Prebunking interventions based on “inoculation” theory can reduce susceptibility to misinformation across cultures

This study finds that the online “fake news” game, Bad News, can confer psychological resistance against common online misinformation strategies across different cultures. The intervention draws on the theory of psychological inoculation: analogous to the process of medical immunization, we find that “prebunking,” or preemptively warning and exposing people to weakened doses of misinformation, can help cultivate “mental antibodies” against fake news. We conclude that social impact games rooted in basic insights from social psychology can boost immunity against misinformation across a variety of cultural, linguistic, and political settings.


Essay summary
• We designed an online game in which players enter a fictional social media environment.In the game, the players "walk a mile" in the shoes of a fake news creator.After playing the game, we found that people became less susceptible to future exposure to common misinformation techniques, an approach we call prebunking.• In a cross-cultural comparison conducted in collaboration with the UK Foreign and Commonwealth Office and the Dutch media platform DROG, we tested the effectiveness of this game in 4 languages other than English (German, Greek, Polish, and Swedish).• We conducted 4 voluntary in-game experiments using a convenience sample for each language version of Bad News (n = 5,061).We tested people's assessment of the reliability of several fake and "real" (i.e., credible) Twitter posts before and after playing the game.
• We find significant and meaningful reductions in the perceived reliability of manipulative content across all languages, indicating that participants' ability to spot misinformation significantly improved.Relevant demographic variables such as age, gender, education level, and political ideology did not substantially influence the inoculation effect.• Our real-world intervention shows that social impact games rooted in insights from social psychology can boost psychological immunity against online misinformation across a variety of cultural, linguistic, and political settings.• Social media companies, governments, and educational institutions could develop similar largescale "vaccination programs" against misinformation.Such interventions can be directly implemented in educational programs, adapted for use within social media environments, or applied in other issue domains where online misinformation is a threat.
• In contrast to classical "debunking," we recommend that (social media) companies, governmental, and educational institutions also consider prebunking (inoculation) as an effective means to combat the spread of online misinformation.

