators for age group, sex, marital status, education, urban proximity, and impressions to improve precision and is estimated using robust SE. We did not include number of impressions as a preregistered control variable; results are similar with it excluded. We estimate the average treatment effect (ATE) by running a WLS regression of Y on T, and demeaned covariates interacted with T using the lm_lin command in the estimatr package in R (37, 38).e We consider engagement rate, click rate, comment rate, and share rate separately and regress these outcomes on indicator variables for each video. Because we have no control condition in the Facebook study, we treat the costs of terrorism video as the baseline and estimate differences in engagement measures relative to it. We also estimate differences in engagement between the communal harmony and peaceful religions videos directly. Controlling for other covariates or estimating the lm_lin model is not possible in the split condition because outcomes are reported at an aggregate level; we simply run a WLS model controlling for age and impressions. For clarity, we only report the results from the cluster design below. Split test results, which are consistent with those reported below, appear in the Supplementary information (Tables S16–S17 and Fig. S4).

We analyze our survey data using ordinary least squares (OLS) regression with robust SE to estimate the treatment effects of exposure to each video relative to a placebo condition. A lasso selection procedure is employed to determine which pretreatment covariates to include in each model (39). Pretreatment covariates mirror those used for the Facebook study and include age, gender, marital status, education, and region, along with pretreatment measures of outgroup trust and pro/anti-extremism. We also include an indicator variable for exposure to the protest video. We separately estimate models with an interaction between treatment video and protest video exposure to estimate how exposure to the protest video moderates the effects of CVE video exposure. Our Supplementary information provides additional technical details.

Results

In total, our CVE videos on Facebook generated 39.8 million impressions across two waves of data collection (2022 April 23–30 and May 25–June 7).f Engagement was also high: across all three videos, the mean engagement rate (a composite of liking/reacting, clicking, commenting, viewing, and sharing) was 3,610 per 10,000 impressions in the cluster design and 3,306 per 10,000 impressions in the split design. We provide additional descriptive tables and figures for each outcome measure across the cluster and split designs in the Supplementary information.

We present our main results in Fig. 1.

Figure 1.

Figure 1 illustrates the effects of our three CVE messages on Facebook outcomes (left panel) relative to the costs of terrorism condition and on survey outcomes (right panel) relative to a placebo condition. We compare Facebook outcomes across two waves of our study (during and after Ramadan). Our survey outcomes include measures for hope and pride, attitudes toward the outgroup, and opposition to extremism.

CVE video treatment effects on Facebook engagement (left panel) and prosocial attitudes (right panel). Estimates from Tables S8 and S20.

Our preregistered hypothesis that the “peaceful religions” message would generate the highest Facebook engagement and greatest improvement in prosocial attitudes and antiextremism is not supported. Instead, appeals to communal harmony outperformed peaceful religions and costs of terrorism messages in Facebook engagement (Fig. 1, left panel) and attitudinal measures on our survey (Fig. 1, right panel) despite our first wave being conducted during Ramadan.

We begin with the results from the Facebook study. The communal harmony video generated significantly higher levels of overall engagement than the costs of terrorism video (d=0.53, P<0.005 in wave 1; d=0.49, P<0.005 in wave 2). By contrast, the peaceful religions messaging actually recorded lower overall engagement relative to the costs of terrorism baseline message (d=−0.12, P<0.005 in wave 1; d=−0.09, P<0.05 in wave 2). As a result, the communal harmony video also significantly increased engagement vs. the peaceful religions video (d=0.64, P<0.005 in wave 1; d=0.59, P<0.005 in wave 2; results disaggregated by engagement type are provided in Table S8). We observe similar results in waves of our split test sample (see Tables S16, S17, and Fig. S4).

The magnitude of these effects is substantial. To illustrate using pooled data across waves, the mean cluster-level engagement rate was 3,528 engagements per 10,000 impressions for the baseline costs of terrorism message. The estimated treatment effect of viewing the communal harmony video instead increases the engagement rate to 4,088 per 10,000 impressions (an increase of 0.51 SD). Engagement declines slightly on average for those viewing the peaceful religions video, decreasing to 3,408 engagements per 10,000 impressions (a decrease of 0.11 SD relative to the costs of terrorism message).g

The communal harmony message also generated more frequent positive emojis than the other videos, suggesting that these differences in engagement reflected differences in positive sentiment. Specifically, the communal harmony video received 133 “likes” and 44 “loves” per 10,000 impressions compared to 107 and 25, respectively, for peaceful religions and 74 and 15, respectively, for costs of terrorism (see Tables S11, S13, and S14; these differences are statistically significant). Notably, the videos provoked very few “angry” or “sad” emojis, suggesting a lack of harmful backlash.

