We began collecting and curating a COVID-19 Twitter data set right at the beginning of the pandemic, in January 2020, to continuously track, in real time, public discourse about the coronavirus pandemic. We have made the data set publicly accessible to the wider research community [12]. This study uses publicly available data, and the data collection and analysis are approved by the University of Southern California Institutional Review Board (protocols UP-17-00610 and UP-21-00005).

Our complete data set, as of this writing (mid-July 2021), contains 1,497,893,426 tweets from January 21, 2020, through July 9, 2021 (release v2.55). While we provide a brief overview of our data set here, a full description of our data set can be found elsewhere [12]. We leveraged release v2.45 for this paper, which contains 1,443,871,621 tweets from January 21, 2020, through April 30, 2021. All our tweets were collected in real time using Twitter’s streaming application programming interface (API), which gave us access to a 1% stream of tweets [13]. We leveraged a manually curated list of keywords to filter for tweets that contained content related to the COVID-19 pandemic and surrounding issues. We list a sample of the keywords we tracked in Table 1. The full list of up-to-date keywords can be found in our GitHub repository [14]. While we did our best to capture as much discourse as we could in our collection, a limitation of our data set is that our keywords were all in English and were manually selected for tracking. This may have influenced the collected tweets and our subsequent observations. A language breakdown for the tweets found in release v2.45 can be found in Table 2.

To identify conspiratorial and questionable tweets, we used the following 2 lists compiled by MBFC: conspiracy-pseudoscience sources and questionable sources. MBFC is “an independent website that rates the bias, factual accuracy, and credibility of media sources” [17]. MBFC classifies domains as conspiracy-pseudoscience if the domain “may publish unverifiable information that is not always supported by evidence. These sources may be untrustworthy for credible or verifiable information” [17]. For the sake of brevity, we also refer to these conspiracy-pseudoscience domains as simply conspiracy or conspiratorial domains. MBFC states that questionable sources are domains that “exhibit one or more of the following: extreme bias, consistent promotion of propaganda/conspiracies, poor or no sourcing to credible information, a complete lack of transparency and/or is fake news. Fake news is the deliberate attempt to publish hoaxes and/or disinformation for the purpose of profit or influence” [17].

We also obtained a set of randomly selected sources by taking a random sample from the set of media sources that appeared in the full data set. We called this set of sources “random sources.” The set of random sources has the same number of elements (URLs) as conspiratorial and questionable sources. The random sources served as a baseline for comparison with conspiratorial and questionable sources.

MBFC also classifies media domains by their political affiliations, with the following 5 political affiliation categories: left bias, left-center bias, least biased, right-center bias, and right bias. We used their lists of domains to identify tweets with a particular political affiliation. Left and right bias sources are “moderately to strongly biased,” may be untrustworthy, and can “publish misleading reports and omit reporting of information that may damage [their] cause” [17]. Left-center and right-center bias sources have “slight to moderate” bias and are “generally trustworthy for information but may require further investigation” [17]. MBFC goes on to describe sources tagged as least biased as sources with “minimal bias,” “factual and usually sourced,” and “the most credible media sources” [17].

For every user in our misinformation subset, we tabulated the number of times they shared domains and identified the political bias of these domains. This gave us a proxy of each user’s political lean. The political lean was determined by the political lean of the majority of a user’s shared domains. In the case of a tie between 2 political biases, we randomly assigned the user a political bias. Any user who shared one or more questionable or conspiracy-pseudoscience domains (as identified by MBFC) within our data set was considered to have engaged with misinformation. This does not mean that a user in our misinformation subset exclusively or mostly shared misinformation content. We restricted our analysis to only users who had shared more than five URLs.

Taking advantage of the retweeting dynamics of Twitter, we constructed a network to conduct social network analysis on the users in our misinformation subset. Nodes represent users and links (or ties) represent retweets between users. If user A (retweeting) retweets user B (retweeted), then the strength of their tie increases with the frequency of retweets. To visualize this network, we adopted a force-based algorithm, Force Atlas [18], which plots nodes that share strong links close together. For the sake of clarity, the ties are not explicitly shown. There were a total of 4,164,572 users and 22,894,165 unique ties between users in our misinformation subset. We labeled the most prominent users, sorted by their highest out-degree. 

This retweet network is constructed from the tweets of users who had retweeted at least one tweet that contained a domain that MBFC had classified as a questionable or conspiracy-pseudoscience domain. This means that each link between a retweeted and retweeting user does not necessarily mean that the retweet contained a misinformation domain or that the retweeted user engaged with a misinformation domain. Thus, the entire retweet network (contained within our dataset) included users who had interacted with a misinformation domain at least once.

