Since the breakout of COVID-19 in late 2019, various conspiracy theories have spread widely on social media and other channels, fueling misinformation about the origins of COVID-19 and the motives of those working to combat it. This study analyzes tweets ( N = 313,088) collected over a 9-month period in 2020, which mention a set of well-known conspiracy theories about the role of Bill Gates during the pandemic. Using a topic modeling technique (i.e., Biterm Topic Model), this study identified ten salient topics surrounding Bill Gates on Twitter, and we further investigated the interactions between different topics using Granger causality tests. The results demonstrate that emotionally charged conspiratorial narratives are more likely to breed other conspiratorial narratives in the following days. The findings show that each conspiracy theory is not isolated by itself. Instead, they are highly dynamic and interwoven. This study presents new empirical insights into how conspiracy theories spread and interact during crises. Practical and theoretical implications are also discussed.
Introduction
As with the emergence of conspiracy theories during the crisis, the COVID-19 pandemic became a fertile ground for conspiracy theories.1–3 Since the COVID-19 outbreak in 2019, the spread of problematic information like fake news and misinformation through various communication channels has raised serious concerns.4–7 A recent report showed that false claims about the actions of political authorities were “the single largest category” of COVID-19 misinformation.1
Conspiracy theories refer to explanatory narratives about the ultimate causes of significant social and political events, with claims of secret plots by powerful actors.8–11 While philosophical debates have refuted the assumption of absolute falsehood in a conspiracy theory,12 it is widely agreed that conspiracy theories provide a way to reduce uncertainty by attributing the causes and outcomes of an unexpected situation to powerful people or elite groups in society. The primary danger of conspiracy theory lies in its reductionist worldview, which denies unwanted consequences as unavoidable parts of political and social reality.13 While unexpected circumstances have served as fertile ground for conspiratorial thinking throughout human history,1 COVID-19 conspiracy theories have been particularly harmful because they have not only disseminated reductionist beliefs but also lead to serious public health hazards, for example, by increasing vaccination hesitancy or rejecting scientific consensus.6,14,15
Recent literature on COVID-19 conspiracy theories has focused on who is more vulnerable to these theories,16 why people believe them,3,17 societal consequences,14,18 and how to detect COVID-19 conspiracy theories using computational techniques.19 Although the spread of conspiracy theories has received much attention, the academic understanding of how different conspiracy theories interact and evolve into a complex narrative system through social media conversations is relatively scarce. While an individual online message (e.g., a tweet) may not convey a full story by itself, an examination of a sizable collection of these messages as a whole may help reconstruct the narratives for collective sensemaking. Furthermore, the narratives of conspiracy theories about a crisis event (e.g., COVID-19) are never static.20 Especially, in the social media space, transmitting conspiratorial messages involves various users with different levels of engagement and even automated agents, making the process even more dynamic.
This exploratory study focuses on one well-known set of conspiracy theories, that is, the role of Bill Gates during the COVID-19 pandemic.21 We investigate the types of narratives and information spread over time to claim the covert influence of this economic elite on the creation and progress of the pandemic. To do so, this study collected 313,088 tweets surrounding Bill Gates over a nine-month period in 2020 (from March 31 to December 31, 2020). An unsupervised machine-learning technique (i.e., Biterm Topic Model) was used to identify salient topics, and Granger causality tests were used to examine the sequential progress of the narratives about Bill Gates.
