Using scientific and colloquial terms associated with HPAI and specifically H5N1, such as bird flu, avian influenza, and avian flu, we collected approximately 350,000 Facebook and Instagram posts referencing H5N1 and associated terms using both CrowdTangle and a webometric tool called Zeeschuimer for more Instagram data. The data included both textual and multimedia content. We focused on these platforms because they are among the 3 most accessed social media platforms globally. Facebook is the market leader with an estimated active user population of over 3 billion [46]. Although YouTube is ahead of Instagram as the second most popular platform, its content being primarily video-based, with minimal text, excluded it from this study. We also selected Facebook and Instagram to expand the knowledge base on health misinformation beyond the platform X, especially regarding the affordances of social media platforms.

The CrowdTangle data were collected from June 5, 2014, to June 4, 2024, the posts are not limited to geographical demarcations. We collected the additional Instagram data with Zeeschuimer in July 2024. The total Instagram posts were (n=69,551). For both the Facebook and Instagram datasets, we employed a Python script to identify misinformation-related posts using terms such as “plandemic,” “medicalmurder,” and “LuciferShot” [47,48]. We compiled this list of terms based on research on health misinformation from academic literature, [49,50], online exploratory observations, and notes from the research team. Keywords were specifically focused on pandemic and vaccine conspiracy–related terms, demonstrating general connections for misinformation narratives between COVID-19 and vaccines. The study period from June 5, 2014, to June 5, 2024, was selected to ensure a proper review of discussions on the virus and its continued activity and outbreaks. A specific instance occurred in March 2024 when H5N1 infections were reported in dairy cattle in the United States for the first time, highlighting its continued relevance for public discussions [51].

Our final dataset contains 544 posts after removing duplicates. Though the volume of posts flagged as misinformation and misleading appears to be limited, we note several limitations. First, and in searching for specific keywords, it is difficult to confirm the identification of all the available misleading or misinformation-filled posts in the dataset, but the above approach has been effective in identifying a relevant subset of data that we can further explore. In addition, since the two data collection tools are not comprehensive in getting data from Meta, we cannot make any claims about the volume of misinformation on Facebook and Instagram. Moreover, as the initial number of gathered posts (350,000) is far larger than the remaining posts we analyzed, the difference in numbers is due to our inclusion criteria that excluded any duplicates, irrelevant use of keywords, and posts beyond our scope. This reduction also represents the inclusion criteria to ensure relevance.

Employing a general inductive approach [52] as part of systematic content analysis approach, 3 coders carried out a close reading of a 10% sample of the dataset to outline the emerging and dominant arguments. The analysis was multimodal, which involves the text and any associated multimedia content. Following a process of several rounds of phased discussions, the researchers organized the broad themes into 2 categories (targets and topics). These categories were used to link the analysis to the research objectives of understanding the prevailing misinformative concepts on HPAI to inform effective public health agenda. The target is the organization or individual discussed in the post. The topic concerns the main subject or issue in the post, ensuring that subcategories were mutually exclusive and that the themes of the various subcategories are unique to each category, while the central messages do not overlap. Two researchers then used this initial codebook on a sample of the first 100 posts to test the codebook and test intercoder reliability. After the first coding round, the researchers adjusted the target section of the codebook, adding corporations as a category to account for multinational and other commercial entities, along with a category for irrelevant posts. The codebook has nine categories under target: (1) government had the highest occurring target references (n=149), (2) public health authorities had 108 references, (3) international organizations (n=74), (4) elites (n=64), (5) pharmaceutical industries (n=52), (6) mainstream media (n=43), and (7) corporations (n=23). The final categories for targets were (8) other/unclear (n=113) and (9) irrelevant (n=53). We distinguished between these 2 categories because while some posts were focused on H5N1 but did not address a target (other), others did have targets but had topics that were unrelated to avian influenza (irrelevant).

