We used Prolific, a UK-based online crowdsourcing platform that connects researchers to participants, to collect data from US citizens over a 4-week period. Prolific has been demonstrated to yield high-quality data and more diverse participants than student samples or other major crowdsourcing platforms [24]. In addition, it allowed for recruitment that balanced age, gender, and ethnicity to approximate the US general public, via stratification using US census data [25]. Recruitment for the initial baseline wave started on April 27, 2020.

A total of 1212 individuals participated in the study at baseline (T0), for which they received £0.45 (~US $0.56). A total of 1089 individuals participated in the first follow-up wave (T1), 1070 individuals participated in the second follow-up wave (T2), and 1028 individuals participated in the final wave (T3); see Table 1 for total sample size, exclusions, and final sample size per wave. Participants received £0.33 (~US $0.41) for participation per follow-up survey. Each of the waves was separated by approximately one week (mean_T0-T1 6.98 days, SD _T0-T1 14.92 hours; mean_T1-T2 7.01 days, SD _T1-T2 12.72 hours; mean_T2-T3 7.06 days, SD _T2-T3 13.18 hours). The sample size was determined by the available resources. This study is part of a research project that was reviewed and approved by the Ethics Committee Social Science at Radboud University (reference No. ECSW-2018-056).

Participants were randomly assigned to either receive the intervention (ie, the boost condition) or no intervention (ie, the control condition). All surveys started with the measure of belief accuracy, for which participants were presented with 10 (T0), 14 (T1), 18 (T2), or 22 (T3) statements about the coronavirus and COVID-19 (see Figure 1). Participants indicated to what extent they believed each statement to be true. An attention check was included among these statements (see Multimedia Appendix 1). Subsequently, participants reported their behavior aimed at preventing the spread of the coronavirus and completed the other measures. The boosting intervention, an infographic presenting three steps that can be used to evaluate a claim, was included at the end of T0, T1, and T2. Only participants in the boost condition were presented with the infographic, allowing them to apply their boosted consensus reasoning skill in the week leading up to the next wave. At T3, all participants completed a manipulation check. At the end of T0, all participants entered demographic information and completed a seriousness check (see Multimedia Appendix 1). All surveys took about 3 to 6 minutes.

The final sample roughly reflects US census data [25] on gender, age, and ethnicity, indicating that the balanced sampling worked well. See Table 2 for more details.

Mean scores of belief accuracy were very high for all waves, with scores reflecting low belief in false statements and high belief in true statements. There was substantial variation in the accuracy of responses between statements, although none of the statements was ever interpreted with less than 0.25 accuracy, on average; see Multimedia Appendix 4 for a complete overview of scores per statement per wave.

There was a modest increase in belief accuracy over time, looking at each set of statements separately (first 10: estimate=0.02, SE<0.01; t_3202.59=13.82, P<.001; T1 set: estimate=0.01, SE<0.01; t_2041.94=4.80, P<.001; T2 set: estimate=0.02, SE<0.01; t_1003.22=3.40, P<.001). This increase was positive for all three sets of statements that were asked more than once (see Figure 2 [34,35]), indicating that participants became more accurate in their interpretation of the statements over time.

Accurate beliefs were correlated with self-reported behavior aimed at preventing the coronavirus from spreading; r at all waves was between 0.26 and 0.29 and all P values were less than .001. This small but robust correlation suggests that accurate beliefs could be important for coronavirus-related behavior. We explored potential evidence of any causal effects in the data using a random-intercept cross-lagged panel model. This yielded a tentative indication that accurate beliefs might be predictive of behavior, with belief accuracy at T2 predicting coronavirus-related behavior at T3. However, with all other paths showing no sign of significant predictive effects, the results regarding causality are largely inconclusive (see Multimedia Appendix 5).