Conspiracy theory, interactions, and beliefs
The term “conspiracy theory” typically has a negative connotation, and is often weaponized to attack counterparts, especially in political campaigns and unrests.20 In practice, people usually deny that their explanations are conspiracy theories, even though those explanations clearly qualify.22 Conspiracy theories have been ubiquitous throughout human history, especially when people are faced with crisis situations, and their messages have been closely suggestive of prejudice, witch hunts, revolutions, and genocide.22–24
Related to conspiracy theories, another important concept that should be noted is conspiracy belief, which generally refers to belief in a specific or set of conspiracy theories.25 Studies of conspiracy belief have shown that a higher level of belief in conspiracy theories is associated with sharing these theories both privately and in public spaces.26,27 The core motive underlying conspiracy beliefs is human beings' natural desires to make sense of the world, but conspiracy beliefs may also stem from a range of psychological, political, and social factors.28,29
Meanwhile, two perspectives have explained why beliefs across different conspiracy theories are correlated.30–32 On the one hand, studies have suggested “conspiracy mentality,” referring to people's general tendency to endorse conspiracy theories, as an overarching psychological mechanism for cross-conspiracy belief.33–35 According to this perspective, people's beliefs in conspiracy theories are interrelated, and several studies have shown that sometimes people even believe mutually contradictory conspiracy theories, alluding to a conspiracy mentality at play.36 Studies based on this perspective have shown that the conspiracy mindset is relatively stable.33,37
On the other hand, other studies have criticized the conspiracy mentality proposition, partly because it is difficult to distinguish between measures of belief in a specific conspiracy theory and those tapping into a more general conspiracy mindset.31 Instead, these studies maintain the “monological belief system” thesis that each conspiracy belief adopted by an individual reinforces other conspiracy beliefs, making the individual more receptive to conspiracy theories that he or she may encounter later.25,31 In the case of COVID-19, people who believe Bill Gates, for example, is associated with the origin of the virus would be prone to finding other conspiracy theories to be plausible. To summarize, the former suggests that a one-size-fits-all type of mindset breeds beliefs across various forms of conspiracy theories, whereas the latter suggests that belief in each of the different conspiracy theories will reinforce one another.3,22,25
While studies have examined these perspectives by taking an individual's belief as a unit of observation, little is known about the narrative systems in the “rabbit hole,”3 in terms of how conspiratorial stories are interwoven, interact, and evolve through social media conversations. Thus, by using conspiracy theories about the role of billionaire and philanthropist Bill Gates during the pandemic as a case study, this study tested the distribution and evolution of different conspiracy theories over a 9-month period. The following research questions guided this inquiry:
RQ1: How were the narratives of “the Bill Gates theory” told on Twitter? RQ2: How did different narratives interact with one another on Twitter?
Methods
Data collection and processing
We collected Twitter data in real time using the streaming application programming interface from March 31 to December 31, 2020. The beginning of data collection corresponds to the COVID-19 pandemic declaration by the World Health Organization in March 2020.38 We first collected COVID-19-related tweets, from which we filtered data to only include tweets with the word “Gates.” We then manually inspected a random sample of 6,000 tweets to refine our sampling strategy. For example, in some contexts, the word “gates” was used to mean “doors.” We iteratively adjusted our search parameters by adding filtering criteria. The final dataset included 313,088 tweets created by 148,605 accounts. On average, each account has posted about two tweets related to Bill Gates (M = 2.107, SD = 7.778, range: 1 to 1763). Figure 1 shows the daily volume of tweets collected during the sampling period.
FIG. 1. Time-series graph of Twitter frequency over the sample period.
Analytic strategy
This study uses a mixed-method approach, including topic modeling, thematic analysis, and statistical modeling. Topic modeling, the main analytic framework of this study, is an unsupervised machine learning method that creates coherent semantic themes from words' co-occurrence patterns in textual samples.39 Topic modeling is widely used when dealing with large samples of textual data.
Many studies show that conventional topic modeling approaches (e.g., Latent Dirichlet Allocation) do not perform well with short texts, such as tweets and short answers in surveys.39–41 To address this concern, this study employed Biterm Topic Modeling (BTM), a word co-occurrence (biterm) based topic method.42 BTM was used for this study because biterms directly model the co-occurrence of words, which increases performance for sparse-text documents and is particularly well suited for Twitter research.42,43 BTM analysis is done by setting the BTM topic number (k) and “n” words (for the first round of analysis, we set k = 10 and n = 15 to cover possible topics). A coherence score is also used to measure how strongly the top words from each topic correspond to their respective topics. After the iterative steps of running BTM and manual inspection of outputs, we selected 10 topic clusters (k = 10) as our final modeling output.