We now turn to results from the online survey experiment. The right panel in Fig. 1 presents estimated treatment effects of each video on feelings of hope and pride, attitudes toward religious outgroups, and opposition to extremism relative to the placebo condition in the survey experiment. These main effect estimates average over randomized exposure with P=0.5 to a subsequent protest videos, which we control for in our models along with prognostic covariates (see Table S20 for full results). All videos increased feelings of hope and pride, though the effects are again greatest with the communal harmony video (d=0.90 vs. placebo). None of the videos measurably affected attitudes toward religious outgroups targeted by extremist organizations, but we do find small positive effects of the communal harmony video on our combined measure of opposition to extremism, which averages the self-reported likelihood of engaging in anti- and proextremist actions with the latter reverse-coded (d=0.04, P<0.05). By contrast, the effects of the other two messages are not statistically significant, though we cannot directly reject the nulls of no difference vs. the communal harmony video.h

Exposure to counter-messages

Our design was constructed to test whether the effects of CVE messages are robust to respondent exposure to extremist counter-messages. In the real world, citizens live in complex media environments in which they are exposed to multiple political narratives. As a result, the main effects reported above for the three CVE videos represent an average of treatment effects between respondents who were shown a 20 s protest clip and those who were not. However, these counter-messages may alter the effects of CVE messages on respondent attitudes. We therefore estimate models of attitudes toward religious outgroups and opposition to extremism that interact each treatment with an indicator denoting whether an individual watched the protest video afterwards (40). The resulting estimates, which are presented in Fig. 2, represent the treatment effects of each CVE video with and without subsequent exposure to the protest video for attitudes toward religious outgroups (left panel) and opposition to extremism (right panel). Full results are in Table S21.i

Figure 2.

Figure 2 consists of two panels: the effects of our three treatment videos on outgroup attitudes relative to a baseline placebo (left panel) and on opposition to extremism (right panel). We compare the effects of each video with and without exposure to a message that approximates extremist messaging. Estimates are derived from the full survey sample.

Treatment effects on outgroup attitudes (left panel) and opposition to extremism (right panel) with or without exposure to counter-message and in the full sample. Estimates from Table S21.

Watching the protest video had no measurable effect on attitudes toward religious outgroups or opposition to extremism in the placebo condition (see Table S21). It also did not significantly moderate the effect of the CVE videos on attitudes toward outgroups.

However, as the right panel of Fig. 2 suggests, exposure to the protest video moderated the effects of the peaceful religions and costs of terrorism videos on opposition to extremism (see Table S21 for full results). While the peaceful religions and costs of terrorism videos increase opposition to extremism among respondents not exposed to the protest video (d=0.09, P<0.005 and d=0.07, P<0.01, respectively), watching it significantly reduced these effects (P<0.005 for both), driving them to null among respondents in the protest condition.

Similarly, the effects of the communal harmony video on opposition to extremism are also significant among respondents who did not watch the protest video (d=0.07, P<0.05) and null among those who did watch it. However, we cannot reject the null hypothesis that protest video exposure does not moderate the effect of the communal harmony video (unlike for the other two CVE videos). As a result, the pooled estimate of the effect of the communal harmony video on opposition to extremism reported in Table S20 (which averages over respondents who did and did not see the protest video) is the only one that is positive and significant (d=0.04, P<0.05).

To aid in interpretation of these findings, we use a bootstrap approach to estimate the average treatment effect of the CVE videos on attitudes toward outgroups and opposition to extremism as exposure to the protest video in the estimation sample increases. We vary P from 0.01 to 0.99 and compute the average treatment effect of each video at each level for a population in which respondents are exposed with probability P to the protest video. Figure 3 includes 95% confidence intervals computed using 1,000 nonparametric bootstrap resamples (see SI Section 12.2). This procedure allows us to determine how the results vary between Fig. 1, which averages over protest video exposure (P=0.5), and Fig. 2, which separately estimates treatment effects by subsequent protest video exposure (ie P=0 or P=1). The results, which are presented in Figs. S5 and S6, suggest that the positive effects of the communal harmony video on opposition to extremism are less sensitive to protest video exposure than are the peaceful religions and costs of terrorism videos.