We explored the relationship of trust in scientists at T0, political orientation, and the primary news source with belief accuracy. The mixed model yielded a significant positive relationship between belief accuracy and trust (estimate=0.07, SE<0.01; t_1200.23=16.44, P<.001) and a significant negative correlation with political orientation (estimate=–0.02, SE<0.01; t_1199.62=–6.78, P<.001). These main effects indicated that participants with higher trust in scientists scored higher on the measure of belief accuracy and that liberal, Democratic participants held more accurate beliefs than conservative, Republican participants. Moreover, these main effects were partially qualified by an interaction effect among trust and political orientation (estimate=–0.01, SE<0.01; t_1195.05=–3.62, P<.001). Plotting of this interaction effect demonstrated that trust in scientists had a stronger relationship with belief accuracy for liberal, Democratic participants than it had for conservative, Republican participants (see Figure 3).

Two of the five most chosen primary news sources were associated with a worse understanding of the facts regarding the COVID-19 pandemic than others (see Figure 4). Participants who reported CNN (estimate=–0.03, SE=0.01; t_1194.49=–2.33, P=.02) or Fox News (estimate=–0.05, SE=0.02; t_1202.49=–3.05, P=.002) as their main news source scored below average on belief accuracy.

We conducted a manipulation check and, as expected, when asked how they evaluated claims, participants in the boost condition (136/600, 22.7%) mentioned consensus, or something similar, more often than participants in the control condition (26/602, 4.3%; χ²₁=85.2, P<.001, N=1202).

We hypothesized that our boosting intervention would lead to more accurate beliefs about the COVID-19 pandemic than would the control condition. However, the interaction effect between condition and wave on belief accuracy was not significant (estimate<0.01, SE<0.01; t_1074.36=0.22, P=.83). This means that the boosting intervention did not significantly alter belief accuracy of participants over time, compared to the control condition (see Figure 5). This was also the case when we explored effects of the intervention on inaccurate statements only (P=.48), accurate statements only (P=.49), only the original 10 statements that were included in all waves (P=.61), and only included participants who scored relatively low on belief accuracy at T0 (belief accuracy_T0<0.76; P=.32).

When the subsample of participants in the boost condition who mentioned consensus, or similar, in the manipulation check was compared to participants in the control condition, we again found that the intervention did not increase belief accuracy (P=.21). However, there was a main effect of condition (estimate=0.06, SE=0.02; t_725.66=3.50, P=.001), indicating that participants in the boost condition who did use consensus while evaluating claims scored higher on belief accuracy than participants in the control condition. This difference was already present at T0, so was not caused by the intervention.

We explored the effect of the boosting intervention on trust in scientists as a source of information about the coronavirus. The mixed-effects model, similar to the hypothesis test but with the repeated measure of trust as the dependent variable and including the interaction term between condition, wave, and political orientation, yielded a significant three-way interaction effect between condition, wave, and political orientation (estimate=0.02, SE=0.01; t_1088.42=2.39, P=.02). Trust in scientists was very high in all four waves (means between 6.11 and 6.19), but investigation of the two-way interaction effects per condition indicated a significant interaction effect among wave and political orientation in the control condition (estimate=–0.02, SE=0.01; t_562.01=–3.24, P=.001), while there was no such significant interaction effect in the boost condition (P=.90). As illustrated by Figure 6, there was a clear overall difference in trust in scientists between participants related to their political orientation. More interestingly, trust remained stable for all participants in the boost condition, but decreased slightly for conservative, Republican participants in the control condition. This could indicate that the boosting intervention inhibited a decline of trust in scientists as a source of information about the coronavirus among more conservative, Republican participants.

The aims of this study were to gain insight into the beliefs of the US public about the COVID-19 pandemic and to investigate whether a boosting intervention could improve people’s belief accuracy. Interestingly, the average scores on belief accuracy over the surveyed 4-week period were high, indicating low belief in false statements and high belief in true statements. Looking at each set of statements, we found a small but significant increase in belief accuracy over time. This indicates that the general public is quite able to figure out what is true and what is not in times of crisis. Moreover, a small but robust correlation suggests that accurate beliefs about the pandemic might be important for coronavirus-related behavior. Associations with belief accuracy suggest that the processes of belief formation and correction might be affected by individuals’ trust in scientists and political orientation, as well as their news habits. Finally, the boosting intervention yielded no significant increase in belief accuracy over the control condition, demonstrating that the boosting infographic was not successful in helping people figure out what is true and what is false. Exploratory analyses suggested that the intervention did, however, inhibit a decline in trust in scientists as a source of information about the coronavirus among more conservative, Republican participants.

