Conspiracy in the time of corona: automatic detection of emerging COVID-19 conspiracy theories in social media and the news

Abstract

Rumors and conspiracy theories thrive in environments of low confidence and low trust. Consequently, it is not surprising that ones related to the COVID-19 pandemic are proliferating given the lack of scientific consensus on the virus's spread and containment, or on the long-term social and economic ramifications of the pandemic. Among the stories currently circulating in US-focused social media forums are ones suggesting that the 5G telecommunication network activates the virus, that the pandemic is a hoax perpetrated by a global cabal, that the virus is a bio-weapon released deliberately by the Chinese, or that Bill Gates is using it as cover to launch a broad vaccination program to facilitate a global surveillance regime. While some may be quick to dismiss these stories as having little impact on real-world behavior, recent events including the destruction of cell phone towers, racially fueled attacks against Asian Americans, demonstrations espousing resistance to public health orders, and wide-scale defiance of scientifically sound public mandates such as those to wear masks and practice social distancing, countermand such conclusions. Inspired by narrative theory, we crawl social media sites and news reports and, through the application of automated machine-learning methods, discover the underlying narrative frameworks supporting the generation of rumors and conspiracy theories. We show how the various narrative frameworks fueling these stories rely on the alignment of otherwise disparate domains of knowledge, and consider how they attach to the broader reporting on the pandemic. These alignments and attachments, which can be monitored in near real time, may be useful for identifying areas in the news that are particularly vulnerable to reinterpretation by conspiracy theorists. Understanding the dynamics of storytelling on social media and the narrative frameworks that provide the generative basis for these stories may also be helpful for devising methods to disrupt their spread.

Keywords: 4Chan; 5G; Bill Gates; Bio-weapons; COVID-19; China; Conspiracy theories; Corona virus; Data visualization; Machine learning; Narrative; Networks; News; Reddit; Rumor; Social media.

Fig. 1

Automated pipeline of processing data and discovering narrative networks in social media and news reports

Fig. 2

Overview graph of the largest thirty communities in the social media corpus. Nodes are colored by community, and sized by NER score. Narrative frameworks are drawn from these communities, each of which describes a knowledge domain in the conversation. Nodes with multiple community assignments are colored according to their highest ranked community. An overarching narrative framework for a conspiracy theory often aligns subnodes from numerous domains

Fig. 3

Progressive attachment of “coronavirus” to “conspiracy theory” in the co-occurrence network of news reports conditioned on entities found in social media: The orange-outlined nodes represent the two concepts, as they gravitate toward one another over time and form new simple paths. From top to bottom, 5-day intervals starting on January 20, 2020, March 15, 2020, and April 10, 2020. Nodes are colored as follows: Celebrities in yellow, media outlets in red, important actants in pink (manually colored), places in green and corporations/entities in black

Fig. 4

Number of common neighbors between “coronavirus” and “conspiracy theory” over time in the news reports: Across all 101 segments of 5-day intervals, the number of simple paths empirically increases rapidly, suggesting the closer ties between the two entities across time

Fig. 5

Cross-Correlation of Relative Coverage Score for Word-Level Community Hits in social media against the news reports: The mean and standard deviation of the relative coverage score are computed per time stamp across 20 trials with 500 community members each. The peak at 0 days offset suggests that social media and the news are intertwined in a very responsive manner

Fig. 6

Scores of (a) Homogeneity, (b) Completeness, (c) Coverage and (d) V-Measure are provided to compare News based communities with Social Media communities. Here we used $Y_{\mathit{pred}}$ and $Y_{\mathit{gr}}$ derived in algorithm 3 as our cluster label and classes. Completeness measures how members of a given class are assigned to the same cluster, while homogeneity measures how each cluster contains only members of a single class. Their harmonic mean is the V-Measure [73]. Coverage percentage is the fraction of actants in news report communities that also are found in social media network communities

Fig. 7

Communities with index 5, 56 and 82 sequentially describe the conspiracy theory surrounding “Bill Gates” and “5g”. The words in bold are the sub-nodes present in the narrative network and the yellow-highlighted phrases are automatically extracted relationships between the sub-nodes. The blue-highlighted sub-node is a key actant that exists in all 3 communities and is one of the connecting components between “Bill Gates” and the conspiracy theory around “5g”. Community 5 describes Gates’s supposed obsession with population control along with his funding of faulty research. The same research is alleged to have created “5g” as a means of spreading the “virus” which is allegedly intended as a “bioweapon”. Community 56 takes it a step further tying “5g” to its carrier frequency and the associated interactions of this frequency with the human body. Community 82 concludes the origin story of the virus (back to the “faulty” research conducted by “Gates”) and mentions the cell-level interaction between the virus and the body

Fig. 8

The histogram of threat scores across the sub-nodes from the phrase classifier. The bi-modality encourages binary classification thresholds around 0.2. In our networks, we use 0.25 which is at the 57^th percentile of sub-nodes classified as threats

Fig. 9

The sub-node “CCP” has associated noun phrases shown in the gray box. The noun phrases have descriptive SVCop relationships, whose descriptive phrases are sampled in the light red and green blobs. The phrases in the red blob are classified as threats by our majority classifier and the phrases in the green blob are classified as non-threats. The highlighted and bold descriptive phrases are sample phrases for which the nearest neighbors are shown. The kNN classifier reasonably clusters phrases that are syntactically different but semantically similar using the BERT embedding. Darker nearest neighbors occur more frequently

Fig. 10

A conspiracy theory narrative framework that links the virus to 5G, Bill Gates, and vaccination. Nodes have been scaled by NER mentions; those with fewer than 250 mentions have been filtered for the sake of clarity. Nodes are colored by community, and outlined with red if they represent a threat

Fig. 11

Communities comprising the narrative framework suggesting that the virus is a result of Chinese wet markets and deliberate information cover-ups. The narrative framework focuses heavily on markets, exotic animals such as pangolins, and the role of Chinese Communist Party in hiding information about the initial outbreak. Nodes are colored by community, and outlined with red if they represent a threat. The graph is filtered to show nodes with degree geq 2

Fig. 12

Communities comprising the COVID-19-as-bioweapon narrative framework. The narrative framework focuses heavily on laboratories and the potential role of the virus as a weapon. Nodes are colored by community, and outlined with red if they represent a threat. The graph is filtered to show nodes with degree geq 2

Fig. 13

The communities comprising the globalist hoax narrative framework: Here, a globalist cabal has conspired to foist the hoax of the Corona virus on the world, with the virus presenting with mild flu-like symptoms. Trump and his allies are fighting against the Democrats and their surrogates to stave off the economic impact of the hoax. Nodes are colored by community, and outlined with red if they represent a threat. The graph is filtered to show nodes with degree geq 2. Two nodes, “filmyourhospital” and “hoax,see global warming” have been highlighted in yellow

Fig. 14

A narrative framework that can be deployed by multiple groups. The framework focuses on the relationship between the virus, SARS, the flu and the testing regimen. Also included are nodes representing research on the virus and questions of immunity. A small disconnected component on the filtered graph provides a critique of QAnon and Glenn Beck. Nodes are colored by community, and outlined with red if they represent a threat. The graph is filtered to show nodes with degree geq 2