We analyzed the content coming from the following 3 groups of sources, each containing 250 URL domains: conspiracy sources, questionable sources, and random sources. Conspiracy and questionable sources were domains classified as such by MBFC, whereas random sources were chosen from a set of URLs selected at random to serve as a baseline for comparison.

To calculate the temporal trends in the amount of news coming from unreliable sources, we performed 2 multiple linear regression analyses using standard ordinary least-squares models. The first model estimated the association between the number of conspiratorial URLs and time, adjusting for an average number of URLs observed on a platform. The model can be represented as follows: V_C ~ tβ₁ + V_R β₂, where V_C is the number of conspiratorial URLs shared, t is time measured in days, and V_R is the number of random URLs shared on Twitter. The second model estimated the association between the number of questionable URLs and time, adjusting for an average number of URLs on a platform. Similarly, it can be represented as follows: V_Q ~ tβ₁ + V_R β₂, where V_Q is the number of questionable URLs shared.

To better understand the relative impact of unreliable sources, we looked at their respective audiences and the communities that formed around sharing these unreliable sources. It is important to quantify the community structure and relationships between the consumers of certain kinds of information, as the strength of these communities can be indicative of the potential of an idea within the community to grow and become dominant over time. According to organizational theory, interpersonal networks that exhibit densely configured ties have a greater likelihood of attaining their goals and retaining the network structure (committed to staying together). Networks of strong ties are also significantly more robust with respect to the connectivity and small world property of social networks [19,20].

To quantify the relative strength of a connection between information sources that spread unreliable information about COVID-19, we constructed 3 networks of the following group of domains as defined earlier: conspiracy, questionable, and random sources. The nodes in the network represent the domains, and a link was drawn between 2 domains if a user shared content from both domains. The weight of a link was set to the number of users who shared both domains. To quantify the density of connections in these networks, we calculated the average clustering coefficient [21] and the average link weight for each respective audience network.

We used a clustering strategy based on dipm-SC [15], described in the Methods section, to identify topics that exhibit similar temporal behaviors and group them into distinct clusters. The detected clusters are visualized in Figure 2. We found that all clusters exhibited distinct peaks, suggesting minimal overlap between distinct clusters and hence robust and reliable clustering results. We now briefly describe the key topics that were detected in the 11 clusters we identified.

We then turned to investigating misinformation and questionable narratives that spread in the context of COVID-19. We used our misinformation data subset, which contains tweets with URLs whose domains were deemed to be from a conspiracy-pseudoscience or questionable source according to MBFC, and leveraged both dipm-SC [15] and LDA [16] to cluster tweets by general topic. From the topics found in both clustering methods, we identified the following 3 major misinformation narratives that encapsulate the tweets that spread questionable media content on Twitter: (1) hydroxychloroquine and alternative medicines, (2) US officials and governing agencies, and (3) COVID-19 prevention interventions.

For each narrative of interest, we filtered our misinformation data set based on several defining keywords (Table 3). We identified the keywords in Table 3 by first isolating the most used keywords and bigrams in each narrative’s cluster, and then manually selecting neutral keywords most reflective of the 3 narratives. This enabled us to isolate subsets of tweets that specifically mentioned keywords related to each misinformation narrative. We then plotted the volume of tweets from each narrative over time (Figure 3) to understand temporal trends in each narrative. We found that a constant flow of misinformation exists, despite Twitter’s efforts to mitigate its spread. However, when we isolated tweets by narratives, we saw that each narrative experiences differing levels of engagement over time. Most of these spikes are driven by active retweeting of viral posts and/or articles that are sometimes related to real-time events. For each narrative, we also found the top hashtags that were used and grouped them into their relevant categories. We did a manual inspection of the tweets during these peaks and describe a few of the prominent topics that drove the volume surges in each narrative as seen in Figure 3.

After identifying and describing the misinformation narratives permeating online discourse, we looked to understand the audience that is more susceptible to misinformation and the trends within the kind of misinformation that is being consumed. In the following text, we used network science as a lens to understand the structure and characteristics of misinformation echo chambers on Twitter, and suggest this as a possible mechanism to explain the spread of misinformation in specific communities.

In this paper, we provide a comprehensive overview of public COVID-19 discourse on Twitter by analyzing 1.4 billion COVID-19–related tweets that spanned the course of over a year. We make several important contributions in this work.

First, using temporal clustering of bigrams, we report 11 major topics of discussion. Aside from 1 topic with general COVID-related phrases that had sustained interest throughout our study period, the rest of the 10 topics were bursty and closely aligned with the progression of current events. We observed 2 types of topics. The first type included political topics that arise due to congregation, such as the protests that occurred in the wake of George Floyd’s death, Trump’s rallies, and India’s national exams. The second type encompassed news events that generated significant online traction, such as Trump testing positive, vaccine updates, and the relief bill. This demonstrates that observing Twitter usage is a valid way to monitor public sentiment and important events as they unfold in the real world.