There is a great deal of worry about the prevalence of misinformation during the current pandemic, which is reflected in popular media (eg, [36,37]), as well as among scientists (eg, [38,39]) and public health agencies (eg, [4,40]). The supposed COVID-19 infodemic is not reflected in US citizens’ beliefs. The finding that most Americans hold quite accurate beliefs about the COVID-19 pandemic is in line with emerging work on perceptions of the pandemic that shows that belief in COVID-19 misperceptions and conspiracy theories is quite low [41-44]. Consequently, this calls into question the prevalence of misinformation or the public’s susceptibility to misinformation.

A convincing body of empirical work on the prevalence of misinformation surrounding the COVID-19 pandemic is not yet available. Research from before the COVID-19 pandemic indicates that the prevalence of misinformation might be lower than many believe [45-47]. Still, it is possible that the current pandemic has led to an increase in misinformation compared to the information landscape from before the pandemic. However, when looking for potential explanations of this study’s findings, we should consider the possibility that COVID-19 misinformation is not as prevalent as expected. Perhaps misinformation makes up only a small portion of the average US citizen’s media diet.

The second possibility is that we are indeed facing a COVID-19 infodemic, but that the public is not very susceptible to it. Misinformation campaigns regarding other topics, such as climate change and the health effects of tobacco [18,48], have demonstrated that misinformation can contribute to misperceptions about important matters. In these cases, however, misinformation campaigns have been carefully organized and executed, continually misinforming the public for decades. In contrast, the COVID-19 pandemic is a novel issue and, at least in the relatively early months that we investigated, did not yield many such coordinated misinformation campaigns. Moreover, the COVID-19 pandemic originated in a very different media landscape than the climate change and tobacco misinformation campaigns. Fake news, misinformation, and disinformation have been discussed widely and frequently in popular media since the 2016 US presidential election and the 2016 United Kingdom European Union membership referendum. This might have resulted in the public being more aware of misinformation campaigns targeting them. Perhaps the widespread discussion of misinformation in popular media has worked as a large-scale media literacy intervention, putting people on guard against false information. In support of this idea, recent research has demonstrated that simply asking one to consider the accuracy of a claim improved subsequent choices about what COVID-19 news to share on social media [49].

A third possibility that should be considered is that the public is more careful in forming beliefs in times of crisis, especially in the relatively early days of a crisis, making a well-informed public not unique to the COVID-19 pandemic. In times of crisis, people are likely to increase news consumption (eg, [50]). This was also the case in the United States during the first months of the COVID-19 pandemic, with people reporting increased news consumption [51]. Although we found that those who reported CNN or Fox News as their main news source scored below average on belief accuracy, the general increase in news consumption may lead to a better understanding of the crisis situation, including more accurate beliefs.

Turning to the finding that political orientation is associated with individuals’ belief accuracy, we see that this is in line with other emerging work [41]. There is likely a multitude of explanations for this evolving partisan divide [52] regarding perceptions about the pandemic, such as political party cues in the news affecting opinion formation [53], the difficulty of correcting false beliefs for the ideological group most likely to hold those misperceptions [17], as well as the differences in news consumption that are reflected in this study. A second variable that is even more strongly related to belief accuracy is trust in scientists as a source of information about the coronavirus, demonstrating that higher trust is related to more accurate beliefs. Interestingly, the associations of political orientation and trust with accurate beliefs were partially explained by an interaction effect among political orientation and trust. The stronger association of trust with belief accuracy for more liberal, Democratic individuals might mean that they rely more on scientists’ perceptions in forming beliefs, while relatively more conservative, Republican individuals might rely more on other cues. Relatedly, the inhibited decline in trust in scientists among conservative, Republican participants in the boost condition could indicate that information about the scientific process might resonate more with them than just hearing the results of this process. Though this exploratory finding should be replicated, it could provide a fruitful avenue for further research on trust in scientists and political orientation.