Results
Topic identification
As mentioned, after the iterative steps of running BTM and manual inspection of outputs, we selected 10 topic clusters as our final modeling output that addresses RQ1. Texts associated with each topic cluster collectively reflect a coherent theme with some references to actors and actions that are important elements of a narrative.
Topic 1 alluded that the outbreak of COVID-19 is a cover-up for Bill Gates to use vaccines to implant microchips for surveillance purposes. Topic 2 centered around who had created the virus, and it linked Bill Gates to China's Wuhan Lab. Related to Topic 1, Topic 3 comprised the narratives that 5G technology helped spread the COVID-19 virus and Bill Gates is involved. Topic 4 (trolling) predominantly contained expletives and emotional expressions that antagonized Bill Gates. For example, “Bill Gates is not a f**king doctor or scientist. He is a eugenics proponent who wants to decrease world population for the reset.” Topic 5 is a remix of stories about pandemics, broadly questioning who created the pandemic, the purpose of vaccines, and different actors' roles during the pandemic.
The narrative that Bill Gates was involved in vaccine tests in foreign countries, such as South Africa, was salient in Topic 6. Topic 7 revolved around the narrative that the COVID-19 vaccine will modify people's DNA, and Bill Gates is behind it. Topic 8 focused on the narrative that Bill Gates anticipated the coronavirus pandemic, and it linked the collusion between the World Economic Forum and the Gates Foundation to the outbreak of the pandemic.44
Topic 9 focused on the narrative that Gates played a role in developing the vaccines, which leads to severe consequences. Topic 10 focused on Gates reaping benefits from the crisis, including allegations that Bill Gates negotiated a $100 billion contact-tracing contract even before the pandemic began.45 To further validate the interpretation, a small sample of tweets from each topic (n = 100) was manually coded by two trained coders,46 Cohen's kappa = 0.87. Figure 2 shows the topic proportions in the dataset.
FIG. 2. Topic proportions (by percentage).
Topic distribution over time and Granger causality
A time-series analysis was performed to address RQ2, which asks how different narratives interacted with one another on Twitter. To do so, we calculated the relative salience of each topic on a particular day, based on which Granger causality was run to test the temporal relationships among topics. Our data consist the daily volume of tweets. Following previous studies, we consider days as an appropriate unit because our interest is in capturing the dynamics of attention to each topic over time. Figure 3 shows the relative topic salience across the sampling period.
FIG. 3. Topic salience over time.
To test the temporal interaction between different topics, we rely on the methodological framework of Granger causality. The Granger causality test is a statistical hypothesis test for determining whether one-time series is useful for forecasting another.47,48 A measure x (Topic A) can generally be said to “granger cause” a measure y (Topic B) if y can be predicted more accurately from previous values of x and y together than from past values of y alone.47 It is noteworthy that Granger causality provides forecasting information, while not indicating the presence of true causality between two variables.48 Furthermore, an important assumption of time-series analyses is stationarity (e.g., statistical properties such as mean and variance are constant over time). A visual inspection of our time series combined with the results of the Augmented Dickey-Fuller test suggests the series are stationary.49 Based on Granger causality premises, this study utilized statistical tests to confirm the appropriate number of lags.50
The results of Akaike information criterion criteria, which are commonly applied for lag selection,47 showed that 4 days represent an optimal lag for the analyses. The test is implemented in R with the function grangertest, as part of the package lmtest for testing linear regression models. Table 1 provides the results of the Granger causality tests and displays significant results with p value and the F-statistic.47Figure 4 shows the significant interactions among the topic pairs, with arrows indicating the directionality of Granger causality.
FIG. 4. A Granger causality graph of topics. Note. ***p < 0.001, **p < 0.01, *p < 0.05.