Implications
There is a loud call for educational interventions to help citizens navigate credible, biased, and false information.According to the World Economic Forum (2018), online misinformation is a pervasive global threat.UNESCO underscored how all citizens need better up-to-date knowledge, skills, and attitudes to critically assess online information (Carlsson, 2019).
Common approaches to tackling the problem of online misinformation include developing and improving detection algorithms (Monti et al., 2019), introducing or amending legislation (Human Rights Watch, 2018), developing and improving fact-checking mechanisms (Nyhan & Reifler, 2012), and focusing on media literacy education (Livingstone, 2018).However, such interventions present limitations.In particular, it has been shown that debunking and fact-checking can lack effectiveness because of the continued influence of misinformation: once people are exposed to a falsehood, it is difficult to correct (De keersmaecker & Roets, 2017;Lewandowsky et al., 2012).Overall, there is a lack of evidence-based educational materials to support citizens' attitudes and abilities to resist misinformation (European Union, 2018; Wardle & Derakshan, 2017).Importantly, most research-based educational interventions do not reach beyond the classroom (Lee, 2018).
Inoculation theory is a framework from social psychology that posits that it is possible to pre-emptively confer psychological resistance against (malicious) persuasion attempts (Compton, 2013;McGuire & Papageorgis, 1961).This is a fitting analogy, because "fake news" can spread much like a virus (Kucharski, 2016;Vosoughi et al., 2018).In the context of vaccines, the body is exposed to a weakened dose of a pathogen-strong enough to trigger the immune system-but not so strong as to overwhelm the body.The same can be achieved with information by introducing pre-emptive refutations of weakened arguments, which help build cognitive resistance against future persuasion attempts.Meta-analyses have shown that inoculation theory is effective at reducing vulnerability to persuasion (Banas & Rains, 2010).
Traditional inoculation research tends to focus on individual arguments against specific persuasion or disinformation attempts (McGuire, 1961;van der Linden et al., 2017).In a significant departure from this paradigm, we have already shown that, using a game-based intervention, participants can also improve in their latent ability to spot misinformation techniques as opposed to just individual instances of misinformation (see Basol et al., 2020;Roozenbeek & van der Linden, 2018, 2019).The main premise of our approach is that fake news stories themselves constantly change and evolve so building immunity against the underlying tactics of misinformation is a more durable strategy.
The research we present here focuses on an "active" psychological inoculation intervention called Bad News.Bad News is a free choice-based browser game in which players take on the role of fake news creators and learn about 6 common misinformation techniques2 .In our experiments, we showed that the game is effective at building resistance against misinformation (Basol et al., 2020;Roozenbeek & van der Linden, 2019) 3 .The Bad News game was widely covered in the media and has won several awards (BBC News, 2018a;CNN, 2019;Reuters, 2019) 4 .Roozenbeek and van der Linden (2019) co-developed and wrote the game's content and implemented a survey within the game that players can choose to participate in.This initial evaluation, which involved asking participants about the reliability of several "fake" and "credible" Twitter posts pre-and postgameplay, showed consistent and significant inoculation effects.This evidence was presented as part of the parliamentary Inquiry on Fake News in the UK (van der Linden et al., 2018) and-in light of further randomized trials and empirical testing-subsequently referred to as "one of the most sustainable paths to combating fake news" by the European Commission (Klossa, 2019, p.23).
In a continuation of the Bad News project, we collaborated with DROG, a Dutch media literacy platform, as well as the United Kingdom's Foreign and Commonwealth Office (FCO), to translate and adapt Bad News in a variety of languages in order to promote media literacy worldwide.The purpose of this collaboration was to scale the Bad News intervention in a variety of languages free for anyone to play and access.Thus far, the intervention has reached about a million people.In the research presented here, we evaluate the cross-cultural effectiveness of the game and find support for its efficacy across a range of different cultural and political contexts.
Practically, our game is used as a digital literacy teaching tool (information for educators is available on the game's website).Importantly, game-based interventions are scalable (i.e. they can reach millions of people worldwide) and offer potential for combating misinformation in other domains as well.For example, we have developed an intervention in collaboration with WhatsApp called "Join this Group".The game inoculates players against misinformation frequently encountered within the context of direct messaging apps, which are a growing problem in countries such as India, Mexico, and Brazil (Phartiyal et al., 2018;Roozenbeek et al., 2019).Similarly, people can be inoculated against the techniques used in extremist online recruitment by preemptively warning and exposing them to weakened versions of these techniques.In collaboration with the Lebanese Behavioral Insights Unit (Nudge Lebanon) we have developed a game-based intervention called Radicalize, which we recently presented at the United Nations (UNITAR, 2019).
However, it is important to note that digital literacy is complex (Nygren, 2019) and inoculation interventions on their own do not offer a silver bullet to all of the challenges of navigating the post-truth information landscape.Future research may investigate how inoculation interventions can be combined with other interventions to support critical thinking (Lutzke et al., 2019), evaluations of authentic news feeds (Nygren et al., 2019), and civic online reasoning (McGrew, 2020).Nonetheless, our results show that this real-world intervention can significantly boost people's ability to recognize online misinformation and improve their resistance against manipulation attempts in a variety of cultural, linguistic, and political settings.