In addition, this study demonstrates that some news sources might be doing a worse job of informing their consumers about the COVID-19 pandemic than others, or perhaps that better-informed news consumers turn to different news sources than less well-informed consumers, again in line with other emerging work [41]. Most likely, a combination of both selection and influence (eg, [54]) explain the differences in belief accuracy found in this study. Interestingly, considering the role of social media in the spread of misinformation (eg, [55]), with about 26% to 42% of tweets in the data collection period containing unreliable facts [56], participants who reported social media sites as their main source of news about the coronavirus did not display significantly worse belief accuracy than others. However, it is possible that participants who reported social media sites as their main source followed major news outlets via the social media site, thereby being exposed to similar news content as the other participants.

Finally, this study demonstrates the difficulty of crafting interventions aimed at increasing belief accuracy. Recent work demonstrates that simple, short media literacy interventions can work [14,15], while other work highlights the difficulties of crafting these interventions [16]. We argue that the divergent findings can be explained by the fact that in the former work the interventions were paired with corrections, while in this study participants had to put their new skill to use outside of the study context. Considering that cues signaling the existence of consensus in relevant news content are very rare [57], participants likely had to search for information about scientific consensus themselves. The results from the manipulation check indicated that only a relatively small portion of participants actually applied this strategy. However, those individuals who indicated that they did apply a strategy related to consensus reasoning scored higher on belief accuracy than the control group. This difference highlights the potential of the intervention in situations where individuals can be empowered to actually apply it.

There are two notable limitations to this study. First, our belief accuracy measure consisted only of science-based statements. We incorporated only science-based claims in our study to ensure that there was sufficient empirical evidence stating that a claim was either true or false. However, this decision did exclude some coronavirus-related claims that were not based on science (eg, “Bill Gates patented the coronavirus”) or were unresolved at the time (eg, “A vaccine will be available before the end of the year”). It should have been harder for participants to figure out whether such unresolved issues were true or not, yielding different responses from participants for a measure reflecting non-science-based, unresolved issues about the pandemic.

A second limitation is the fact that the recruitment platform that we used, Prolific, is known as a platform for research. Although participants on the platform receive financial incentives for completing studies, they might be more interested in scientific research than the average US citizen. This could lead to them also having a higher trust in science than the general population, even though our sample was balanced regarding age, gender, and ethnicity. As trust in science was highly related to belief accuracy, it could be possible that this led to an inflated belief accuracy score. Future research should attempt to replicate this study with a sample that represents the US population better than our balanced sample.

Our work demonstrates that most people are quite able to figure out the facts in this time of crisis, but also that it is difficult to craft an intervention targeting these beliefs. However, in cases where people do not immediately have a clear understanding of the facts, they are capable of figuring them out over time. There are some factors that might make it easier or harder for one to figure out the facts. We found that the accuracy of participants’ beliefs was related to political orientation, as well as their primary news source. This suggests that, even in the relatively early days of the pandemic, political polarization and media diet had a grip on US citizens’ factual beliefs, leading to polarization along party lines. Another factor strongly related to accurate beliefs about the pandemic was trust in scientists. It is unclear whether an already-high trust led to accurate beliefs or whether being able to figure out the facts increased trust in scientists, but the importance of expert communication is underlined by these findings.

Although a small but robust correlation suggests that accurate beliefs about the pandemic might be important for coronavirus-related behavior, the role of misinformation in the pandemic seems to be relatively small, either because it is rare or because it is unable to persuade. However, we note that even if misinformation is not prevalent and only accepted by a small portion of the receivers, it can still be dangerous. To illustrate, we found that almost all participants in this study disregarded the statement that injecting or ingesting bleach is a safe way to kill the coronavirus, but this false claim is reported to have cost at least one life [58]. Additionally, with the antivaccine community launching coordinated misinformation campaigns against potential coronavirus vaccines [59] and politicization of the pandemic looming, the infodemic might become a much bigger threat.