Researchers also adjusted the topics section following the first round of coding to address vagueness and overlapping categories. After the second round of intercoder reliability, further adjustments were made, including renaming a category to religious allusions and prophecies to include a broader range of religious content. Moreover, we removed an alternative medicine category as there were only 2 posts, and they were better classified under other. The adjusted seven categories are (1) pandemic engineering, which accounted for 362 topics, (2) societal destruction (n=121, 22%), (3) anti-vaccination (n=84, 15%), (4) elections/democracy (n=66, 12%), (5) unjustified food insecurity (n=59, 11%), (6) religious allusions and prophecies (n=26, 5%), and (7) other/irrelevant (n=77, 14%). The final category contained posts that were not focused on H5N1 but on other pathogens and diseases such as SARS-CoV-2/COVID-19. Others did not have a clear message and sometimes also used contradictory hashtags [49]. There were also 2 posts advancing misinformative pseudoscientific remedies to H5N1 infections by advertising plant-based vaccines. After 3 attempts involving the modifications described above, intercoder reliability was statistically acceptable at 0.887 for targets and 0.899 for topics using Krippendorff Alpha [53]. Though limited contextual insight was identified due to lack of context, we also examined the nature of the social media posters by reviewing their profiles and previous posts. Since the evidence we collected on users’ profiles was only anecdotal, we did not report these findings in this study. Finally, as we analyzed publicly available social media data, all posts used have been deidentified and paraphrased to prevent re-identification and to protect user privacy.

This project received an exemption from the requirements of Research Ethics Review at Simon Fraser University (30002857). All analysis is based on publicly available and deidentified data. All publicly available social media content used in this study was deidentified to prevent reidentification of posts or individuals.

The findings from this analysis of misinformation on 2 major social media platforms raise implications for public health policy and practice on the evolving H5N1 situation. Some evidence suggests that efforts to counter misinformation, such as debunking or fact-checking, may unintentionally amplify misinformation by elevating its perceived significance [32]. As the sample of posts analyzed represents a small portion of the initial dataset, our study does not aim to show prevalence on various platforms but to describe misinformation narratives and their types. The qualitative analysis of these narratives can demonstrate the construction and development of such narratives on this topic.

Importantly, findings suggest the need for the public health community to carefully consider such misinformation. Evidence shows that misinformation on social media has a differential diffusion pattern compared to traditional media [23]. Moreover, studies show that misinformation garners more engagement than facts. As Vosoughi et al [28] found, “falsehood diffused significantly farther, faster, deeper, and more broadly than the truth in all categories of information.” The top 1% of the most popular false news was disseminated among 1000 to 100,000 people. Verifiable information, however, barely reached more than 1000 people. Misinformation had a 70% higher chance of being retweeted compared to facts. By contrast, the truth took about 6 times as long as false information to reach 1500 people. This more rapid and far-reaching diffusion pattern of misinformation is also reported by studies of medical large language models, which have been found to be vulnerable to “data poisoning” attacks [55]. Large language models, trained on web-scraped data, are increasingly being explored to support health care [56]. However, as Alber et al [47] wrote, “replacement of just 0.001% of training tokens with medical misinformation results in harmful models more likely to propagate medical errors”. Thus, a deeper understanding is needed over time of their diffusion pattern and potential role in the spread of misinformation.

Our findings provide a baseline from which ongoing monitoring for trends in patterns of spread can be undertaken and lay the foundation for future work exploring misinformation on the HPAI virus in online spaces. For example, findings in this study about the use of scientific terminology in anti-vaccination posts showed similar points to the investigation by Faasse et al [57] of the dynamics of language use in pro- and anti-vaccination commentary on Facebook. They found that messages against vaccination used linguistic markers that resembled scientific language. Under the elections/democracy topic, posts analyzed reflected patterns noted in other studies regarding the belief in multiple conspiracy theories at the same time [36]. For instance, respondents in Miller’s study asserted that COVID-19 was a bioweapon manufactured by China, and the United States accidentally released it into society. In the study by Hellsten et al [58] on X communication networks during avian influenza outbreaks in the Netherlands from 2015 to 2017, they found frequent instances of concurring themes such as the rejection of health regulations #kippendans (chicken dance), #vuurwerk (firework), and #jacht (hunting). Thus, a principal recommendation is that public health systems need to use early warning data of social and behavioral factors, including social media monitoring of misinformation, to inform the development of public health messaging.

