Abstract
Incels (involuntary celibates) base their identity on the inability to form romantic relationships. We conceptualize the ideology promoted by incels as misogynist extremism and explore the impact of the first COVID-19 lockdown on the radicalization of this online community. Based on computational measures, we conducted a multi-perspective exploration, comparing the prevalence of and participation in threads dealing with extremism, ideology and mental health on the incels.is forum between pre-lockdown, lockdown and post-lockdown periods. We found evidence of long-term and temporary radicalization. Moreover, we found that, specifically, older forum members increasingly post in extremist-themed threads triggered by the lockdowns. Crucially, we show that activity on mental health–themed threads temporarily decreased during the lockdown. These findings indicate that real-world social isolation reduces mental health complaints among incels but, at the same time, exacerbates misogynist extremism among active community members.
Introduction
It has been suggested that the COVID-19 pandemic, characterized by repeated and prolonged lockdowns, exacerbated polarization and radicalization on the Internet (Davies et al., 2021). Existing studies typically measure the average number of posts in extremist forums and infer radicalization from the increase in activity during lockdowns (Davies et al., 2021; Vu, 2020). More recent work has complemented these findings, questioning whether an increase in posts on the forum level should be interpreted as an increase in radicalization (Baele et al., 2023). Similarly, we argue that increased activity is a predictable outcome of limited leisure options. We propose an operational definition of online radicalization that focuses on the substance of online activity rather than its mere presence. Our study provides novel information on radicalization signals during global crisis events such as the COVID-19 pandemic through a theoretically grounded operationalization of extremism and affiliated themes (e.g. toxicity, ideology, mental health), a methodological approach that goes beyond mere dictionary approaches and a longitudinal study of these signals.
Our research centres on incels (involuntary celibates), a radical online group that defines its identity around the inability to form romantic relationships. Incels are considered the most radical subgroup among misogynist online groups, exhibiting signs of violent extremist ideology (Baele et al., 2019; Tomkinson et al., 2020). Furthermore, misogynistic ideology, specifically, has been shown to lead to violence online and offline and contribute to radicalization processes (e.g. Cottee, 2021). Therefore, our research focuses on the potential for radicalization among incels, specifically on their most prominent forum, incels.is.
Our analysis was carried out on all posts published in the largest incel forum (incels.is) between 1 January 2020 and 31 July 2020 (N = 1,206,287). We compared the prevalence of and participation in extremist and non-extremist threads before, after and during the first lockdown. To analyze user participation, we only considered the sample of users (n = 421) who had been active in all three time periods. This allows us not only to evaluate the general development of the community, as in previous research at the forum level (Baele et al., 2023), but also to discuss the effects at the individual user level.
We conceptualized the ideology of the incels.is forum as misogynist extremism and grounded our understanding of extremism on Berger (2018), who describes extremism as a measurable attribute of an ideology. To measure misogynist extremism and related dimensions, we used a topic model (BERTopic), and grouped the resulting topics into theoretically and inductively derived categories representing an exclusively extremist topic (extremism) and three major topics within the incel cosmos (black pill, manosphere and mental health). In addition, we employed two established computational measures, one for toxicity and one for misogyny. This multi measurement approach allows us to avoid the pitfalls of pure dictionary approaches which, compared to language models such as BERTopic, are bound to previous vocabulary knowledge for the dictionary-making and cannot take the context in which a term is used into account. By comparing the prevalence of extremist threads and users’ participation in those across different statistically derived time periods (pre-lockdown, lockdown, post-lockdown) with multivariate regression models, we investigate whether the lockdown amplified radicalization, both during the lockdown and in its aftermath.
Incels, radicalization and misogynist extremism
Radicalization is often defined as ‘the process whereby [groups or] individuals come to adopt an extremist mindset’ (Marwick et al., 2022: 13–14). Some scholars (e.g. Alimi et al., 2015; Malthaner, 2017) restrict radicalization to violent radicalization. For this research, we consider cognitive radicalization (Whittaker, 2022), and adopt the definition of radicalization as a process of ‘acceptance of far reaching changes in society, which may or may not constitute a danger to democracy and may or may not involve the threat of or use of violence to attain the stated goals’ (Borum, 2011: 12). This is in line with a broader conception of radicalization as the process of adopting extremist beliefs, such as far-right extremism or jihadism (Marwick et al., 2022; Neumann, 2012).
