COVID-19 vaccine conspiracy theories, discourses of liberty, and “the new normal” on social media

1 Vaccine conspiracy theories and dis/misinformation

Vaccine confidence is underpinned by public trust that vaccines work, are safe, and are produced by a trustworthy source. Vaccine confidence thus encompasses “trust in the vaccine (the product), trust in the vaccinator or other health professional (the provider), and trust in those who make the decisions about vaccine provision (the policy-maker)” (Larson et al. 2015). Vaccine hesitancy may emerge as a result of a loss of confidence in vaccines with respect to “(i) the effectiveness and safety of vaccines; (ii) the system that delivers them, including the reliability and competence of the health services and health professionals; and (iii) the motivations of policy-makers who decide on the needed vaccines” (MacDonald 2015: 4162). As such, if public trust in vaccines is undermined, it may influence vaccine confidence and result in vaccine hesitancy.

Public distrust in government, pharmaceutical companies, healthcare professions, and medical science and technology has been consistently linked with vaccine hesitancy and rejection and thus policymakers are keen to understand it (Attwell et al. 2017). Some modern examples of vaccine hesitancy or refusal (often resulting from contact with, or stated belief in, misinformation or rumours) have seen the resurgence of – or inability to inoculate against – life-threatening illnesses. Polio saw a resurgence in Nigeria following boycotts based on, amongst other factors, rumours suggesting the oral polio vaccine was linked to cancer, HIV, and sterility (Ghinai et al. 2013). Polio remains endemic in Pakistan and Afghanistan due to “parental refusal of polio vaccination, conspiracy theories, and misinformation which have rendered the polio eradication initiatives futile” (Ittefaq et al. 2021: 480). Finally, numerous countries previously free of endemic measles (Albania, Brazil, the Czech Republic, Greece, Mongolia, the UK, and Venezuela) have reported a resurgence in transmissions (Bozzola et al. 2020; Durrheim 2020). In the case of the UK, Andrew Wakefield’s now-debunked research that suggested links between the MMR (measles, mumps, and rubella) vaccine and autism has resulted in widespread anxieties, increasing vaccine refusal and by extension measles outbreaks (Larson et al. 2015).

As noted by Hardaker et al. (2024]: 164), “resistance to vaccination has existed as long as vaccination itself” and this resistance has taken the form of organized protests (and even riots) at the local and national levels. The internet has enabled the rapid distribution and consumption of (dis/mis)information about vaccines and made possible the worldwide coordination of anti-vaccination movements, which can serve to undermine public trust in vaccination and result in widespread vaccine refusal. Social media specifically has proven to be a prominent avenue for the distribution of vaccine misinformation (Suarez-Lledo and Alvarez-Galvez 2021).

Related to vaccine dis/misinformation, vaccine conspiracy theories may also serve to undermine confidence in vaccine products, providers, and policymakers. Drawing on Grieve and Woodfield (2023]: 14), we position conspiracy theories here as a form of fake news as they may draw variously on misinformation (Grieve and Woodfield’s Type I Fake News, i.e., unintentionally inaccurate information such as errors), disinformation (Type III Fake News, i.e., true but intentionally deceptive information such as omissions), or both simultaneously (Type II Fake News, i.e., intentionally false news, that is, lies). Conspiracy theory beliefs may be premised on little access to (or ignorance of) information that is either veracious or honest but represent a rational attempt to understand complex phenomena and deal with feelings of powerlessness, for example. Studying the effects of anti-vaccine conspiracy theories on intentions to vaccinate, Jolley and Douglas (2014]: 6) have found that “anti-vaccine conspiracy theories may have more than a trivial effect on vaccination intentions” and that “anti-vaccine conspiracy theories appear to introduce undue suspicion about vaccine safety, and increase feelings of powerlessness and disillusionment, whilst decreasing trust in authorities, which in turn introduce reluctance to vaccinate”. Anti-vaccination misinformation and conspiracy theories feature a range of tropes, including that vaccines are: toxic (containing foreign DNA, aborted fetal tissue, or formaldehyde), thus harmful (citation of previously harmful vaccines or medical treatments such as thalidomide is common) and linked to autism (Kata 2012); linked to genetic modification (Lyons et al. 2019) and DNA alteration (Reuters Fact Check team 2020); and a means of population control through sterilization (JitsuVAX n.d.) or even genocide (Genocide Watch 2021). Underpinning these tropes in anti-vaccination arguments runs a general distrust of science and government but, particularly, pharmaceutical companies who are accused of bribing researchers “to fake their data, cover up evidence of the harmful side effects of vaccines, and inflate statistics on vaccine efficacy” for financial gain or some other sinister motive (Jolley and Douglas 2014: 1).