Second, our systematic analysis of the content of the posts deemed misinformation suggests that claims about H5N1 are an extension of dominant narratives, by topics and targets, put forth during the COVID-19 pandemic and other issue areas such as climate change. This could be seen, for example, in patterns of blaming a foreign entity or minority communities, such as blaming the Chinese government for H5N1, for a major disease outbreak, which has occurred regularly in history. Examples found included the massacre of Jews during the Black Death of the 15th century and during The Great Plague and the Spanish Flu of 1918‐1919; the lesbian, gay, bisexual, transgender/transsexual, queer, and other minority sexual orientations and gender identities community during the global AIDS pandemic; and Asian people during the COVID-19 pandemic [59].

An overarching theme as well was mistrust of the motivations for public health messaging and interventions. Our findings suggest that views sustaining mistrust persist, framed around a broader narrative of control by powerful elites to achieve a new world order favorable to their interests [60]. H5N1 has become subsumed within this world view, with public health concerns about the virus attributed to self-interested “pandemic engineering.” [61] Misinformation about the food system also demonstrated patterns of such mistrust in authorities following measures to control H5N1 infections such as bird depopulation. Bird depopulation can appear drastic because it involves both healthy and infected birds within a given perimeter. It has, however, shown to be the most effective method for stemming infection spread [8,62,63]. Misinformation against this measure typically has ominous labels such as murder and mass slaughtering. Additionally, while health authorities and governments have announced that H5N1 does not currently present food safety risks if animal products are appropriately processed or cooked [55,64], some misinformative posts called the safety of dairy products into question.

Another pattern of mistrust found was under religious allusions and prophecies topics, as users connected prominent public health communications, such as one made by the WHO Director-General [65], with religious narratives seen as the realization of prophetic predictions. This topic, in general, offers insights among certain religious groups that may resist mainstream information. Christian beliefs and those of other religious and insular communities are primarily grounded in faith rather than empirical evidence [66]. Such groups, particularly the fundamentalist ones, reject mainstream and external information, reinforcing their worldviews and possibly a greater risk of accepting misinformation [67-69]. Therefore, the credibility of such religious declarations relies on people’s acceptance of the parallels with Bible verses, which could reduce the need for expositions based on objective reality. Douglas further contends that the growing relationship between misinformation and religion is due in part to religious communities being principal targets [70]. This relates to the historical foundations of misinformation that was facilitated by Christian fundamentalists seeking alternative news sources as they rejected mainstream media’s inclination toward what they deem as biased expert knowledge, which usually deviates from religious dogma [6].

Given this broader framing, rebuilding and protecting public trust in the state, market, and civil society may require efforts beyond the public health community. Instead, strategies to sustain public trust in health practitioners and scientists, which remains high in most countries [71], as sources of quality information can be supported. For example, “localizing” public health efforts by working in meaningful partnership with community leaders can “humanize” and thus increase trust in public health messaging [72].

Third, our findings demonstrate the importance of protecting information quality as the H5N1 situation evolves while balancing the need to not draw undue attention to misinformation. Unlike COVID-19, the public health community has been managing the risks from H5N1 since the late 1990s, so much is already known about the virus [73]. Since 2022, the spread to a growing number of mammal species including dairy farms has increased media attention and public awareness [2]. Evidence suggests that there is thus an opportunity for public health messaging to facilitate public understanding of this issue while it is still emerging. For example, Vos and Buckner [74] found that sensemaking plays a key role in the success of public health interventions during health crises. Sensemaking involves the way the public interprets and frames health information they receive about emerging health events. The misinformation identified in this paper suggests H5N1 is already interpreted using existing conspiracy-related worldviews (eg, New World Order). To counter misinformed sensemaking, there is a need to avoid information vacuums. Moreover, public health messaging must not only be factually accurate but in a form that facilitates sensemaking. Accordingly, there is a need to provide appropriate information for the public to understand, and not just be informed, about H5N1. The insights into online health misinformation on HPAI that this study provides are instrumental for supplementing surveillance structures, guiding communication mechanisms with an understanding of the misgivings by some, which could be fueling hesitation toward and rejection of interventions and ultimately bolstering pandemic readiness measures.