Closely related to radicalization is the concept of extremism, a highly community-based and relational attribute (Berger, 2018). A broad definition, such as ‘a belief that an in-group’s success or survival can never be separated from the need for hostile action against an outgroup’ (Berger, 2018: 79), is helpful for extremism research, as it is measurable and encompasses all prominent cases of extremism (e.g. Nazism, Islamist terrorism) under one theoretical umbrella term. Adopting a broader conceptualization of radicalization that focuses on adopting extremist beliefs (Marwick et al., 2022; Neumann, 2012) rather than solely on the willingness to commit violence highlights the importance of community and engagement with extremist content – encompassing radicalization processes beyond terrorism but also culture wars or societal polarization (Berger, 2018; Sunstein, 2018).
Incels have long been dismissed as a minor, anomalous group (Tomkinson et al., 2020). Only in recent years have they been recognized as an important contemporary phenomenon among extremist communities (e.g. Baele et al., 2019, 2023). Incels are part of the manosphere, often cited as an example of culture wars in news articles (Bosman et al., 2019) but, as shown by Tomkinson et al. (2020), these proclaimed culture wars are closely linked to the rise of political violence. The manosphere is a group of loosely coordinated misogynist online communities, its roots dating back to the 1970s (Messner, 2016). The four main groups are pick-up artists (PUAs), Men Going Their Own Way (MGTOW), men’s rights activists (MRAs) and involuntary celibates (incels), all examples of gender-based extremism (Berger, 2018; Zimmerman et al., 2018). Most subgroups of the manosphere are connected through a common base ideology, the so-called red pill in reference to the movies The Matrix (Bratich and Banet-Weiser, 2019).
The red pill ideology describes the beliefs of all manosphere groups except incels. Within this ideology, it is a fact of nature that women are only attracted to the most attractive and wealthiest men. Furthermore, the red pill ideology proclaims a reality in which women rule the world without bearing any responsibility, and men, their victims, are not allowed to complain or protest. Although every group has a different interpretation of their grievances, the binding idea is that boys and men are victimized and that feminists are to blame for this (Champion and Frank, 2021).
The black pill ideology, the ideological basis of incels, sets them apart from the rest of the manosphere. They follow the same initial assumptions as the red pill ideology. However, the two ideologies come to different conclusions based on these a priori beliefs. In the tradition of neoliberalism, the red pill postulates that self-optimization can lead to success despite unfortunate predispositions. In contrast, the black pill ideology concludes that no effort can overcome the evolutionarily determined fate incels are confronted with (Bratich and Banet-Weiser, 2019; Hoffman et al., 2020). They blame the feminist movement and society at large for their grievances. The only solution is seen in a structural change in society and, given the poor probability of such a change, violence against yourself and others (Ging, 2019; Zimmerman et al., 2018). Such violence within the incel ideology has been identified on three levels: personal, interpersonal and societal (Aiolfi et al., 2024). Together with a sense of despair, this is why researchers have conceptualized the black pill radicalization pipeline as the radicalization process rooted in the incel ideology (Green et al., 2023).
Following such conceptualizations, the incel radicalization process is characterized by violence. Yet incels are not necessarily engaging in this; instead, ‘the prevailing condition among incels is one of apathy, not violent rebellion’ (Cottee, 2021: 108) The conceptualization of the black pill ideology is after all an idealization of a belief system: ‘incels are not ‘deviant’ others, they are not separate from other men’ (Sugiura, 2021: 94). Consequently, previous applied incel research defines extremism on online incel plat forms accordingly, in a broader sense with violent extremism defined as ‘the endorsement of violence against dehumanized outgroups’ rather than as the enactment of violence (Baele et al., 2023: 383). Furthermore, the approval of violence might even serve as a coping mechanism, preventing violence rather than contributing to it (Sugiura, 2021).
Drawing on Berger’s (2018) definition of extremism, which emphasizes a worldview that divides people into an ingroup and an outgroup and the ‘belief that an in-group’s success or survival cannot be separated from a hostile action against an out-group’, with gender as a possible basis for the ingroup/outgroup divide, we operationalize the ideology of incels as misogynist extremism. Misogyny, defined as ‘an unreasonable fear or hatred of women that takes on some palpable form in any given society’ and ‘sexual prejudice that is symbolically exchanged (shared) among men’ (Gilmore, 2010: 9)), is a defining factor of incel ideology (Cottee, 2021; Halpin et al., 2023). Following this, we take the direct anti-woman agenda that is part of the incel ideology as the basis to conceptualize the extremist part of the incel ideology as misogynist extremism. We measure the concept of misogynist extremism, which we operationalize along four dimensions: (1) toxicity, (2) misogyny, (3) general extremism and (4) the incel-specific black pill ideology. Misogyny, extremism and the black pill ideology are an integral part derived from our theoretical reflections above. Toxicity has been suggested to correlate strongly with hate speech and offensive language (Lees et al., 2022), which are good predictors of extremist language (Kleinberg et al., 2021). Therefore, we include toxicity as an additional proxy for misogynistic extremism.
