Research note: Bolsonaro’s firehose: How Covid-19 disinformation on WhatsApp was used to fight a government political crisis in Brazil

---

Implications 

In this study, we analyzed disinformation about Covid-19 on WhatsApp and its connection to the political debate in Brazil. Brazil was heavily hit by Covid-19, partially because of the different discourses from government authorities about the seriousness of the pandemic. From the beginning, for example, Brazil’s far-right president Jair Bolsonaro minimized Covid-19. He publicly called it a “little flu,” advocated for the end of measures such as social distancing and quarantine (which were put in place by state governors and mayors), and constantly appeared in public hugging his supporters without wearing a mask. The Health Minister at the time, Luiz Henrique Mandetta, on the other hand, strongly followed WHO recommendations, establishing measures of social distancing and publicly asking for people to use face masks to underline the seriousness of the pandemic. Most state governors and mayors supported Mandetta. In this context, two opposing discourses emerged: one reinforcing WHO recommendations about the seriousness of the situation and following scientific evidence, and another minimizing the pandemic (Sandy & Milhorance, 2020). Amidst this dispute, many people refused to wear masks, and multiple protests against the Covid-19 containment measures and social distancing protocols erupted, which were lead predominantly by Bolsonaro’s supporters. The pandemic quickly escalated in Brazil, with cases rapidly climbing from fewer than 100 cases in early March to more than 100,000 cases and over 8,000 deaths by early May. By the end of 2020, Brazil had one of the highest numbers of cases (over 7.5 million) and deaths (almost 200,000) in the world.¹See https://www.nytimes.com/interactive/2020/world/americas/brazil-coronavirus-cases.html

Brazil’s case demonstrates how disinformation can undermine public health strategies used by authorities and create a context in which people may choose not to follow scientific evidence and recommendations due to their political views. Political groups on WhatsApp may have played a central role in promoting the defiance of Covid-19 guidelines and regulations, since they already had part in spreading disinformation during elections and other key political moments (Reis et al., 2020). WhatsApp is the most popular communication app in Brazil and an important space for users to find and share information (Resende et al., 2018). According to the Reuters Institute Digital News Reports (Newman et al., 2020), 48% of Brazilians use WhatsApp for news consumption. Similarly, Baptista et al. (2019) found that 57% of Brazilians consider WhatsApp an important or very important source of political information. WhatsApp is also a key channel for spreading false information (Newman et al., 2020), particularly because it can’t be traced or debunked publicly by fact-checkers.

The problem of disinformation and its connections to partisanship and, particularly, the far-right has been explored by researchers in different contexts. Hindman and Varash (2018) have shown that the far right and Russian trolls have created communication strategies for disseminating elections-related disinformation on social media in the United States. Pierri et al. (2019) have described similar strategies in Italy. Alves (2016, 2019) has also pointed out how Brazil’s far right has taken control of social media strategies to massively create and share disinformation and hate campaigns targeted at political adversaries. 

In this article, we discuss disinformation about Covid-19 on WhatsApp in Brazil. In particular, we look at the topics and types of disinformation messages spread in political public groups on the app. Our findings imply that WhatsApp may have had a role in spreading disinformation about the disease through these groups, particularly by aligning untruthful content with political views. The disinformation spread may also have negatively influenced how people within the groups we monitored responded to the containment measures proposed by health authorities, particularly if these people were politically aligned with Bolsonaro. We understand disinformation as distorted, manipulated, or entirely false information that is created to mislead and achieve political gains (Benkler et al., 2018; Fallis, 2015). It is created intentionally to influence public opinion. Accordingly, disinformation is dangerous because it might negatively affect the public sphere (Tucker et al., 2018).  