Table 1. Statistically Significant Granger Causality Relationships Between Topics From To F-statistic p Vaccine & Microchip Gates & 5G 3.0689* 0.0170 Vaccine & DNA change 3.9986** 0.0036 Gates & China Vaccine & microchip 5.2128*** 0.0005 Vaccine & DNA change 2.9019* 0.0224 Gates tests in foreign countries 5.4557*** 0.0003 Tracing deal before pandemic 2.4733* 0.0449 Gates & 5G Vaccine & DNA change 3.6609** 0.0064 Trolling Gates & 5G 2.7298* 0.0297 Vaccine & DNA change 2.5589* 0.0392 Gates & vaccine concerns 3.7069** 0.0059 Remix with stories about Pandemic 5.3163*** 0.0004 Remix with stories about Pandemic Gates tests in foreign countries 5.2713*** 0.0004 Vaccine & microchip 4.0394** 0.0034 Vaccine & DNA change 3.0036* 0.0189 Gates & vaccine concerns 2.4408* 0.0473 Vaccine & DNA change Gates & 5G 4.0642** 0.0032 Vaccine & microchip 5.5577*** 0.0003 Gates' prediction before pandemic Vaccine & DNA change 3.1672* 0.0145 Gates & vaccine concerns Vaccine & DNA change 2.972* 0.0200 Tracing deal before pandemic Gates & China 3.4897** 0.0085 Remix with stories about pandemic 2.9766* 0.0199
Overall, this study found 21 significant Granger-causal pairs out of the potential 90 topic pairs. The results suggest that emotional expressions (Topic 4) and baseline conspiracy narratives (e.g., Topic 1, Topic 2, Topic 8) preceded the distribution of more “advanced” conspiracy narratives in the following days. The findings that frustration and anger triggered conspiracy theories, as opposed to the other way around, allude that conspiracy theory is fundamentally emotional in nature, rather than analytical.20 When people face uncertainty and insecurity during a crisis, they seek a simplistic explanation to reduce anxiety, which often leads to the creation, belief, and distribution of conspiracy narratives.
In addition, the results showed that instead of a static process, the evolution of the conspiracy narrative somehow reflects the progress of reality. For instance, general conspiracy theories about the origin of viruses (Topic 2, Topic 8) tended to lead to more detailed allegations about vaccines, microchips, and 5G (as shown in Fig. 4). Furthermore, as the pandemic progressed, emerging issues were constantly being brought into the storytelling, while the main actor (Bill Gates) was still at the center.
Discussion
Social media platforms provide excellent opportunities to keep people updated about a risk or a crisis. That being said, unregulated content sharing also amplifies the infodemic and cultivates a breeding ground for falsehoods and conspiracy theories.2–4 By analyzing large-scale tweets (N = 313,088) over a 9-month period in 2020, this study dives into the understanding of narrative components of popular conspiracy theories about Bill Gates during the pandemic. Specifically, this study identified ten salient narratives around Bill Gates and analyzed how they interacted over the sampling period.
The results of Granger causality tests showed that there is a strong interaction among different conspiratorial narratives. Such a finding is in line with the existing literature,20,22,36,51 which suggests that the distribution of one conspiracy theory is more likely to trigger another. Not just triggering another theory, our findings allude that different conspiratorial stories converge to create a more complex narrative (e.g., stories of China combined with Bill Gates). That is, the findings demonstrated that each conspiracy theory is not isolated by itself. Instead, they are highly dynamic and interwoven in ever-evolving digital social conversations.
Moreover, the finding also showed that some “baseline” conspiracy narratives (e.g., Topic 2, Topic 8) preceded the distribution of more “advanced” conspiracy narratives in the following days. We cautiously interpret this finding in several ways. First, instead of a static process, the evolution of the conspiracy narrative mirrors the progression of events in reality. As the pandemic progressed, emerging issues (e.g., vaccines, microchips, and 5G) were constantly incorporated into the storytelling process. Second, with the increasing cases of COVID-19, people's attention may have been shifting from the origin of viruses (Topic 2, Topic 8) toward things that are more related to their lives, especially the safety and effectiveness of vaccines.
