Research note: Fighting misinformation or fighting for information?

---

Implications

In psychological experiments, participants are approximately as likely to accept a piece of fake news as they are to reject a piece of true news (Altay et al., 2021a; Pennycook et al., 2020; Pennycook & Rand, 2021), suggesting that the acceptance of fake news and the rejection of true news are issues of similar amplitude. Such results, combined with the apparent harmfulness of some fake news, have led to a focus on fighting misinformation. However, studies concur that the base rate of online misinformation consumption in the United States and Europe is very low (~5%) (see Table 1). Most of the large-scale studies measuring the prevalence of online misinformation define misinformation at the source level: news shared by sources known to regularly share fake, deceptive, low-quality, or hyperpartisan news is considered to be online misinformation (see the ‘definition’ column in Table 1). In the United States, misinformation has been calculated to represent between 0.7% and 6% of people’s online news media diet (Altay et al., n.d.; Grinberg et al., 2019; Guess et al., 2018; Guess, Lerner, et al., 2020; Osmundsen et al., 2021), and 0.15% of their overall media diet (Allen et al., 2020). In France, misinformation has been calculated to represent between 4 and 5% of people’s online news diet (Altay et al., n.d.) and 0.16% of their total connected time (Cordonier & Brest, 2021). Misinformation has been calculated to represent approximately 1% of people’s online news diet in Germany (Altay et al., n.d.; Boberg et al., 2020), and 0.1% in the UK (Altay et al., n.d.). In Europe, during the 2019 EU Parliamentary election, less than 4% of the news content shared on Twitter came from unreliable sources (Marchal et al., 2019). Overall, these estimates suggest that online misinformation consumption is low in the global north, but this may not be the case in the global south (Narayanan et al., 2019). It is also worth noting that these estimates are limited to news sources, and do not include individuals’ own posts, group chats, memes, etc.

To illustrate our argument, we developed a model that estimates the efficacy of interventions aimed at increasing the acceptance of reliable news or decreasing the acceptance of misinformation. Our model shows that under a wide range of realistic parameters, given the rarity of misinformation, the effect of fighting misinformation is bound to be minuscule, compared to the effect of fighting for a greater acceptance of reliable information (for a similar approach see Appendix G of Guess, Lerner, et al., 2020). This doesn’t mean that we should dismantle efforts to fight misinformation, since the current equilibrium, with its low prevalence of misinformation, is the outcome of these efforts. Instead, we argue that, at the margin, more efforts should be dedicated to increasing trust in reliable sources of information rather than in fighting misinformation. Moreover, it is also crucial to check that interventions aimed at increasing skepticism towards misinformation do not also increase skepticism towards reliable news (Clayton et al., 2020). Note that our model does not compare the effect of existing interventions, but the effect that hypothetical interventions would have if they improved either rejection of misinformation or acceptance of reliable information.

Improving trust in sound sources, engagement with reliable information, or acceptance of high-quality news is a daunting task. Yet, some preliminary results suggest that this is possible. First, several studies have shown that transparency boxes providing some information about the journalists who covered a news story and explaining why and how the story was covered enhances the perceived credibility of the journalist, the story, and the news organization (Chen et al., 2019; Curry & Stroud, 2017; Johnson & St. John III, 2021; Masullo et al., 2021). Second, credibility labels informing users about the reliability of news sources have been shown to increase the news diet quality of the 10% of people with the poorest news diet (Aslett et al., n.d.), but overall, such labels have produced inconsistent, and often null, results (Kim et al., 2019; Kim & Dennis, 2019). Third, in one experiment, fact-checks combined with opinions pieces defending journalism increased trust in the media and people’s intention to consume news in the future (Pingree et al., 2018). Fourth, in another experiment, fact-checking tips about how to verify information online increased people’s acceptance of scientific information from reliable news sources they were not familiar with (Panizza et al., 2021). Finally, a digital literacy intervention increased people’s acceptance of news from high-prominence mainstream sources but reduced acceptance of news from low-prominence mainstream sources (Guess, Lerner, et al., 2020).

More broadly, interventions fostering critical thinking, inducing mistrust in misinformation, and reducing the sharing of misinformation (Cook et al., 2017; Epstein et al., 2021; Roozenbeek & van der Linden, 2019; Tully et al., 2020), could be adapted to foster trust in reliable sources and promote the sharing of reliable content.