We refrain from abstracting the misogynist extremist behaviour of incels to gender-based extremism as a whole, acknowledging the distinct nature and mechanisms of these ideologies. Although there are female incels, they have not been shown to express signs of extremist belief or to politicize their frustrations in the same way as their male counterparts (Cottee, 2021). However, there is a growing debate about reactionary feminism, which exhibits parallels to the characteristics of male incels and the broader manosphere, such as bio-essentialism, fatalism and hostility towards liberationist feminism, despite its purported feminist framing (Kay, 2024). Although these parallels exist, the fundamental difference lies in the nature of their extremism: incel ideology approves of direct, violent misogyny, whereas reactionary feminism operates through a subtler, though equally damaging, reinforcement of traditional power structures (Kay, 2024).
COVID-19 and online radicalization
Far-right and incel online communities, experienced a surge in posts per day at the beginning of the global pandemic (Davies et al., 2021). The COVID-19 pandemic has provided opportunities for extremists and terrorists to mobilize and advance their agendas (e.g. Gunaratna and Petho-Kiss, 2023). Kim (2023) argues that pandemic-related anti-Asian rhetoric has served as a conduit for right-wing extremism, leading to an increase in hate crimes against Asians and individuals of Asian descent. In Western countries such as the United Kingdom, the pandemic has facilitated the spread of toxic far-right narratives, creating fertile ground for extremist mobilization (Collins, 2023). In several far-right online communities, Schulze et al. (2022) found significant increases in all radicalization-related signals.
The global COVID-19 pandemic led to social instability and uncertainty and has been described as being exploited by extremists to spread disinformation and increase distrust among the public Ivanov (2022). Evidence of increased activity has been reported in far-right and misogynistic forums during the first COVID-19 lockdown (Davies et al., 2021; Schulze et al., 2022). The reasons for this may lie in the selective (ideological) reinterpretation of science and pseudoscience practised among incels (Cannito and Ferrero Camoletto, 2022; Daly and Reed, 2022; Johansson Wilen et al., 2024). Furthermore, political communication during the COVID-19 era had a particular masculinist property: ‘a particularly White, masculinist genre of political “truth-telling” has assisted the confusion, false belief and casualties of COVID-19’ (Harsin, 2020: 1061). This could have been a key aspect of the tendency among especially men to ignore the guidelines of health officials and the increased activity in far-right and misogynistic forums, which are connected by an ideal of patriarchal masculinity (Coffe, 2018; Kovats, 2017; Mellstrom, 2020). Furthermore, it is suggested that people who show cognitive traits such as self-uncertainty, low self-esteem and lack of analytical thinking are more susceptible to believing in misinformation and radicalization (Roberts-Ingleson and McCann, 2023), traits that are present especially among incels (Colliver et al., 2022; Moskalenko, Gonzalez, et al., 2022; Moskalenko, Kates, et al., 2022). Finally, loneliness, often due to social rejection during adolescence, is another key aspect that has been reported to lead to inceldom (Colliver et al., 2022). The pandemic increased the potential for loneliness and, therefore, perhaps the chance for new and former forum users to find their way (back) to the incel community.
Initial empirical findings support the above theorizing: Davies et al. (2021) concludes that this particular global crisis was well suited to promote far-right and misogynistic narratives. For incels, this finding is supported by Vu (2020) and Baele et al. (2023), who also found indications of increased use of hostile and violent language. Davies et al. (2021) observed an increase in activity in far-right and incel forums at the beginning of the COVID-19 pandemic. They provided anecdotal evidence that incels were delighted that attractive people were unable to have sex and had claimed that COVID-19 is karma for people having casual sex. Furthermore, incels supposedly viewed masks as an ‘equalizing force’ concerning physical appearance. Such research defines radicalization merely as the number of activities on a particular extremist site or provides only anecdotal evidence. However, the literature lacks a scalable approach to analyse the presence of misogynist extremism online in depth. The closest to this approach was a study by Baele et al. (2023); however, they based their approach on the more general definition of violent extremism and do not distinguish between ideological and extremist slang. Furthermore, the respective research could, at that time, only investigate the impact of lockdowns compared to pre lockdown times (Davies et al., 2021; Vu, 2020) and only looked at the aggregated effects on the forum level Baele et al. (2023). This study introduces the temporal dimension in a more robust way by comparing the pre-lockdown, lockdown and post-lockdown phases as well as looking specifically at user-level effects.