The COVID-19 pandemic presented a range of social, medical, and political challenges stemming from the rapid generation and spread of both novel and recontextualized anti-vaccination tropes and conspiracy theories, such as the idea that COVID-19 was created as a biological weapon in China (Pennycook et al. 2020) or that COVID-19 is a false pandemic engineered to administer vaccines containing microchips to track vaccine recipients (Carmichael and Goodman 2020). Given that conspiracy theories are inherently social phenomena and that their meanings are situated in specific social contexts (Bergmann et al. 2020), the aim of this paper is to explore vaccine-related conspiracy theories in the context of the COVID-19 pandemic. One way to gain a better understanding of social phenomena is to explore language used about them. Given the increasingly important role of social media for our social lives, especially during the lockdowns of COVID-19, this paper investigates tweets about vaccination posted during the COVID-19 pandemic that include hashtags that make reference to various forms of dis/misinformation (including conspiracy theories). It should be noted that while these tweets contribute to narratives specific to various conspiracy theories, they do not necessarily constitute misinformation or disinformation.

2 Data: TRAC:COVID

The data are drawn from TRAC:COVID (Trust and Communication: A Coronavirus Online Visual Dashboard; Kehoe et al. 2021). The primary output of the TRAC:COVID project is a freely accessible online dashboard (https://traccovid.com) that combines tools from corpus linguistics with various visualization techniques and enables users to query a corpus of 84,138,394 tweets (i.e., posts on the social media platform Twitter, now known as X) containing reference to COVID-19. Only tweets published in the English language and in Great Britain between 1 January 2020 and 30 April 2021 were collected. TRAC:COVID is intended for use by a wide audience, including members of the general public, researchers, policymakers, and journalists to better understand how social media was used during the pandemic to talk about COVID-19. TRAC:COVID, therefore, facilitates exploration of language use in Britain during the COVID-19 pandemic.

Although TRAC:COVID can be used to explore myriad linguistic and social phenomena as realized in tweets about COVID-19 posted during the pandemic, this paper concentrates specifically on tweets discussing vaccines and conspiracy theories as a way to understand vaccine-related conspiracy theories. This paper builds upon the work of McGlashan et al. (2021), which inductively identified 276 dis/misinformation-related hashtags in tweets containing reference to vaccines from the underlying data of TRAC:COVID. As such, where McGlashan et al. (2021) identified the scale and variety of these hashtags, this paper investigates in more detail the contents of those tweets that contain both reference to vaccines and dis/misinformation (via these hashtags).

Specifically, we concentrated on tweets and retweets that include the terms *vaccin* or *vax*. Figures 1 and 2, visualizations from https://traccovid.com for the searches “vaccine” and “vax” respectively, show that use of both terms began to increase in frequency of use per day after 9 November 2020. Increased frequency would suggest an increased interest.

Figure 1:

TRAC:COVID timeseries: percentage of tweets per day including the term “vaccine”.

Figure 2:

TRAC:COVID timeseries: percentage of tweets per day including the term “vax”.

In terms of the size of our corpus (see Table 1), there were 2,658,495 tweets and retweets containing the terms *vaccin* or *vax*, but duplicate tweets accounted for 22.63 % of the corpus. Removing duplicates left 2,056,747 remaining and, following removal of retweets (which further add to duplication in the corpus), the corpus comprised 1,499,885 unique tweets (or 56.42 % of the initial corpus of 2,658,495 tweets and retweets). Unique refers here to tweets that are not total duplicates of any other tweet in the corpus.