Findings

We developed a simple model with two main parameters: the share of misinformation (the rest being reliable information) in the environment and the tendency of individuals to accept each type of information when they encounter it. Reliable information refers to news shared by sources that, most of the time, report news accurately, while misinformation refers to news shared by sources that are known to regularly share fake, deceptive, low-quality, or hyperpartisan news. With this broad definition, misinformation represents approximatively 5% of people’s news diets, with the remaining 95% consisting of information from reliable sources (Allen et al., 2020; Cordonier & Brest, 2021; Grinberg et al., 2019; Guess et al., 2019; Guess, Nyhan, et al., 2020; Marchal et al., 2019). The rate at which people accept reliable information or misinformation when exposed to it is less clear. Here, we take as a starting point experiments in which participants are asked to ascertain the accuracy of true or fake news, suggesting that they accept approximately 60% of true news and 30% of fake news (Altay et al., 2021a; Pennycook et al., 2020; see Appendix A for more information). As shown below, the conclusions we draw from our models are robust to variations in these parameters (e.g., if people accept 90% of misinformation instead of 30%).

The goal of the model is to provide a broad picture of the informational environment, and a rough index of its quality. Although it has some clear limitations (discussed below), it captures the main elements of an informational environment: the prevalence of reliable information vs. misinformation, and people’s propensity to accept each type of information. While more elements could be included, such simple models are crucial to put the effects of any type of intervention in context.

In our model, exposition to news is drawn from a log-normal distribution, with few agents (i.e., the individuals simulated in the model) being exposed to many pieces of news (reliable and unreliable) and the majority of being exposed to few pieces of news, mimicking the real-life skewed distribution of news consumption (e.g., Allen et al., 2020; Cordonier & Brest, 2021). Due to the low prevalence of misinformation, we compare extreme interventions that bring the acceptance rate of misinformation to zero (Figure 1, left panel) to a counterfactual situation in which no intervention took place (black dotted line) and to interventions that increase the acceptance rate of reliable information from a range of one to ten percentage points. We show that an intervention reducing the acceptance rate of misinformation from 30% to zero, increases the overall information score as much as an intervention increasing acceptance of reliable information by one percentage point (i.e., from 60% to 61%).

On the right panel of Figure 1, we plotted how much more efficient in improving the global information score is an intervention on reliable information, compared to an intervention reducing acceptance of misinformation to zero. The only situations in which the intervention on misinformation has an advantage is when the proportion of misinformation is (unrealistically) high, and the improvement in the acceptance rate of reliable information is very low (i.e., at the bottom right corner of the plot). Overall, minute increases in acceptance of reliable information have a stronger effect than completely wiping out acceptance of misinformation. A one percentage point increase in reliable information acceptance has more effect than wiping out all misinformation for all realistic baselines of misinformation prevalence (i.e., 1 to 5%).

In these simulations, the baseline acceptance of misinformation was set to 30%. This percentage, however, was obtained in experiments using fake news specifically, and not items from the broader category of misinformation (including biased, misleading, deceptive, or hyperpartisan news). As a result, the acceptance of items from this broader category might be significantly higher than 30%. We conducted simulations in which the baseline acceptance rate of misinformation was raised to a (very unrealistic) 90%. Even with such a high baseline acceptance of misinformation, given the disproportionate frequency of reliable information with respect to misinformation, an intervention that brings the acceptance rate of misinformation to 0% would only be as effective as increasing belief in reliable information by 4% (for a prevalence of misinformation of 5%).

This basic model was extended in two ways. First, despite its low prevalence, online misinformation can have deleterious effects on society by eroding trust in reliable media (Tandoc et al., 2021; Van Duyn & Collier, 2019; although see Ognyanova et al., 2020). In the first extension of our model, we tested whether misinformation could have a deleterious effect on the information score by decreasing trust in reliable information. In this model, when agents accept misinformation, they then reject more reliable information, and when agents accept reliable information, they then reject more misinformation. In such scenarios, losses in the global information score are mostly caused by decreased acceptance of reliable information, not by increased acceptance of misinformation. Manipulating the relevant parameters shows that even when the deleterious effect of misinformation on reliable information acceptance is two orders of magnitude stronger than the effect of reliable information on rejection of misinformation, the agents keep being more likely to accept reliable information than misinformation (Figure 2, bottom left). Even in this situation, modest interventions that improve acceptance of reliable information (by 1%) are more effective than bringing acceptance of misinformation to zero (Figure 2, bottom right).

In the model so far, the relative proportion of misinformation and reliable information has been used as a fixed parameter. However, the acceptance of misinformation, or, respectively, reliable information, might increase its prevalence through, for example, algorithmic recommendations or social media sharing. In a second extension, accepting misinformation increases the prevalence of misinformation, and accepting reliable information increases the prevalence of reliable information. Similar to the results of the previous extension, given that reliable information is initially much more common, we find that misinformation becomes prevalent with respect to reliable information only when the effect of accepting misinformation on misinformation prevalence is two orders of magnitude larger than the effect of accepting reliable information on reliable information prevalence, which is highly unrealistic. This shows that a sharp increase in the prevalence of misinformation—which would invalidate our main results—requires unrealistic conditions. Moreover, the basic simulation shows that modest increases in the prevalence of misinformation do not challenge our main conclusion: even with a 10% prevalence of misinformation, improving the acceptance of reliable information by three percentage points is more effective than bringing acceptance of misinformation to zero.