Importantly, the ability to monitor the content and volume of misinformation on social media, and proactively message about the evolving H5N1 situation, is likely to be more important yet more challenging amid changing practices by social media platforms to limit access to data and to protect information quality. One of the strengths of this study is the novel expansion of online health misinformation analysis beyond the textual content of X platform. Our multimodal analysis of textual and multimedia data from Instagram and Facebook broadens the scope of knowledge of online public narratives, particularly about infectious diseases. Despite Meta’s claims about the strict moderation of posts,75 the majority of posts identified as misinformation were not flagged by both Facebook and Instagram for fact-checking. There were only 5 posts in our dataset that were flagged as inaccurate (less than 1%). The generic notification stated they were not completely factual, and they contained information similar to other posts that third-party fact checkers had reviewed. In 2018, CEO Mark Zuckerberg stated that Meta’s research showed, “that no matter where we draw the lines for what is allowed, as a piece of content gets close to that line (demarcating prohibited content which contains misinformation and harmful material), people will engage with it more on average—even when they tell us afterwards they don’t like the content” [76]. The announcement in January 2025 that Meta would end third-party fact-checking, in favor of allowing people to “express themselves freely,” could have implications for protecting information quality [77]. Given the ubiquity of social media and its centrality in sustaining social connection, especially during times of uncertainty such as public health emergencies, the public health community can benefit from understanding and navigating this evolving landscape as a core component of future pandemic preparedness and response.

A key limitation was the inability to include an analysis of the volume of misinformation on Meta platforms due to the lack of data collection tools. Moreover, we were unable to analyze the ideological leanings of users to further contextualize the narratives. We conducted a brief review of the profiles of users in our dataset to explore their potential ideological or political affiliations. Though some users’ profiles suggested specific political ideologies, this analysis was not included in the study, as confirming such affiliations would require an ethnographic approach, which falls outside the scope of this paper.

Additionally, our paper was limited to English-Language and region-specific search terms that mostly covered North America, along with a focus on some US public figures such as Anthony Fauci. This could have excluded similar discussions from other regions and countries such as the United Kingdom and Germany, where different terms were used. Future research could address this by including additional relevant keywords from different regions such as “flockdown” and “raw milk.” Another limitation was the prioritization of pandemic and vaccine-related terms in our dataset, which meant that other misinformation narratives about transmission and treatment, for example, may not have been captured. This limitation should be considered when examining our findings, as the topic categories identified were a result of our keyword search focus. Thus, findings reflect a specific focus of H5N1 misinformation.

Misinformation on HPAI on social media has implications for public health as recurrent avian influenza outbreaks in animal populations have contributed to fears about the pandemic potential of the disease for humans. Notably, H5N1 in human populations is associated with several clinical symptoms; thus, decisive and swift actions are needed to stem any further infections. This finding underscores the importance of reviewing contextual factors like online misinformation, which could interfere with health interventions.

Our online investigation revealed that the targets of H5N1 misinformation narratives were largely people or entities who wield varying levels of political or other authority such as government and public health officials. These individuals received attention for their varying roles in health-related decision-making and in facilitating public health interventions.

We also found that the most prevalent misinformation themes were accusations of pandemic engineering, societal destruction, anti-vaccination sentiments, impact on elections or democracy, causing food insecurity, and religious allusions and prophecies. These assertions took various forms, including accusations that the virus itself did not occur naturally but claimed that the virus was artificially created as a bioweapon. The messages that contained anti-vaccination sentiments concluded that health interventions for avian influenza would include vaccines that would insert digital biometric identification into people, which could be used to either instigate population control or enforce a digital economy.