Figure 3.

This figure illustrates the treatment effect of our three videos (communal harmony, peaceful religions, and costs of terrorism) on anti-extremist behavioral intentions relative to a baseline placebo condition. We plot three values for each video: the estimated effect without a protest video being shown, the overall effect, and the estimated effect after a protest video has been shown. All three videos increase anti-extremist behavioral intentions. The communal harmony video proved the most robust to possible extremist counter-programming, especially when compared with the costs of terrorism video.

Treatment effects on antiextremist behavioral intentions with or without exposure to counter-message and in the full sample. Estimates from Table S25.

Disaggregating net opposition to extremism

To further investigate these results, we conduct an exploratory analysis investigating whether increases in net opposition to extremism can be attributed to reduced intent to engage in proextremist behaviors or increased willingness to engage in antiextremist behaviors. To do so, we disaggregate our combined opposition to extremism measure into intentions to engage in antiextremist and proextremist behaviors and again estimate main effects and conditional effects by protest video exposure.

The main effects results, which are summarized in Fig. 3 (“Overall”), show that the communal harmony and peaceful religions messages elicited increased intentions to engage in antiextremist behaviors (d=0.04, d=0.04, P<0.05 for each) but did not measurably change intentions to engage in proextremist behavior (full results for each separate item are provided in Table S24). We also conduct principal component analysis for all items in these indices (see Tables S29–S33).

When we disaggregate the results, we observe that all three videos measurably increase intention to engage in antiextremist behaviors while having no effect on proextremist behavioral intentions in the absence of protest video exposure. However, we again observe a statistically significant negative interaction between the protest video and the costs of terrorism video that drives the effect of the latter to zero (P<0.05; see Table S25). The negative interaction is similar, though not quite significant, for peaceful religions. As a result, we cannot reject the null of no effect for any CVE video tested when they were followed by the protest videos. However, the communal harmony video performs best, measurably outperforming the costs of terrorism video (d=0.05, P<0.05), which performs worst.

Discussion

Despite the threat posed by VEOs, the messaging campaigns that are central to many CVE programs have rarely been rigorously evaluated. Our study represents an initial attempt to place the evaluation of CVE campaigns on a more rigorous foundation. We offer at least three contributions to scientific knowledge on CVE.

First, our findings reinforce the small set of existing studies that find religious authorities can be effective endorsers of CVE messaging. We extend this literature, however, by testing two nonreligious messages, including a communitarian message that outperformed religious and costs of terrorism messages. Moreover, we uncover important differences across messages despite holding religious authorities constant: not all messages were equally well-received, with our costs of terrorism appeal failing to shift views despite being endorsed by religious authorities. Unlike Ref. (21), we find very little evidence of attitudinal blowback among viewers, suggesting that naturalistic treatments can be fielded at scale without inadvertently undermining CVE efforts. We also discovered that not all messages are equally robust to counter-messaging. Our communitarian message proved more resilient than its alternatives, a finding that raises a host of new questions about how to design messages for competitive media environments. In short, communitarian messages, though neglected to date in the existing literature, can be an effective CVE tool when designed by locals and delivered by religious authorities.

Given the nature of our research design, we can only speculate about why we observe differences in the effectiveness of our three messages. We offer two interpretations. First, it is plausible that the communitarian message outperformed our religious video because it invoked a wider set of social enforcement mechanisms. In particular, our message may have prompted feelings of belonging and obligation and, in turn, raised the specter of reputational loss and possible social exclusion for failing to defend shared values. We note, for example, that our communitarian message elicited the most positive emoji reactions on Facebook and the highest levels of prosocial emotions like hope and pride (Fig. 1). By contrast, our religious message invokes a potentially weaker moral claim that is more procedural (“peaceful coexistence”) than stirring. Second, it is possible that subtle differences in the videos themselves could have shaped their effectiveness. Our costs of terrorism message, for example, might have inadvertently struck a more negative tone due to its use of longer, and more violent, animation, which produced lower levels of prosocial emotions. By highlighting the dangers associated with VEOs, it may have suppressed, rather than encouraged, opposition to extremism.