Disinformation has been categorized by many. Wardle (2019), for example, considers seven categories of information disorder: satire and parody, false connection, misleading content, false context, imposter content, manipulated content, and fabricated content. Similarly, Brennen et al. (2020) divide disinformation into three categories: satire and parody, reconfigured, and fabricated. We use categories similar to Brennen et al. (2020); however, instead of reconfigured, we use distortion. We created a specific category for conspiracy theories, and we do not use satire and parody as a category. We removed satire and parody because, as we understand this category, it is about the “form” of disinformation, not the “content.” Thus, our classification system has three categories of disinformation: 

(1) distortion—content based on factual information that is distorted to mislead, such as information out of context, false connections, information biased reconfigured in some way;

(2) fabricated information—completely false content, such as fake audios and made up information;

(3) conspiracy theory—a narrative based on a conspiracy plot from a sinister group that lacks any evidence. 

Conspiracy theory differs from the two other categories because its key characteristic is the narrative describing obscure plots of social manipulation (for example, the conspiracy that China created the virus to win World War III)—and it does not matter whether the narrative is based on distorted factual information or completely false content. 

As we aimed to explore the connections between Covid-19 disinformation and political issues, we looked at the topics mentioned in the messages and their connection to the mainstream media political agenda during the period covered by our analysis (March and April 2020). Furthermore, we compared shares distribution and topic frequency with the political debate in Brazil to evaluate how the political debate impacted on disinformation. As other studies have noted, polarization fuels disinformation and leads to different responses, even in the health context of the Covid-19 pandemic (Allcott et al., 2020; Recuero & Soares, 2020; Recuero et al., 2020). 

As we discuss in the next section, disinformation about Covid-19 in Brazil was connected to political discourse within the country. Disinformation seems to have been connected to partisanship and, particularly, to Bolsonaro (see Finding 1, below). This implies that the support for certain political figures may be a strong motivator for circulating disinformation. Moreover, it also implies that legitimizing discourse of political leaders may also help spread disinformative content. Therefore, public authorities should find ways to punish political and other prominent public figures for spreading and legitimating untruthful content. That is, public authorities should act to develop legal procedures to address disinformation spread by political leaders. This could reduce the amount of disinformation spread by political leaders. In addition, our findings show how WhatsApp groups function as spaces that incite the spread of disinformation to other, more closed groups. This suggests that the only way WhatsApp could limit or avoid the spread of disinformation from is by flagging disinformation as such through debunked fact-checking stories, similar to what Reis et al. (2020) have also proposed.

Findings 

Finding 1: Covid-19 related disinformation on WhatsApp was largely political and provided a pro-Bolsonaro framing of the pandemic.

On March 17, Brazilians started banging pots from their windows every day at the same time to protest against Bolsonaro’s lack of action against the pandemic, calling for his impeachment (“Coronavirus protest in Brazil”, 2020). In the same period, several governors and mayors announced social distancing measures against Bolsonaro’s will (Marques, 2020). This crisis resulted in Bolsonaro’s worst approval rating since the beginning of his presidency. His response to the pandemic was condemned by 65% of Brazilians (Marcello & Brito, 2020). At the same time, political players that produced opposing discourse saw an increase in their approval ratings, such as governors and mayors (54% approved their response to the pandemic) and the Brazilian Health Minister Luiz Henrique Mandetta (55% approval rating).

As a response to this crisis, Bolsonaro made a public pronouncement on live television on March 24. In his speech, he criticized social distancing measures and, thus, governors, mayors, and Mandetta. He also called for Brazilians to go back to their normal lives, arguing the virus was not as serious as it had been made out to be. His speech started a disinformation campaign that connected social distancing measures, health authorities, and governors and mayors to the “left” and to “communism,” and the pandemic to a conspiracy.

In Figure 2, we show the distribution of disinformation over time. Right after Bolsonaro’s pronouncement, there was a huge spike in Covid-19 disinformation within the WhatsApp groups we studied. The topics from Bolsonaro’s speech were also central in the messages we identified that contained disinformation. The prevalence of these topics within the messages studied indicates that Bolsonaro’s speech might have fueled disinformation spread. Another public pronouncement on March 31, in which Bolsonaro affirmed he wanted to save the economy and claimed people needed to go back to work, also increased disinformation spread the day after.

