Aggressive behaviour of anti-vaxxers and their toxic replies in English and Japanese

The study investigates how anti-vaccine advocates (anti-vaxxers) use Twitter replies—direct messages that can reach beyond follow-follower ties—to spread aggressive, toxic, and emotionally charged content during the COVID-19 pandemic. Prior work has documented misinformation, echo chambers, and polarization in vaccine discourse, largely in English, but reply behavior as a targeted mechanism remains underexplored and its cross-linguistic characteristics are unknown. Given that replies can reach anyone and may have significant impact, the authors ask: how active are anti-vaxxers’ replies, whom do they target, and how toxic/emotional are these replies across English and Japanese contexts? Understanding these dynamics can inform countermeasures against vaccine misinformation and harassment.

The paper situates its inquiry within research on COVID-19 misinformation and the anti-vaccination movement, including studies of narratives, emotions, behavioral patterns, echo chambers, temporal spread of fake news, and motivations for sharing misinformation. It highlights that vaccination discourse forms ideologically segregated communities (echo chambers), often aligned with political identities, and that conservative ideologies are associated with higher susceptibility to vaccine conspiracies. Most prior analyses focus on English content; little is known about directed reply behaviors or cross-language comparability. The authors also reference work on the persuasive power of emotional narratives and the tendency for influential social media actors to express negative sentiments, framing why toxic replies might be impactful.

Data collection: Using the Twitter Search API, the authors continuously collected COVID-19-related tweets (English and Japanese) from February to December 2020, querying generic COVID-19 terms (e.g., coronavirus, COVID19, SARS-CoV-2 and Japanese counterparts). They then filtered tweets containing vaccine-related terms (e.g., vaccine, vax, vaccination and Japanese equivalents). English: 8,579,728 tweets (80.2% retweets; 3.43% replies), 2,799,034 unique users. Japanese: 1,952,376 tweets (81.5% retweets; 2.64% replies), 576,894 unique users. User stance clustering: They constructed a retweet (RT) network over February–December 2020, connecting users with at least two RTs (including mutual RTs) to reinforce endorsement signals. They applied k-core decomposition (k=3) to retain users engaged in primary discussions and then used the Louvain method to detect communities. Five large clusters emerged in both languages and were labeled via most-retweeted accounts and keyword analysis: Pro-Vax, Anti-Vax, Left (anti-Trump/left-leaning), Right (pro-Trump/right-leaning), and Neutral (predominantly media and informational sources; vaccine makers, universities). Word clouds and representative accounts (in SI) validated labels. Toxicity measurement: Google’s Perspective API (English-only) scored tweet toxicity (0–1). Japanese tweets were translated into English via Google Translate prior to scoring. Emotion measurement: The authors counted words from lexicon-based resources. Positive/negative emotions via LIWC 2015. Valence and arousal via Warriner et al. (2013). They counted words per tweet above the dictionary median score for each metric. Japanese tweets were translated with Google Translate before lexicon matching. Analyses: They compared reply activity levels by cluster; quantified intra- vs inter-cluster reply patterns; identified reply targets by cluster; computed proportions of replies targeting large accounts (≥10,000 followers); compared toxicity between inner- and inter-cluster replies and across clusters; examined which targets received the highest toxicity; and analyzed the relationship between target follower counts and received toxicity. Emotional content (positive, negative, valence, arousal) was compared across clusters for inter-cluster replies using Mann–Whitney U-tests with Bonferroni corrections.