These findings suggest several avenues for future research. Scholars should, for example, test the mechanisms that underpin the effectiveness of CVE messages, the conditions under which counter-messaging alters these effects, and why proextremist behavioral intentions appear more difficult to change. Our RCT framework could also be extended to consider how varying an endorser’s identity to include political leaders, athletes, and celebrities might change the reception of these messages. It is possible, for example, that our costs of terrorism video might have performed better if endorsed by government officials, police, or victims. Deliberate engagement with counter-programming by VEOs, though subject to ethical constraints, should also be pursued to understand the resilience of different messages. Moreover, our treatments were single-shot in nature; we cannot speak to the durability of these effects or how they might vary with repeated exposure. Future work should consider designs with multiple exposure to messages as well as repeated measurement over time. Similarly, as with all other CVE evaluations, we cannot directly measure real-world behavior, an important challenge for future research.

Our multiplatform design has several advantages over the single-platform, often nonexperimental, designs that currently dominate the study of online CVE campaigns. Most notably, our paired Facebook and survey experiment allowed us to combine real-world engagement data with rigorous outcome evaluations of survey outcomes. By drawing on multiple platforms and measures, we provide a template that can be modified and extended to assess the comparative efficacy of different CVE messages at reasonable costj across multiple platforms used by VEOs, including TikTok, Telegram, X, and Instagram.

In addition, we move beyond existing CVE outcome metrics by considering pro- and antiextremist behavioral intentions. Our exploratory results suggest that CVE messages may be more effective at rallying opposition to extremism than at undermining support for it. This nuanced framework suggests that (video) treatments should be targeted at influencing an individual’s calculations about taking action rather than focusing on already-radicalized individuals. Subsequent research might also experiment with different treatments designed to encourage, and perhaps model, the most efficacious strategies for engaging VEOs safely.

Finally, we offer some tentative lessons for practitioners. Communitarian messages, for example, should be added to the CVE toolkit. Video messages also offer a low-cost and scalable intervention that can supplement, or even replace, existing radio- and survey-based treatments. We also note the existence of a tension between naturalistic and universal treatments. We opted for naturalistic treatments that would prove engaging to online respondents. Doing so, however, meant introducing stylistic differences in the videos that might have contributed to their differential reception. Shorter messages, or imposing tighter uniformity on form and content to prevent blurring of treatment conditions, might prove valuable in future research even if these steps sacrifice some engagement.

There are, of course, limits to single country research designs. Determining the relative efficacy and generalizability of these messages, and whether communitarian messages outperform more traditional messages, will require cross-national testing. We note, however, that Bangladesh shares several traits thought to foster violent extremism as other countries facing VEO threats, including Indonesia, Pakistan, India, and Nigeria. These conditions include similar (high) levels of income inequality, weak democracy, young populations, government crackdowns, and religious authorities that command widespread trust. While the specific content of these messages will vary by context, an emphasis on shared identity, a common threat, and a duty to protect the political community may resonate in Bangladesh and beyond.

Supplementary Material

pgag142_Supplementary_Data

Acknowledgments

This research would not have been possible without the help of many individuals and organizations. At the International Republican Institute (IRI), we thank Bret Barrowman, Geoffrey Macdonald, Rohullah Niazi, and Saimum Parvez. At Beatnik Digital, we thank Mufrad Mahmud Choudhury, Samia Kader, Samir Muhammad Korban, and Tanjir Bin Jamil. At TGM Research, we thank Luiz Gustavo Pelegrini Marongio and Ronen Singha. At Meta, we thank Michael Zoorob, Andrew Gruen, and Rachel Davis Mersey. We also thank our research assistants at Dartmouth, including Ahnaf Aronno, Benjamin Aronson, Mimi Majumder, Adelia Rattray, Hannah Tanenbaum, and Ethan Weber. Finally, we thank Alicia Cooperman, Alex Coppock, Blair Read, and participants in the Experiments in Governance and Politics conference at Cornell University for helpful comments. This article was checked for errors by Claude Opus 4.6 and 4.7, Google Gemini 3.1 Pro, and OpenAI’s ChatGPT 5.2 Pro. The findings and opinions stated here do not reflect the official position or policy of IRI or the US government.