Tweets in our corpus were then tagged based on their inclusion of hashtags relating to six areas of COVID-19 dis/misinformation (see Table 2) as identified by McGlashan et al. (2021). Our final corpus comprises 35,956 unique tweets containing at least one hashtag relating these categories of dis/misinformation (see Table 3).

3 Methods: keyword co-occurrence analysis

The principal method employed here is keyword co-occurrence analysis (KCA), which, as described in Clarke (2023) and Sha and Clarke (2025), is a method for examining keywords – lexical items (types) that are found to occur with an unusual token (in)frequency when comparisons are made between a target corpus and a reference corpus – by using multiple correspondence analysis (MCA). MCA is a geometric data analytic method which identifies relationships between three or more categorical variables. In the context of KCA, these categorical variables are a binary measure of whether a keyword is present or absent in the texts of a corpus and MCA is used to identify keywords that co-occur often in the texts of the corpus.

Sha and Clarke (2025) outline the following four-step approach for carrying out KCA, with the first three steps representing methodological procedures to enable later analysis (i.e., step 4):

Step 1: Compute keywords using a traditional keyword analysis

Step 2: Analyse each text in the corpus for the occurrence of these keywords and record in a categorical data matrix (see Table 4

Table 4:

Dummy data matrix.

Tweet	Keywords			Quali_sup						Quanti_sup
	KW1	KW2	KW …	Anti	Conspiracy theories	COVID-19	Pharma	Politics	Science and technology	Text length
1	A	P	A	FALSE	FALSE	FALSE	FALSE	TRUE	TRUE	29
2	P	P	A	FALSE	FALSE	TRUE	TRUE	FALSE	FALSE	48
3	A	A	A	TRUE	FALSE	FALSE	FALSE	FALSE	FALSE	32

)

Step 3: Subject the data matrix to MCA to reveal dimensions comprising the most common patterns of co-occurring keywords

Step 4: Interpret these dimensions of keyword co-occurrence, guided by the principles of linguistic co-occurrence (Biber 1988) and the indicative nature of keywords in discourse (Baker 2023)

As such, step 1 involved producing a keyword list for our target corpus by comparing tweet text (minus #hashtags and @mentions) against the same Twitter reference corpus used to generate keywords on the TRAC:COVID platform. Keywords were identified using log-likelihood as a test for significance and because infrequent and overly frequent features can adversely affect MCA (step 3), we only considered keywords that occurred within at least 1 % (≥359.56) and fewer than 95 % (≤34,158.2) of all 35,956 unique tweets in our corpus. This approach produced a list of 175 keywords.

In line with step 2, using this keyword list, we produced a data matrix that records whether the keywords identified in step 1 are present (P) or absent (A) in each tweet in our corpus and also appended some supplementary information to the matrix. We specified qualitative supplementary variables (“Quali_sup” in Table 4) indicating whether dis/misinformation hashtags occurred in these tweets and which category or categories of COVID-19 dis/misinformation (Table 2) these hashtags belonged to, as well as information about the length of each tweet as a quantitative supplementary variable (“Quanti_sup” in Table 4). Table 4 gives some dummy data to visualize the structure of this matrix.

The third and final methodological step was to subject this matrix to MCA using the FactoMineR package in R (Lê et al. 2008) to identify dimensions comprising the most common patterns of keyword co-occurrence across the corpus. The MCA assigned each category of keyword (present or absent) a score (as a percentage) for how much they contribute to a dimension. Categories of keywords that make stronger contributions to a dimension therefore account for a larger share of the total percentage of contributions to a dimension. Our analysis below focuses on only those categories of keywords that made the strongest contributions to a dimension (and therefore are most representative of the patterns of variation within a dimension; Le Roux and Rouanet 2010: 52) by considering those with a contribution score equal to or greater than the average expected contribution for a dimension. This was achieved by using the following equation, which divides the total possible contribution to a dimension (100 %) by the total number of categories of keywords in our feature set. Given that we have 175 keywords and two possible categories of keyword, the average contribution score is calculated as follows:

Furthermore, keywords are also assigned coordinates, which show how closely associated keywords are to each other based on their distribution in texts; “keywords with strong contributions and positive coordinates co-occur often together in many texts, while keywords with strong contributions and negative coordinates co-occur often together in a different set of texts” (Sha and Clarke 2025). For more detail on MCA for identifying dimensions, see Sha and Clarke (2025) and Clarke et al. (2021).

Concerning step 4, which focuses on the interpretation of dimensions to explain why keywords co-occur across many tweets, our approach draws broadly from theory and methods in corpus linguistics, discourse analysis, and corpus-assisted discourse studies. To wit, we concentrate on keywords so as to identify lexical items that are both distinctive to and generalizable across (Egbert and Biber 2019; McGlashan and Krendel 2023) the texts that make up our target corpus. In other words, keywords enable us to identify language use that is statistically salient within and across – and therefore typical of – the texts that make up our corpus. Furthermore, our consideration of keyword co-occurrence enables us to operationalize the conception of discourse as “a cluster of context-dependent semiotic practices that are situated within specific fields of social action” (Reisigl and Wodak 2016: 27) and to see how specific dimensions (and potentially discourses) are revealed through the co-occurrence of keywords that also contain reference to vaccines and dis/misinformation-related hashtags.

4 Findings

Due to limitations of space, this analysis concentrates on the first dimension of our KCA, which, using Benzécri’s (1992) modified rates, explains 23 % of the variance in the dataset. The 94 keywords contributing most strongly to Dimension 1 are presented in Table 5. Unlike traditional keyword analysis, wherein a specified number of keywords – or set of keywords – would be isolated from a larger keyword list for independent analysis, KCA is interested in how keywords co-occur as part of larger functional constructs and so keywords need to be interpreted in the context of the texts in which they occur to understand how they function within those constructs.

Dimension 1 distinguishes between texts on the negative side of the dimension, which include specific forms of COVID-19 dis/misinformation and conspiracy theories, and texts on the positive side of the dimension, which include more general concerns about freedom, control, and “normality” stemming from ongoing (and future) interventions to curtail the spread of COVID-19.

Texts associated with negative Dimension 1 make (and intend to signal/boost/spread) specific forms of COVID-19 disinformation, misinformation, and conspiracy theory (including reference to specific individuals or external links), as exemplified in the text most strongly associated with negative Dimension 1 (see Table 6). Many of the texts associated with negative Dimension 1 promote conspiracies about Bill Gates’s role in the COVID-19 pandemic. The conspiracies suggest that Bill Gates (along with other elite individuals) orchestrated the COVID-19 pandemic to simultaneously profit from the manufacture of vaccines, and harm and control the global population through adulteration of these vaccines to include “microchips” and DNA-altering substances. Such tweets often share supporting content, including URLs to videos, images, and other additional content, as a way to add evidence to their claims.

By contrast, texts associated with positive Dimension 1 draw from an overarching “discourse of liberty”, which “shelters” a range of “subordinate” discourses (Sunderland 2004: 69) wherein liberty is construed in relation to perceived norms of freedom and restriction in the specific context of the COVID-19 pandemic.

These discourses articulate a variety of perspectives on the tensions between negative liberty (i.e., the desire to be free from outside interference; freedom from imposition) and positive liberty (i.e., the ability to self-determine; freedom to act without imposition) evaluated (thus, ideologically construed) in terms of what is (un)desirably “normal”.

Some texts (as exemplified in Table 7) include a “discourse of delayed gratification” in which impositions on negative liberty (e.g., using COVID passports, taking vaccines, taking tests, #jabmeup) are accepted despite expressing immediate desires for positive liberty (“getting back to normal”, seeing and hugging family). The achievement of positive liberty (“normal”) is therefore construed as possible but only through temporary acceptance of some imposition on negative liberty.