Communication on the incel.is forum is not just extremist. Conversations about the manosphere (the overarching online space incels are part of) and mental health are generally the major, and not necessarily clear-cut extremist, topics discussed on platforms such as incels.is (Baele et al., 2023; Moskalenko, Gonzalez, et al., 2022; Sparks et al., 2022). Mental health and its connection to incel ideology and radicalization have in particular been explored before (Moskalenko, Gonzalez, et al., 2022; Moskalenko, Kates, et al., 2022). Consequently, mental health struggles are commonly found among incels but not necessarily specifically connected to the potential for radicalization among those suffering (Marwick et al., 2022; Moskalenko, Gonzalez, et al., 2022).
Our analysis aims to provide a comprehensive picture of incel radicalization, not only through our robust conceptualization of misogynist extremism, but also through the incorporation of further important thematic dimensions. Therefore, while considering dimensions that are part of our operationalization of misogynist extremism (toxicity, misogyny, extremism, black pill ideology), we also include the topics of mental health and the manosphere in our analysis. The latter two are not part of our conceptualization of misogynist extremism but are included as comparative thematic dimensions. We extend the existing literature by (a) applying a robust theoretically derived definition of incel-specific extremism; (b) providing a more holistic view of radicalization processes by incorporating additional thematic scopes; (c) comparatively exploring the forum and user-level effects on a statistically sound definition of pre-, during and post-lockdown phases; and (d) exhibiting how topic models can be employed to overcome issues of static language use measures, such as dictionary-based approaches.
For our research, misogynist extremism encapsulates the unity of the two topics, extremism and black pill ideology, as well as toxicity and misogyny scores. To thoroughly investigate the dynamics of radicalization across pre-lockdown, lockdown and post-lockdown periods, we are first interested in the general change in content on the forum (thread-level), we ask:
RQ1. How did the first global COVID-19 lockdown impact the prevalence of misogynist extremism, mental health and manosphere content topics on incels.is?
Second, we are interested in the same users’ trajectories (user-level) throughout the pre-lockdown, lockdown and post-lockdown phases:
RQ2. How did the first global COVID-19 lockdown impact the participation of incels.is users in extremist misogynist threads, threads on mental health and threads on the manosphere?
Finally, we control for user seniority on the forum as a confounding variable:
RQ3. How does the seniority of users impact the participation of incels.is users in misogynist extremist threads, threads on mental health and threads on the manosphere?
This set of research questions allows for an understanding of the radicalization processes on incels.is before, during and after the first lockdown period.
Methodology
Data and the relevant timeframe
The analysis is based on all visible posts within the incel forum incels.is retrieved in November 2022 (Wedel, 2023). To investigate the specific case of the COVID-19 pandemic, we analysed only posts between 1 January 2020 and 31 July 2020, encompassing the beginning and end of the first global lockdown period (Onyeaka et al., 2021). To identify the periods that we would consider the pre-lockdown, lockdown and post-lockdown phases, we focused on the increase in activity on the forum, as previous research has suggested a clear relationship between lockdown and forum activity (Davies et al., 2021; Vu, 2020). A piecewise regression applied on the daily number of posts identified a steep increase in activity between 23 March 2020 and 1 April 2020 and a clear decline in the number of posts between 13 and 29 May 2020 (see Figure 1). The first set of breakpoints coincides closely with 11 March 2020, when the World Health Organization declared COVID 19 a global pandemic and subsequent lockdowns ensued worldwide (Onyeaka et al., 2021).
Fitted piecewise regression on the number of daily posts on incels.is between 1 January 2020 and 31 July 2020.
Based on these results, we adopted three time frames: (T1) pre-lockdown (1 January 2020 until 23 March 2020), (T2) lockdown (01 April 2020 until 13 May 2020) and (T3) post-lockdown (29 May 2020 until 31 July 2020). Posts that fall into the transition phases between the time frames were not considered further, leaving 992,548 out of 1,206,287 posts (78, 47%).