Findings
Finding 1: An online social impact game, Bad News, is effective at conferring resistance against misinformation techniques in Sweden.
As a first step, we developed and evaluated the Swedish-language version of the Bad News game5 using an in-game survey as a cross-cultural pilot test and a direct replication of the English version.Accordingly, the measures and methodology were therefore kept similar to the initial evaluation of the game by Roozenbeek and van der Linden (2019) with three fake news Twitter posts (making use of the impersonation, conspiracy, and discrediting techniques) and 3 "credible" news control items that did not contain any deception techniques ("Brexit, the United Kingdom's exit from the European Union, will happen in 2019", "US President Donald Trump wants to build a wall between the US and Mexico" and "Only one in ten students can distinguish between real news and advertisement"6 ).We showed participants (n = 134 to n = 379 for the full pre-post survey) 3 fake and 3 credible (control) Twitter posts before and after playing the game and asked them to rate their reliability on a 1-7 Likert scale (see the "Methods" section for more details).We hypothesized that playing the game would significantly reduce the perceived reliability of fake (but not credible) items.
We find that playing Bad News significantly reduces participants' susceptibility to simulated fake Twitter posts for impersonation (t(173) = -2.37,p = 0.018, d = 0.18), conspiracy (t(378) = -3.22,p = 0.001, d = 0.17) and discrediting (t(375) = -2.74,p = 0.001, d = 0.15).Averaged across all fake items, the mean reduction in reliability judgments is also significant (t(172) = -3.38,p = 0.001, d = 0.24).On the other hand, people did not reduce or change their ratings of the credible news control items (for which we hypothesized no significant change), including the survey items about Brexit (t(133) = -0.88,p = 0.38), Donald Trump (t(379) = -0.28,p = 0.78), and media literacy (t(243) = -0.35,p = 0.72).These results are consistent with earlier findings by Roozenbeek & van der Linden (2019) and are visualized in Figure 2   shows that the credible news items were perceived as equally reliable before and after playing, whereas the "fake" items are seen as significantly less reliable.All mean pre and post scores can be found in Table 3 in the appendix.
Finding 2: In a cross-cultural sample, the Bad News game also improves people's ability to recognize misinformation strategies in Germany, Greece, and Poland.
The Swedish pilot study demonstrated the potential cross-cultural effectiveness of the Bad News game.The survey items were subsequently extended and standardized across languages to include an item for each of the six manipulation techniques people learn about in the game (impersonation, conspiracy, emotion, polarization, discrediting, and trolling; see Roozenbeek & van der Linden (2019) for a more thorough explanation of these techniques).To corroborate these initial results, we also implemented prepost surveys in the game's German, Greek and Polish versions7 .As such, the exact same six items were administered pre-and post-gameplay in Germany (n = 2,038), Greece (n = 1,518), and Poland (n = 1,332) along with standard socio-demographic questions (age, gender, education, and political ideology).
Following Roozenbeek and van der Linden (2019) and Basol et al. (2020), we created a single fake news scale (averaged across all items) to reduce multiple testing (though item-level results are also presented by country in Tables 4, 5, and 6 in the appendix).This allowed us to compare participants' performance across the three countries.Overall, the aggregate inoculation effect across countries is significant (Mpre = 2.50, Mpost = 2.09, Mdiff = -0.41,[95% CI -0.38, -0.44], (t(4887) = -28.47,p < 0.01, d = 0.37), indicating that the game effectively inoculates players against misinformation techniques across a variety of cultural and linguistic settings8 .The observed effect-size can be described as close to "medium" or "moderate" and is very similar to the English version of Bad News (Roozenbeek & van der Linden, 2019).
Nonetheless, it is possible that the average masks significant between-country variability.A multi-level model revealed an intraclass correlation of just 0.001, which means that the proportion of variance in fake news scores that is accounted for by country is close to zero.In other words, most variability lies within groups (Gelman & Hill, 2007).Accordingly, the likelihood-ratio test indicated a multi-level model was not preferred over a standard linear model (χ2 = 2.59, p = 0.05).A standard Ordinary Least Squares (OLS) model with country as a dummy variable shows that compared to Germany, the average inoculation effect is somewhat (but not much) lower in Greece (β = -0.083,p = 0.01, [95%CI -0.01, -0.15]) and Poland (β = -0.09,p = 0.01, [95%CI -0.02, -0.16]).Figure 3    The importance of education level and cognitive abilities has been noted as crucial to navigating online information in constructive ways (Nygren, 2019;Nygren & Guath, 2019;Pennycook & Rand, 2019).Previous research has also acknowledged a problematic digital divide between groups in society where the productive use of online information is associated with socio-economic status (Scheerder et al., 2017).Accordingly, we conducted a robustness check by controlling for several demographic covariates for the German, Greek, and Polish versions of the game (the sample size for the pilot in Sweden was too small to control for these variables).Adjusting the model for socio-demographics revealed no main effect for either gender or education but a small significant effect of political ideology and age (see Table 7 in the appendix).Adjusting for covariates slightly increased between-country differences for Greece (β = -0.09,p = 0.01, 95%CI -0.02, -0.16) and Poland (β = -0.12,p = 0.01, 95%CI -0.05, -0.20,Table 7, Model 2).
We observe some indications in the pre-test that liberals and people with higher education were a bit more sceptical towards misinformation to begin with.The Polish sample shows the most variation across age, political ideology, and education (see Figures 5-8 in the appendix).For example, conservatives performed slightly better than liberals (meaning: the pre-post-test difference is slightly larger), although this difference is very small and likely explained by conservatives' higher pre-test scores enabling greater learning effects (see appendix for details).On average, we still see that participants with different political ideologies display a significant inoculation effect, highlighting how the intervention may not be hindered by confirmation bias.An exploratory analysis found no significant interaction effects between the intervention, country, and demographics, but since a-theoretical interactions are often underpowered and sample-dependent (Hainmueller et al., 2019) we refrain from further reporting on these here.
These results indicate that the game may work as a broad-spectrum vaccine, as participants with different levels of education, gender, age, and ideologies all learned to better spot misinformation.Nonetheless, it is important to point out that our sample is self-selected and not representative of the respective populations within each country.It is therefore possible that the lack of variation across demographics observed here does not accurately describe the general population.For example, because more liberals opted into the study than conservatives in most countries, it remains unclear to what extent differences in political ideology can be generalized.
We do, however, find a small consistent decrease with age, which is in line with Guess et al. (2019), who found that older people may be more susceptible to online misinformation.Another explanation is that the Bad News game was designed with an audience in mind between 15 and 35 years of age, indicating that older people may not have experienced the same level of cognitive involvement as younger people.A third explanation could be rooted in differences in internet literacy, which our survey did not measure.Considering that older internet users are a diverse group with high variance in internet literacy (Hunsaker & Hargittai, 2018), it is possible that the game is effective among older users with high internet literacy and less so for older users who are less familiar with the internet.Our present analysis is unable to disentangle these potential differences.