More common in positive Dimension 1, however, are texts that bring into question any COVID-19 prevention and containment measures. For example, texts in Table 8 illustrate what we interpret to represent a “discourse of barriers to positive liberty” wherein the existence of COVID-19 is not questioned but measures that cause any imposition on freedom to act (e.g., lockdown, social distancing) are construed as ineffective. Any concession of negative liberty is viewed with suspicion. And so, perceptions of what might be ideally “normal” here are those that favour positive liberty.

More radical scepticism is seen throughout texts associated with positive Dimension 1 and include outright rejection of prevention and containment measures and what we interpret to be a more radical “discourse of libertarianism”. Examples range from texts that include some general reference to (and the rejection of) government/state control enabled by COVID-19 prevention and containment measures (Table 9) to texts that express fears of coercive government/state control (Table 10). Within such texts, COVID-19 measures are construed as direct acts of control rather than a byproduct of COVID-19 that enable governments/states to pursue more insidious agendas such as eugenics programmes and population control (example (9)), and government/state tyranny (examples (10)–(12)). Here, articulation of the “discourse of libertarianism” is achieved through direct references to “control” (examples (9)–(10)) and “permanent control over our liberties” (example (10)), “#tyranny” (example (11)), and intertextual reference to George Orwell’s Nineteen eighty-four – a dystopian novel about government totalitarianism – which is recontextualized to the COVID-19 context through the hashtag “#COVID1984” (example (12)).

Though brief, this analysis of a specific subset of COVID-19 tweets containing reference to both vaccines and dis/misinformation-related hashtags identifies an overarching “discourse of liberty” and several subordinate discourses of “delayed gratification”, “barriers to positive liberty”, and “libertarianism” through which “normality” is construed most typically as a life free from government imposition (here, a range of COVID-19 prevention and containment measures). Reading these findings together with a summary of how frequently different categories of dis/misinformation hashtags are used in texts associated with positive Dimension 1 (Figure 3), it appears that these tweets serve to function as forms of visible (and searchable) protest that are informed by a range of conspiracy theories; the hashtag categories “Anti” and “Conspiracy theories” account for 63.5 % (36.71 % and 26.79 % respectively) of the total number of hashtags in positive Dimension 1 texts.

Figure 3:

Tree map visualizing hierarchically (left to right) the frequency of occurrence of each of the COVID-19 dis/misinformation hashtag categories (described in Table 2) as they occur in texts associated with positive Dimension 1.

Overall, Dimension 1 indicates that the texts in the corpus most often vary according to those which present Bill Gates as the villain protagonist in the COVID-19 pandemic and those which call for freedom. This dimension thus represents a continuum of blame: on the one hand, blame is put squarely on Bill Gates as the root cause for violating “health freedoms”, whereas on the other hand, multiple conspiracy theories are co-present to argue for “health freedoms” by questioning the effect of restrictions and their legitimacy. In other words, where the target for blame on the positive side of the dimension can be (and is in our data) more nebulous, the target for blame on the negative side is clear and absolute.

5 Conclusions

This study concentrated on a very small subset of the total number of tweets that make up the TRAC:COVID corpus (35,956/84,138,394 = 0.04 %), which include both references to vaccination (*vax* and *vaccine*) and dis/misinformation hashtags. Any findings, therefore, are not generalizable beyond this restricted sample. However, this study found that tweets posted during the COVID-19 pandemic containing both dis/misinformation hashtags and reference to vaccines draw on “discourses of liberty” in order to protest against the perceived imposition of new norms of behaviour (e.g., mask-wearing), which are viewed with suspicion and as overreaching. In doing so, these tweets function to both reject any infringement on “health freedoms” as well as to identify (through drawing on a range of conspiracy theories) and apportion blame to some cause(s) of these infringements. These findings suggest that changes to perceived norms (especially those that may impede on subjective understandings of personal liberty) may be perceived as threats and, thus, these findings present a challenge for managing change, whether in this specific context or not.