In addition, another notable finding from this study is that emotional expressions (Topic 4) were more likely to trigger the distribution of other conspiracy theories in the following days. In line with several prior studies,3,20,35 our findings partially supported the idea that conspiracy theories were primarily emotional and intuitive. When people are faced with insecurity and ambiguity during a crisis, they tend to actively seek a simple explanation to alleviate negative emotions, which often leads to strong beliefs and the spread of conspiracy stories. Furthermore, such findings are consistent with online virality and diffusion literature, which indicates that online content that evokes high-arousal emotions (e.g., anger or excitement) tends to be more viral than low-arousal emotions (e.g., sadness).52,53 Especially, during uncertain crises like pandemics, tweets that contain aggressive expressions of distress can spread high arousal of negative emotions on a collective level.54,55
The networked public would then search for the reason behind their distressful emotional state, sometimes landing on misleading conclusions like conspiracy theories. Moreover, cultural attractor theory suggests that certain content properties (e.g., content eliciting threat and danger) play critical roles in the distribution of cultural variants (e.g., beliefs, narratives, misinformation).56,57 Compared to other types of content, conspiracy theories often include sensational narratives and high psychological attractiveness; such content characteristics also make them more viral.56
Furthermore, in the case of Bill Gates, this study did not find contradictory conspiracy theories; instead, most conspiracy theories that emerged in our dataset were complementary to one another. Such findings allude that individuals' beliefs in one conspiracy theory may reinforce another, which leads to more sharing behaviors in digital space. However, this study's findings may not invalidate the idea of a conspiracy mentality either, because it is also possible that the presence of multiple conspiracy theories in our dataset may also attest to a shared worldview that social events are orchestrated by powerful forces. Given that the analysis of collective narratives alone is insufficient to conclude the psychological mechanisms of conspiracy beliefs, future research should further explore the relationship between the social sharing of conspiracy theories and beliefs. Integrating individual-level analysis with digital data may shed light on the debate between monological belief systems and conspiracy mentality.30–33
This study presents new empirical insights into how conspiracy theories evolve during crises. However, several limitations are worthy of further attention. First of all, this study focused on Twitter. Compared with other social media platforms, Twitter is a much more open platform and has unique characteristics in its user base.58 Future studies are worthwhile to test whether the findings apply to other platforms (e.g., Facebook, Reddit). Future researchers should also continue exploring whether conspiracy theories on social media will have a ripple effect on the agendas and narratives of traditional media, and vice versa.59
Furthermore, this study did not test the role of super-spreaders in the information distribution process. Recent studies have revealed the role of bots—automated accounts—in the spread of misinformation. Shao et al found that, during the 2016 U.S. presidential election on Twitter, social bots played a key role as “super-spreaders” in distributing articles from low-credibility sources (e.g., false news, conspiracy theories).60 Other studies also showed that conspiracy theory-sharing behaviors also differ significantly by individuals' political leanings.5,31,61 Given the potential role of hyperactive accounts—“super-spreaders”—in distributing information and forming public opinions, understanding who they are, and what they are posting may help combat the spread of harmful information on social media platforms.
Thus, future studies should consider incorporating account-level analysis (e.g., the number of retweets, followers, and comments) to identify the super-spreaders in the online community. Even more intriguing would be to apply social network metrics such as degree centrality and link similarity to identify users who have the most important positions in the discussion network.62–64 Finally, given the importance of visuals in the digital narrative, future studies should consider incorporating visuals into the analysis framework. In an era of intense information warfare, continuing to explore the evolution and interactions of different conspiracy theories remains a worthwhile scholarly pursuit.
Acknowledgments The authors would like to thank the editor and anonymous reviewers for their valuable and insightful comments on our article.
Author Disclosure Statement
No competing financial interests exist.
Funding Information
This project was supported, in part, by the National Science Foundation under grant 2027387, Massachusetts Institute of Technology (MIT) under grant PO 7000506684, and ASU ISSR's internal seed grant.