Methods
The "Bad News" game In the game, players enter a simulated social media environment and go through 6 scenarios, each representing a different misinformation technique: impersonating people or organizations online; using emotional language to evoke fear or anger; using divisive language to drive groups in society apart (polarization); spreading conspiracy theories; discrediting opponents by using gaslighting and red herrings; and baiting people into responding in an exaggerated manner (trolling).After creating a fictitious "news" site, over the course of about 15 minutes, players must gain followers and build credibility for their site by correctly identifying and making use of these techniques.If their credibility drops to 0 (e.g., by lying too blatantly or using misinformation techniques incorrectly), the game ends.Figure 1 above shows an example of what the gameplay looks like.Although the examples in the game are modeled after real-world events (e.g.impersonating celebrities or politicians online), the game exposes people to severely weakened doses of these techniques in a controlled learning environment (often using humor and purposeful exaggeration, see Figure 1).Consistent with the vaccination metaphor, this process is meant to trigger the "immune system" but not overwhelm it.

The present study: Bad News Intervention
Previous work has shown that inoculation theory can be used to combat various types of online misinformation, for example about conspiracy theories (Banas & Miller, 2013), climate change (van der Linden et al., 2017), and immigration (Roozenbeek & van der Linden, 2018).Crucially, recent work indicates that online games that rely on inoculation theory can be effective at conferring psychological resistance against misinformation strategies rather than just individual examples of misleading information (Roozenbeek & van der Linden, 2019).We have dubbed this approach prebunking.The most well-known example is the online choice-based browser game Bad News (CNN, 2019).The game works by preemptively exposing players to weakened doses of misinformation techniques and combining elements of perspective-taking (stepping into the shoes of someone who is trying to deceive you) and active experiential learning (creating your own media content).By "weakened dose", we mean weakened versions of manipulation techniques in a controlled environment.In other words, "weakened" means strong enough to get people to pay attention (i.e.activate the psychological immune system) but not so convincing as to actually dupe them.We achieved this by 1) using fictional examples throughout the game and 2) by using a combination of humor (Compton, 2018) and extreme exaggeration so that the basic point is still preserved but the risk of duping people is minimized.In the game, players earn 6 "badges", each representing a common misinformation technique (see table 2 in the appendix for all items): impersonating people or groups online (BBC News, 2018b;Goga et al., 2015), using emotional language (Bakir & McStay, 2017;Konijn, 2013), polarizing audiences (Bessi et al., 2016;Melki & Pickering, 2014), spreading conspiracy theories (Lewandowsky et al., 2013;van der Linden, 2013), discrediting opponents and deflecting criticism (A 'Beckett, 2013), and online trolling (Griffiths, 2014;McCosker, 2014).
As part of a collaboration with the United Kingdom's Foreign and Commonwealth Office and the Dutch antimisinformation platform DROG (2018), we translated Bad News into various languages with the goal of building media literacy efforts internationally9 .In addition, the game was also translated to Swedish in a collaboration between Uppsala and Cambridge University.Following the approach laid out by Roozenbeek & van der Linden (2019), we implemented a voluntary in-game pre-post survey to assess players' ability to spot misinformation techniques in each country.We showed participants 6 fake Twitter posts (not featured in the game), each matching one of the techniques mentioned above10 , both before and after playing Bad News.Participants rated the reliability of each post on a 7-point scale.These Twitter posts were designed to be realistic, but not real, for two reasons: 1) to exclude familiarity confounds and simple memory-retrieval of "real" fake news, and 2) to have full experimental control over the items' content.This was important because we wanted to make sure the focus of the survey items was on the relevant technique, not the content.Accordingly, we designed each tweet to embed one specific technique (e.g.impersonation) and not another.Figure 4 shows an example of what this looked like in the game environment.We measured the main effects for each language, and conducted paired t-tests and regression analyses between average pre-and post-reliability scores for each Twitter post.We also measured socio-demographic variables including age, gender, education level, and political affiliation.