- RT network communities: Five clusters were identified in both English and Japanese datasets: Pro-Vax, Anti-Vax, Left, Right, and Neutral. Neutral clusters were dominated by news media and vaccine makers; Anti-Vax clusters focused on conspiracies and government criticism; Pro-Vax focused on efficacy and evidence. - Reply activity rates: Anti-Vax users exhibited the highest reply activity in both languages. • English reply rates (RP/TW): Pro-Vax 3.01%, Right 0.70%, Anti-Vax 5.06%, Left 0.94%, Neutral 1.25%. • Japanese reply rates (RP/TW): Pro-Vax 2.24%, Right 1.25%, Anti-Vax 2.84%, Left 2.80%, Neutral 0.97%. - Inter- vs inner-cluster replies: Although most replies were inner-cluster, Anti-Vax groups sent the largest number of inter-cluster replies in both languages (Right was comparable in Japanese), indicating an active outreach to differently aligned users. - Targets of Anti-Vax replies: Anti-Vax users primarily targeted Neutral accounts (notably media and politicians), more so than other groups. They directed fewer replies to Pro-Vax than vice versa, revealing asymmetry in engagement. - Influential targets: Reply receivers tended to have more followers than senders; a high proportion of Anti-Vax inter-replies were directed at large accounts (≥10,000 followers), especially within Neutral clusters. - Toxicity: Inter-cluster replies were significantly more toxic than inner-cluster replies. Anti-Vax inter-cluster replies were significantly more toxic than those from Pro-Vax and Neutral groups in both languages (Mann–Whitney U-tests with Bonferroni correction; p < 0.001). In English, Anti-Vax toxicity was notably directed toward the Right cluster (e.g., aggressive complaints about vaccine policy under the Trump administration). In Japanese, toxic replies from Anti-Vax to Right centered on administration criticism and were not significantly elevated versus other targets. - Toxicity vs popularity: There was a positive correlation between the maximum toxicity a user received from Anti-Vax accounts and the user’s follower count in both languages; users with many followers received highly toxic replies. In English, some small accounts (few followers), including Pro-Vax accounts, also received highly toxic replies—a pattern not evident in Japanese. - Emotion analysis: Anti-Vax inter-cluster replies contained more negative and fewer valence words in both languages; reduced use of positive words was clear in English. These emotional profiles align with higher toxicity and negativity in Anti-Vax replies. - Dataset scale: English COVID-19 vaccine-related tweets totaled 8,579,728 (80.2% RTs; 3.43% replies); Japanese totaled 1,952,376 (81.5% RTs; 2.64% replies).

The study addresses how anti-vaxxers leverage Twitter replies as targeted mechanisms to reach users beyond echo chambers. Across English and Japanese, Anti-Vax clusters were the most active in sending inter-cluster replies, preferentially targeting Neutral (largely media) accounts with high toxicity and negative emotionality. This behavior likely amplifies visibility by leveraging prominent accounts and may increase exposure of broader audiences to anti-vaccine narratives. Elevated toxicity in inter-cluster replies suggests attempts to intimidate, discredit, or emotionally influence targets rather than engage in constructive debate. The asymmetrical engagement—Anti-Vax users directing relatively fewer replies to Pro-Vax than vice versa—indicates strategic neglect of adversarial communities and emphasis on neutral or influential targets for maximum reach. The positive association between target popularity and received toxicity underscores risks for high-profile accounts; the English-specific pattern of high toxicity directed at smaller Pro-Vax accounts suggests additional vulnerability for less prominent advocates in English contexts. These findings inform cross-lingual countermeasures, such as prioritizing moderation tools for reply toxicity, protective settings for targeted users, and proactive engagement strategies aimed at Neutral audiences most likely to be exposed to Anti-Vax replies.

This comparative, cross-linguistic analysis reveals consistent patterns in anti-vaxxers’ aggressive reply behaviors: high inter-cluster activity, focused targeting of Neutral and influential accounts, and elevated toxicity and negative emotionality. The English data additionally show that smaller Pro-Vax accounts can receive highly toxic replies, a difference from the Japanese case. These insights suggest both language-independent and language-dependent countermeasures for platforms, media, and public-health communicators. Practical implications include strengthening automated detection of toxic replies, offering enhanced protections for high-visibility and newly active Pro-Vax accounts, and prioritizing inoculation strategies for Neutral audiences frequently targeted by Anti-Vax replies. Future research could validate findings across more languages, extend beyond 2020 to capture evolving platform features and discourse, assess causality between toxic exposure and vaccine attitudes, and evaluate the effectiveness of intervention strategies in reply dynamics.

- Toxicity measurement relies on Google’s Perspective API, available only for English; Japanese tweets were machine-translated prior to scoring, which may introduce translation bias. - Emotion analysis similarly depends on English lexicons and translation of Japanese tweets, potentially affecting accuracy. - Data collection focused on COVID-19-related tweets and then filtered by vaccine terms, potentially excluding relevant vaccine content outside the COVID-19 context or using atypical terminology; ‘pandemic’ was not included for consistency at the search’s start. - Analyses are limited to Twitter and to February–December 2020; behaviors and platform features may have changed subsequently. - RT network construction applied thresholds (RT ≥ 2 and k-core = 3), focusing on active participants and possibly underrepresenting peripheral users. - Replies constitute a small fraction of tweets; while impactful, their low base rate may affect generalizability to all discourse forms.