Notes

a

We define a VEO as a group that advocates and uses violence for political, religious, or social goals that involve challenging or overthrowing existing governments, institutions, and social norms. Some have global ambitions; others are highly localized. All, however, face the twin challenges of recruiting followers and wielding violence in the service of their goals.

b

See Figs. S2 and S3 for cluster-level engagement rates for each treatment.

c

We plot the distribution of survey outcomes across our three treatments and placebo in Fig. S1.

d

Forty-two out of 1,152 total ads in the cluster condition received 0 impressions. Because our WLS models weight by the number of impressions each cluster received, these ads receive no weight in our analyses. However, we exclude them from the descriptive statistics provided in the Supplementary information.

e

We pre-registered including estimated cluster size S (the variable on which we blocked) as a covariate. We instead control for the demographic indicators that define the clusters and the number of impressions.

f

Wave 1 cluster data were downloaded 2022 May 14–17; all other data were downloaded 2022 June 21–23. Because “reach” (the number of unique users exposed to the videos) is calculated within (rather than across) campaigns, we are unable to measure the total number of unique Facebook users who were exposed to the ads.

g

We explore heterogeneous effects across urban populations and young men in Tables S18 and S19.

h

In an exploratory analysis, we also find no evidence of meaningful differences across the videos when survey respondents are asked if they “would like to learn more about antiextremist activities” (see Tables S26 and S27). We also test for heterogeneous effects across urban and highly educated populations as well as young men in Tables S22 and S23.

i

Given that actual VEO messaging is likely more intense than our 20 s counter-messaging clip, our nonattenuation finding on the communal harmony treatment is suggestive, not definitive. We note, too, that we chose not to test the effects of exposure to a VEO message alone (rather than a counter-message video from a protest) due to ethical concerns.

j

We generated nearly 40 million impressions on Facebook for under $19,000 USD.

Contributor Information

Jason Lyall, Department of Government, Dartmouth College, Hanover, NH 03755, USA.

Brendan Nyhan, Department of Government, Dartmouth College, Hanover, NH 03755, USA.

Elsa Voytas, Independent Researcher, Brooklyn, NY 11217, USA.

Supplementary material

Supplementary material is available at PNAS Nexus online.

Competing interest

B.N. previously received research funding from Meta for an unrelated project. The research reported in this article was conducted before E.V. became employed by Meta. J.L. declares no competing interest.

Funding

We gratefully acknowledge the financial support of the Wright Center for the Study of Computation and Just Communities at Dartmouth. IRI funded the development of the Facebook videos and contributed to the second wave of Facebook ad buys. J.L. also thanks the Carnegie Corporation of New York (R-F-20-57728) and B.N. thanks the John Simon Guggenheim Memorial Foundation.

Author contributions

J.L., B.N., and E.V. designed the research and wrote the article. J.L. coordinated with partner organizations and supervised the stimuli development. E.V. conducted the Facebook experiments and was the lead data analyst. B.N. supervised the development and fielding of the survey experiment.

Preprints

This manuscript was posted on a preprint: https://osf.io/vb489/files/3c4ab.

Data availability

Our design, hypotheses, survey instrument, and analyses were preregistered with the Open Science Framework. Our Facebook study was preregistered on 2022 April 23 and updated on 2022 May 26 (https://osf.io/hjb56/). Our survey was preregistered separately in the same OSF project on 2022 August 18 and updated on 2022 August 28 (https://osf.io/4udq7/). We identify all deviations from the preanalysis plan and note whether our analyses are preregistered or exploratory. Our study was approved by Dartmouth’s Committee for the Protection of Human Subjects (#00032496). Replication data and materials can be found on the Harvard Dataverse: https://doi.org/10.7910/DVN/8OYWEV.