Data collection and sample
Data was collected online through voluntary in-game surveys for a period of 8 months (March-October 2019) for the German, Greek, and Polish versions, and for a period of 13 months (September 2018-October 2019) for the Swedish version (the Swedish game was launched earlier as a pilot study).Participants were recruited organically by driving traffic through media reports linking to the game, as well as through promotional activities conducted by the Foreign and Commonwealth Office, Uppsala University, and local collaborating media literacy organizations (see the "funding" section for more details).The sample is therefore voluntary and self-selected, and consists of people who played the Bad News game in Swedish, German, Greek, or Polish and consented to participate in a within-subject (prepost) in-game survey.As per our ethics application, all incomplete responses, duplicates, and participants under 18 were excluded from the analyses.Table 1 in the appendix gives a detailed description of the sample characteristics for each country.Because we rely on a convenience sampling methodology, participants were more likely to be male, liberal, younger, and educated.However, randomized trials with the English version of the game have been conducted (Basol et al., 2020), highlighting similar results across diverse groups (Basol et al., 2020), consistent with the current findings.Our decision to conduct surveys "in the wild" was based on the real-world nature of the challenge of misinformation.It was therefore important that the game and its translations were freely available online to the entire population in each target country.Lastly, the intervention is of course, not without limitations.For example, it is possible that the intervention has unintended side effects, such as inadvertently promoting fake news.As Tandoc Jr. et al. (2018) note, "people spread fake news with two primary motivations; ideological and financial".We were careful to not provide either incentive in the game (the game itself is politically neutral and monetary incentives are never mentioned).We therefore deem it unlikely that the game will inspire people to produce fake news, but we cannot fully rule out potential side effects, including decay of the inoculationeffect over time.We encourage future research to further explore these issues."very" and "somewhat" categories).The violin shape represents the distribution of the data and the point estimate is the median (with boxplot).All pre-post differences within each country (by ideology) are statistically significant at the p < 0.001 level.

Figure 1 .
Figure 1.Bad News landing page and gameplay below.

Figure 2 .
Figure 2. Perceived reliability judgments for "real" control items (averaged, panel A), impersonation (panel B), conspiracy (panel C), and discrediting (Panel D), for the Swedish version of the game.The error bars represent 95% confidence intervals.The figureshows that the credible news items were perceived as equally reliable before and after playing, whereas the "fake" items are seen as significantly less reliable.All mean pre and post scores can be found in Table3in the appendix.

Figure 3 .
Figure 3. Average perceived reliability judgments for the fake news scale across countries pre and post gameplay (Germany, Greece, and Poland).Error bars represent 95% confidence intervals.The figure shows that the average perceived reliability of fake news items goes down significantly after gameplay in each country.All pre and post means for the individual items can be found in the methodology appendix.

Finding 3 :
The game's effectiveness may not substantially vary across key demographic variables.

Figure 1 .
Figure 1.Example of a fake Twitter post (impersonation technique)

Figure 3 .
Figure 3. Violin plots for the German, Greek, and Polish versions of the game(pre-and-post-test), by age.The violin shape represents the distribution of the data and the point estimate is the median (with boxplot).All pre-post differences within each country (by age) are statistically significant at the p < 0.01 level with the exception for older people in Poland (p = 0.13).

Figure 4 .
Figure 4. Violin plots for the German, Greek, and Polish versions of the game(pre-and-post-test), by education (lower education = high school or less).The violin shape represents the distribution of the data and the point estimate is the median (with boxplot).All pre-post differences within each country (by education) are statistically significant at the p < 0.001 level.

Figure 5 .
Figure 5. Violin plots for the German, Greek, and Polish versions of the game(pre-and-post-test), by political ideology (combines "very" and "somewhat" categories).The violin shape represents the distribution of the data and the point estimate is the median (with boxplot).All pre-post differences within each country (by ideology) are statistically significant at the p < 0.001 level.

Table 1 .
Sample characteristicsNote: Sample characteristics (proportions) for all respondents who filled out the sociodemographic questions at the end of the in-game survey.Note that because the survey is voluntary, sample sizes differ by item, country, and demographics.

Table 2 .
Survey items (fictitious Twitter posts); participants were asked to assess the reliability of each Twitter post on a 1-7 Likert scale, both before and after playing.