To understand how disinformation was connected to political events, we analyzed the most frequent topics on the messages (categories were not exclusive – Figure 3) and their connections (Figure 4). The majority of disinformation (42%) was connected to how social distancing measures would hurt the economy. Other topics included the governors and mayors and their “hidden interests” against Bolsonaro (36%); how China had engineered the virus (27%); how the media (21%), the Congress (13%), the leftists (16%), the Brazilian Supreme Court (9%), Mandetta (6%) and others were all conspiring to defeat Bolsonaro.

To understand the connections between the themes and their relevance, we calculated the Pearson’s correlation coefficient between the topics and created a graph to visualize this data (see Figure 4) – each topic is a node and the ties connect significant positively correlated topics. Node size represents how many other nodes a topic is connected to. We also used modularity to identify topics that most frequently appeared together (different colors). The data support the strength of the political framing of the pandemic, with “leftists” as the most central concept, connected to almost all political players that produced a discourse opposed to Bolsonaro. Many disinformation messages accused these political actors of being “leftists” or “communists.” This political argument was used to frame all discourses that discredited Bolsonaro’s as created by political opposition. Thus, polarization between “us” (the good) and “them” (the evil) was one key strategy for spreading disinformation. Most clusters connect political actors and health authorities to Bolsonaro and his criticism of social distancing. 

Finding 2: Conspiracy theories and distortion were the most frequent types of disinformation in the messages analyzed.

Most studies have found that distortion is the most frequent type of disinformation on many social media platforms. For example, Brennen et al. (2020) studied Twitter, Facebook, and YouTube, and Recuero and Soares (2020) analyzed Twitter in Brazil; the categories used in these studies were slightly different from ours, but both studies included distortion, fabricated information, and conspiracy theories as categories or subcategories. However, conspiracy theories were the most prevalent category in our dataset (41%). Although some found evidence of conspiracy theories about Covid-19 on Twitter (see Gruzd & Mai, 2020), there were fewer messages with these characteristics compared to other types. For example, Brennen et al. (2020) identified conspiracy theories as only 17% of their dataset. The second most frequent type of disinformation we identified was distortion (39%) and the third, fabricated information (20%). This result highlights the importance of considering platform differences in this type of research.

Conspiracy theories were strongly correlated to the theme “China” (r(800) = .53, p < .001). These messages often accused the Chinese (as the representatives of “communism”) of creating the virus for economical gain (see Figure 5). 

Conspiracy theories were also correlated to other political topics, such as Brazilian Congress (r(800) = .15, p < .001), media outlets (r(800) = .20, p < .001), Brazilian Supreme Court (r(800) = .08, p < .05), and “leftists” and “communists” (r(800) = .18, p < .001). Narratives accusing several actors of plotting against Bolsonaro were frequently found during the political crisis he faced (figure 6).

Distortion and fabricated content were also used to support Bolsonaro. Distorting messages often criticized social distancing measures (r(800) = .29, p < .001), and reinforced Bolsonaro’s positions (figure 7). Fabricated information was used, for example, to state that hydroxychloroquine was the cure (r(800) = .09, p < .05) for Covid-19 (see Figure 8). Bolsonaro claimed the same.

The data show that disinformation about Covid-19 on WhatsApp in Brazil followed a pattern of framing the pandemic as a conspiracy and its combat measures as strategies to undermine Bolsonaro’s government. Disinformation was used to fight the political crisis the president was in, offering “alternative” stories to support his discourse. Moreover, these messages framed the public health issue as a political one (the “leftists” vs. Bolsonaro), taking advantage of the political polarization in Brazil. 

Methods

In this study, we aimed to understand the topics and the types of disinformation about Covid-19 on WhatsApp in Brazil. We analyzed a sample of disinformation messages from public political groups. Most interactions on WhatsApp are private, so it is hard to track data from the app. To create our sample, we used WhatsApp Monitor (Resende et al., 2018) to collect data from more than 500 WhatsApp public groups in Brazil. The public groups tracked are for political discussion and news sharing. The system provides an overview of the content shared on WhatsApp, as it stores multiple content categories such as images, videos, audio, and textual messages and displays the most shared content per day in these groups. We only have information about the content shared in the groups, so we do not know how many times the messages we analyze were shared from the tracked public groups to other groups or one-to-one conversations. The WhatsApp Monitor does not store sensitive data, such as usernames and phone numbers. Therefore, we used only data from public groups and without any identification of the users to avoid ethical issues. 