References

  • 1. Jones  M, Wallner  C, Winterbotham  E. 2021. Lessons for preventing and countering violent extremism: an evidence-based approach. RUSI Occasional Papers.
  • 2. United States Agency for International Development . 2020. Policy for countering violent extremism through development assistance. USAID, Washington, D.C.
  • 3. Stephens  W, Sieckelinck  S, Boutellier  H. 2021. Preventing violent extremism: a review of the literature. Stud Confl Terror. 44:346–361. [Google Scholar]
  • 4. Grip  L, Kotajoki  J. 2019. Deradicalisation, disengagement, rehabilitation and reintegration of violent extremists in conflict-affected contexts: a systematic literature review. Confl Secur Dev. 19:371–402. [Google Scholar]
  • 5. Jones  M, Winterbotham  E. 2020. Research methodology: the prevention project. RUSI Occasional Papers.
  • 6. Lapowsky  I. 2018. Viral political ads may not be as persuasive as you think. Wired, 2018 Aug 7. Downloaded 2022 Oct 13, from https://www.wired.com/story/viral-political-ads-not-as-persuasive-as-you-think/
  • 7. Messing  S. 2020. A correlational analysis confirmed it all. Tweet posted 2020 Dec 14. Downloaded 2022 Oct 13, from https://twitter.com/SolomonMg/status/1338563542308491268?s=20&t=V8pD0X54QFY7Y_lvwQPMvA
  • 8. Bodine-Baron  E, Marrone  J, Helmus  T, Schlang  D. 2020. Countering violent extremism in Indonesia: using an online panel survey to assess a social media counter-messaging campaign. RAND. [Google Scholar]
  • 9. van den Bos  K. 2020. Unfairness and radicalization. Annu Rev Psychol. 71:563–588. [DOI] [PubMed] [Google Scholar]
  • 10. Fair  C, Hamza  A, Heller  R. 2017. Who supports suicide terrorism in Bangladesh? What the data say?  Politics Relig. 10:622–661. [Google Scholar]
  • 11. Hasan  M, Macdonald  G. 2022. The persistent challenge of extremism in Bangladesh. United States Institute of Peace Special Report No. 511. https://www.usip.org/publications/2022/06/persistent-challenge-extremism-bangladesh.
  • 12. South Asia Terrorism Portal . 2025. Datasheet - Bangladesh. https://www.satp.org/datasheet-terrorist-attack/fatalities/bangladesh.
  • 13. Fair  C, Patel  P. 2022. Support for domestic Islamist terrorism in Bangladesh: insights from a novel survey. Politics Relig. 15:673–699. [Google Scholar]
  • 14. International Republican Institute . 2023. National Survey of Bangladesh. Technical Report, Center for Insights in Survey Research, March–April.
  • 15. Parvez  S. 2019. The Khilfah’s soldiers in Bengal: analysing the Islamic state Jihadists and their violence justification narratives in Bangladesh. Perspect Terrorism. 13:22–38. [Google Scholar]
  • 16. Wolf  S. 2022. The Hefazat-e-Islam and the Islamist challenge in Bangladesh. Research Report No. 8, South Asia Democratic Forum, July.
  • 17. Siegel  A, Badaan  V. 2020. #no2sectarianism: experimental approaches to reducing sectarian hate speech online. Am Polit Sci Rev. 114:837–855. [Google Scholar]
  • 18. Blair  G, et al.  2022. Trusted authorities can change minds and shift norms during conflict. Proc Natl Acad Sci U S A. 118:1–6. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19. Condra  L, Isaqzadeh  M, Linardi  S. 2019. Clerics and scriptures: experimentally disentangling the influence of religious authority in Afghanistan. Br J Polit Sci. 49:401–419. [Google Scholar]
  • 20. Masoud  T, Jamal  A, Nugent  E. 2016. Using the Qur’an to empower Arab women? Theory and experimental evidence from Egypt. Comp Polit Stud. 49:1555–1598. [Google Scholar]
  • 21. Grossman  A, Nomikos  W, Siddiqui  N. 2022. Can appeals for peace promote tolerance and mitigate support for extremism? Evidence from an experiment with adolescents in Burkina Faso. J Exp Polit Sci. 10:124–136. [Google Scholar]
  • 22. Bilali  R. 2022. Fighting violent extremism with narrative intervention: evidence from a field experiment in west Africa. Psychol Sci. 33:184–195. [DOI] [PubMed] [Google Scholar]