We then identified misinformation narratives by analyzing latent topics detected from tweets that shared domains that have been identified as unreliable media sources. We found that the following 3 prominent misinformation narratives emerged: hydroxychloroquine and alternative medicines, US officials and governing agencies, and COVID-19 prevention practices. Each of these narratives experienced surges in mentions and engagement, the majority of which occurred in tandem with and in response to real-world events occurring at the same time.

We also characterized misinformation adoption by analyzing the retweet social network structures of users who had retweeted at least one tweet that contained a domain classified as unreliable by MBFC. We found that there exists an alignment between the misinformation topic a user tends to engage in and that user’s political party. A large portion of the left-leaning userbase engaged specifically in COVID-19 prevention misinformation. The right-leaning userbase discussed COVID-19 prevention in the context of alternative medicines (such as hydroxychloroquine), and US officials and governing authorities. Interestingly, we observed a fracture in the right-leaning user base. Some users primarily discussed only 2 of the identified narratives (COVID-19 prevention and US officials), while others engaged with tweets surrounding all 3 narratives.

Lastly, and of great concern, we found that engagement with unreliable sources is increasing at a faster rate compared to engagement with our baseline of random sources. Our results show that, in the space of public health messaging on social media platforms, there is still significant work that needs to be done in order to combat misinformation. Although social media platforms are making efforts to stem the flow of misinformation and raise awareness of its presence, the dangers of misinformation, particularly surrounding public health, are increasingly apparent. In our network, there are dense and highly connected communities that form around unreliable sources (so-called misinformation bubbles [73]), which can serve to further promulgate health misinformation online.

Our study highlights how social media platforms can help us to shed light on the issue and consequences of misinfodemics, particularly during an unforeseen global health crisis. Social media platforms, such as Twitter, currently employ various tactics to counter misinformation, including the use of automated misinformation tags to raise awareness and partnerships with third-party fact checkers. Our research suggests that, while efforts are being made to mitigate misinformation, misinformation continues to be a mainstay on Twitter and is still growing in prevalence in the narratives we detected on online social platforms. We can also continue to understand the kinds of communities that form around sharing unreliable sources. In particular, we found that misinformation echo chambers exist within the COVID-19 misinfodemic landscape, and that the major echo chambers align with users’ political affiliations (as determined by the political lean of the sources they engage with). This has significant implications for how we can use unreliable domain usage to not only identify more communities that are susceptible to misinformation, but also funnel resources and develop strategies to combat misinformation flow in these communities.

While our study leverages a large tweet data set, there are still several limitations that need to be considered when interpreting the results of our study. First, when collecting data through Twitter’s free API, we were only able to collect 1% of all tweets in real time. Even with this limitation, we were able to collect several million tweets each day. We also only conducted our study on Twitter, which has been found to be used in the United States by a more liberal and left-leaning audience [74].

Due to the ever-evolving nature of misinformation, it is difficult to accurately judge and tag individual stories on Twitter as being misinformation or not. Thus, we used MBFC’s list of unreliable domains and the domains a user decides to share as a proxy for misinformation and engagement with a known unreliable source. This, however, does not necessarily mean that every URL shared from these domains has misinformation.

We did not focus on delineating social bots from human users in our analysis [75]. The term social bot generally refers to an account that is automated through software, and detecting and characterizing bot behavior is an active research area on its own [76]. Bots are incredibly salient to the misinfodemics conversation and have been found playing roles in the perpetuation of misinformation on social networks [75,77-79]. However, this study focused on the content and veracity of narratives shared on Twitter, and we hope to explore automated manipulation in the context of infodemics in future expansions of this work.

In this paper, we analyzed over 1 billion tweets posted during the COVID-19 pandemic and about the pandemic, spanning the course of over a year. We described the major topics of discussion that occurred over the broader COVID-19 Twitter discourse and identified the primary misinformation narratives that permeated the Twittersphere. We demonstrated that there are distinct misinformation echo chambers that form around specific topics and narratives, and that these echo chambers are also political echo chambers. This suggests that these echo chambers are driven by not only misinformation narratives, but also political alignment. Finally, we brought awareness to the increasing presence and consumption of unreliable content on Twitter, despite the current efforts being made to mitigate misinformation spread.

The COVID-19 pandemic and subsequent lockdowns around the world forced much of our forms of communication online, creating an environment where misinformation could more easily target a wider audience. We hope that our work will provide valuable insights into which communities are more susceptible to misinformation and contribute to laying the groundwork for other researchers in the field of misinfodemics.