The models described in this article deal with the prevalence and acceptance of misinformation and reliable information, not their potential real-life effects, which are difficult to estimate (although the importance of access to reliable information for sound political decision-making is well-established, see Gelman & King, 1993; Snyder & Strömberg, 2010). Our model doesn’t integrate the possibility that some pieces of misinformation could be extraordinarily damaging, such that even a tiny share of the population accepting misinformation could be hugely problematic. We do note, however, that since the prevalence of misinformation is very low, the negative effects of each individual piece of misinformation would have to be much greater than the positive effects of each individual piece of reliable information to compensate for their rarity. This appears unlikely for at least two reasons. First, every piece of misinformation could be countered by a piece of reliable information, making the benefits of accepting that piece of reliable information equal in absolute size to the costs of accepting the piece of misinformation. As a result, high costs of accepting misinformation would have to be mirrored in the model by high benefits of accepting reliable information. Second, some evidence suggests that much misinformation, even misinformation that might appear extremely damaging (such as COVID-19 related misinformation, or political fake news), mostly seem to have minimal effects (Allcott & Gentzkow, 2017; Altay et al., 2021b; Anderson, 2021; Carey et al., n.d.; Guess, Lockett, et al., 2020; Kim & Kim, 2019; Litman et al., 2020; Valensise et al., 2021; Watts & Rothschild, 2017).

Our model is clearly limited and preliminary. However, we hope that it demonstrates the importance of such modeling to get a broader picture of the potential impact of various interventions on the informational environment. Future research should refine the model, particularly in light of new data, but the main conclusion of our model is that interventions increasing the acceptance of reliable information are bound to have a greater effect than interventions on misinformation.

Methods

In the model, N agents (N = 1,000 for all results described here) are exposed to pieces of news for T time steps, where each time step represents a possible exposure. Exposure is different for each agent, and it is drawn from a log-normal distribution (rescaled between 0 and 1), meaning that the majority of agents will have a low probability of being actually exposed to a piece of news at each time step, and few agents will have a high probability, mimicking the real-life skewed distribution of news consumptions (e.g., Allen et al., 2020; Cordonier & Brest, 2021). 

First, a main parameter of the model (C_m Composition misinformation) determines the probability that each piece of news will be either misinformation or reliable misinformation. The baseline value of this parameter is 0.05, meaning that 5% of news is misinformation. Second, two other parameters control the probability, for each agent, to accept reliable information (B_r Believe reliable) and to accept misinformation (B_m Believe misinformation). These two values are extracted, for each agent, from a normal distribution truncated between 0 and 1, with standard deviation equal to 0.1 and with mean equal to the parameter values. The baseline values of these parameters are 0.6 and 0.3 for B_r and B_m respectively, so that agents tend to accept 60% of reliable information and 30% of misinformation. Finally, a global information score is calculated as the total number of pieces of reliable information accepted minus the total number of pieces of misinformation accepted, normalized with the overall amount of news (and then multiplied by 100 to make it more legible). A global information score of -100 would mean that all misinformation is accepted and no reliable information, and a global information score equal to 100 would mean that all reliable information is accepted and no misinformation.

Finally, a global information scoreis calculated as the total number of pieces of reliable information accepted minus the total number of pieces of misinformation accepted, normalized with the overall amount of news (and then multiplied by 100 to make it more legible). A global information score of -100 would mean that all misinformation is accepted and no reliable information, and a global information score equal to 100 would mean that all reliable information is accepted and no misinformation. 

In the main set of simulations, we first compare (see results in Figure 1 – left panel) the global information score of our baseline situation (C_m = 0.05; B_m = 0.3; B_r = 0.6) with a drastic intervention that completely wipes out acceptance of misinformation (B_m = 0), and with small improvements in reliable information acceptance (B_r = 0.61, B_r = 0.62, B_r = 0.63, etc. until B_r= 0.7). We then explore the same results for a larger set of parameters, including changing C_m from 0.01 to 0.1 in steps of 0.01, i.e., assuming that the proportion of misinformation can vary from 1 to 10% with respect to total information. The results in Figure 1 – right panel show the difference between the global information score obtained with the parameters indicated in the plot (improvements in reliable information acceptance and composition of news) and the information score obtained with the drastic intervention of misinformation for the same composition of news. All results are based on 10 repetitions of simulations for each parameter combination, for T = 1,000. The two extensions of the model are described in Appendix B and C. All the code to run the simulations is written in R, and it is available at https://osf.io/sxbm4/.