Data collection

Two authors accessed the WhatsApp Monitor system and manually collected textual messages that we considered to contain disinformation about Covid-19. We collected textual messages that were shared at least 20 times in one day within the monitored groups from March 1 to April 30, 2020. If a message was shared at least 20 times in two days, then we collected it twice. We decided to keep the duplicated messages because they were potentially shared in different groups which helped us to analyze the distribution of the shares per day. We retrieved a dataset of 810 messages. During the content analysis (detailed below), we removed eight messages that we did not classify as disinformation. Our final dataset comprised 802 messages (727 unique messages) that were shared almost 35,000 times in the groups that WhatsApp Monitor tracks. Our dataset consists of about a third of the total number of messages shared at least 20 times a day in the monitored groups in March and April 2020 (nearly 2,400 messages). We selected this particular timeframe because it was the beginning of the Covid-19 spread in Brazil. Due to Bolsonaro’s response to the pandemic, it was a period of political instability and polarized discourses. With this timeframe, we were able to map the beginning of the disinformation campaign on WhatsApp and its connections to the political debate in the country. We also monitored several news outlets to make inferences about the connection between public discourse, disinformation spread, and the topics mentioned in the messages.

Data analysis

To understand the characteristics of the disinformation we used content analysis. We first made a preliminary analysis of the dataset, looking at the major characteristics of the messages, the topics mentioned, and so on. We then separately coded the types of disinformation and the topics of the messages. The authors of this essay performed the content analysis. We discussed the disinformation categories based on a preliminary analysis of the dataset, but we did not conduct any prior training on a separate data set nor the data set used in this study. Two coders manually annotated all 810 messages to classify the types of disinformation (distortion, fabricated information, and conspiracy theory). The categories were mutually exclusive. For the coding schema, we used the definitions outlined in the arguments and implications section. We calculated Krippendorff’s alpha to measure the reliability of our coding. The overall score was moderate to substantial (α = .601). Types of disinformation are interpretative categories, which explains the agreement score. To have more confidence in our final classification, we used a third coder to solve the disagreements (tiebreaker), so two coders agreed on the final categorization for each message. As mentioned above, eight messages were removed from the original dataset because they were not considered disinformation by the coders. Three coders manually annotated the topics in the 802 messages. The coders were asked whether the message contained each of the topics we created based on the preliminary analysis. Topics categories were not exclusive. The overall Krippendorff’s alpha score was a substantial (α = .675). We used the agreement of at least two coders for the final decision. We used Freelon’s (2010) ReCal2 and ReCal3 to calculate intercoder reliability. For the qualitative inferences we made and to select the examples we presented in the essay, we returned to the data for a more qualitative deep reading of the messages. We used SPSS to run Pearson’s correlation tests on the data. We used Gephi to create a graph of the correlated topics. We manually created an edge list based on the positively correlated topics with statistical significance. Although some of the correlations were small (r < .2), we decided to report them because they helped us identify topics that were mentioned in similar contexts. In Figure 4, each node represents one topic and the ties connect positively correlated topics. We used modularity to identify the clusters and the degree to define the size of the nodes. 

Limitations

This study has some limitations. We only had access to public groups monitored by WhatsApp Monitor. Although it monitors more than 500 public groups, we were not able to measure the spread of the disinformation messages in other groups and interpersonal conversations on WhatsApp. Therefore, our data was based on a sample of the messages on WhatsApp, and we could not measure how prevalent these messages were in the broader community. Also, we only analyzed textual messages. Disinformation is also spread in videos and images, and we likely missed part of the disinformation messages spread on WhatsApp. In addition, there is a bias in the political public groups monitored by WhatsApp Monitor, as many of them are groups for Bolsonaro’s supporters.