  • 23. Blair  G, Littman  R, Paluck  E. 2019. Motivating the adoption of new community-minded behaviors: an empirical test in Nigeria. Sci Adv. 5:eaau5175. 10.1126/sciadv.aau5175 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 24. Mitts  T, Phillips  G, Walter  B. 2022. Studying the impact of isis propaganda campaigns. J Polit. 84:1220–1225. [Google Scholar]
  • 25. Feinberg  M, Willer  R, Kovacheff  C. 2020. The activist’s dilemma: extreme protest actions reduce popular support for social movements. J Pers Soc Psychol. 119:1086–1111. [DOI] [PubMed] [Google Scholar]
  • 26. UNDP Arab States, Nudge Lebanon, and B4Development . 2021. Applying behavioral science to support the prevention of violent extremism: experiences and lessons learned. UNDP, Washington, D.C.
  • 27. Green  S, Proctor  K. 2016. Turning point: a new comprehensive strategy for countering violent extremism. Technical Report, CSIS, Washington, D.C.
  • 28. Goldberg  M, Gustafson  A, Rosenthal  S, Leiserowitz  A. 2021. Shifting republican views on climate change through targeted advertising. Nat Clim Chang. 11:573–577. [Google Scholar]
  • 29. Levy  R. 2021. Social media, news consumption, and polarization: evidence from a field experiment. Am Econ Rev. 111:831–870. [Google Scholar]
  • 30. Asimovic  N, Nagler  J, Bonneau  R, Tucker  J. 2021. Testing the effects of Facebook usage in an ethnically polarized setting. Proc Natl Acad Sci U S A. 118:1–9. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31. Breza  E, et al.  2021. Effects of a large-scale social media advertising campaign on holiday travel and COVID-19 infections: a cluster randomized controlled trial. Nat Med. 27:1622–1628. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32. Speckhard  A, Ellenberg  M. 2020. Breaking the ISIS brand: counter narrative Facebook campaigns in Europe. J Strat Sec. 13:120–148. [Google Scholar]
  • 33. Ryan  TJ. 2012. What makes us click? Demonstrating incentives for angry discourse with digital-age field experiments. J Polit. 74:1138–1152. [Google Scholar]
  • 34. Ryan  TJ, Brader  T. 2017. Gaffe appeal: a field experiment on partisan selective exposure to election messages. Polit Sci Res Methods. 5:667–687. [Google Scholar]
  • 35. Broockman  DE, Green  DP. 2014. Do online advertisements increase political candidates’ name recognition or favorability? Evidence from randomized field experiments. Polit Behav. 36:263–289. [Google Scholar]
  • 36. HHS.gov. 2018. 45 cfr 46. Downloaded 2022 December 13 from https://www.hhs.gov/ohrp/regulations-and-policy/regulations/45-cfr-46/index.html.
  • 37. Lin  W. 2013. Agnostic notes on regression adjustments to experimental data: reexamining Freedman’s critique. Ann Appl Stat. 7:295–318. [Google Scholar]
  • 38. Blair  G, et al.  2018. Package ‘estimatr’. Stat. 7:295–318. [Google Scholar]
  • 39. Bloniarz  A, Liu  H, Zhang  C-H, Sekhon  JS, Yu  B. 2016. Lasso adjustments of treatment effect estimates in randomized experiments. Proc Natl Acad Sci U S A. 113:7383–7390. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 40. Muralidharan  K, Romero  M, Wüthrich  K. 2019. Factorial designs, model selection, and (incorrect) inference in randomized experiments. Technical Report, National Bureau of Economic Research.

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

pgag142_Supplementary_Data

Data Availability Statement

Our design, hypotheses, survey instrument, and analyses were preregistered with the Open Science Framework. Our Facebook study was preregistered on 2022 April 23 and updated on 2022 May 26 (https://osf.io/hjb56/). Our survey was preregistered separately in the same OSF project on 2022 August 18 and updated on 2022 August 28 (https://osf.io/4udq7/). We identify all deviations from the preanalysis plan and note whether our analyses are preregistered or exploratory. Our study was approved by Dartmouth’s Committee for the Protection of Human Subjects (#00032496). Replication data and materials can be found on the Harvard Dataverse: https://doi.org/10.7910/DVN/8OYWEV.


Articles from PNAS Nexus are provided here courtesy of Oxford University Press

RESOURCES