Established toxicity and misogyny measures
Toxicity and misogyny have been measured in the past through two established measures: the Perspective API (2024) for toxicity and a misogyny dictionary (Farrell et al., 2019). The Perspective API is a deep learning language model that represents a minimum estimate of toxicity given its low sensitivity (Trott et al., 2022). We employed the Perspective API in line with previous research on extremist online groups, including incels, using the severe toxicity score (Ribeiro, Ottoni and West, 2021). The developers of the Perspective API define a severely toxic comment as ‘very hateful, aggressive, disrespectful [. . .] or otherwise very likely to make a user leave a discussion or give up on sharing their perspective’ (API, 2024).
The misogyny scores were calculated based on the misogyny dictionary by Farrell et al. (2019). They built a dictionary based on qualitative research to measure the misogyny of online posts. We adapted the dictionary similarly to subsequent research, measuring the misogyny of incel posts as lexicon density (Ribeiro, Jhaver, et al., 2021). This measure describes the fraction of words per post that are part of the dictionary. The concept of misogyny was used as defined by Farrell et al. (2019) for the dictionary construction as a ‘range of hostile and violent activities against women’. For this research, the entire dictionary was used to generate the lexicon density per post.
Both scores have proven helpful in previous research (e.g. Farrell et al., 2019; Ribeiro, Jhaver, et al., 2021; Ribeiro, Ottoni, et al., 2021). To account for the Perspective API’s limit on text input, all posts with more than 2000 characters received neither score. In addition, posts with less than three characters were excluded. Such posts usually contain only a link or media content (video, picture, GIF). A total of 847,346 (82.864%) posts were included and received both scores. The frequency distribution of the excluded posts over time was not biased. For the thread-level analysis, scores were aggregated per thread, and for the user-level analysis, they were aggregated per user, respectively. In both cases, we calculated the mean score for all aggregated posts.
Topic modelling in incel research
Topic models on incel forums have been shown to capture particular topics distinctive from misogynist extremism (Baele et al., 2019). In our topic model, we captured (1) extremism as a topical dimension of content on incels.is. In distinction to pure extremism, we defined three additional thematic categories that are of interest to our research goal in order to capture the content present on incels.is: (2) black pill ideology, (3) the manosphere and (4) mental health. Extremist content refers to an ingroup/outgroup dichotomy, including implications of hostility towards the outgroup, non-ideology-related extremist worldviews and cases of dehumanizing language. The category of extremism is not exclusive to misogynistic extremism. Black pill content refers to the specific ideology of incels, including nihilism and the belief that the social structure is predetermined. This type of content does not necessarily call for violence and is not necessarily extremist but can be. Content on the manosphere includes posts discussing the various beliefs shared among the subgroups of the manosphere, such as men being socially disenfranchised and oppressed. This content does not call for violence and is not necessarily extremist. Finally, mental health content refers explicitly to discussions of suicide and coping mechanisms with mental health problems as a topic of their own (Baele et al., 2019; Regehr, 2022).
We used BERTopic (Grootendorst, 2022) to calculate the topic model. Compared to the more traditional approach of Latent Dirichlet Allocation (LDA), BERTopic allows researchers to more easily arrive at a topic solution that is both inter-thematically coherent and has a sufficient distance between topics from a human and a machine perspective (Egger and Yu, 2022). We calculated topics on a thread level since the purpose and usage of the thread feature commonly imply topic exclusivity. Concatenating all posts in one thread with at least one post in either considered timeframe resulted in a corpus of 49,309 documents. We used the default parameters suggested by the documentation for dimensionality reduction, sentence transformation (all-MiniLM-L6-v2) and outlier reduction with the topic distribution. We removed the English stopwords via the nltk package for topic representations.
We found 385 initial topics which we aggregated manually (e.g. Topic 3 [‘virus’, ‘corona’, ‘coronavirus’, ‘flu’, ‘infected’, ‘china’, ‘chan’, ‘spread’, ‘die’, ‘vaccine’] and Topic 100 [‘virus’, ‘coronavirus’, ‘corona’, ‘immune’, ‘spread’, ‘covid’, ‘ds’, ‘pandemic’, ‘19’, ‘st’]). To do so, we used our initial set of thematic scopes that we expected to find: extremism, black pill ideology, mental health and the manosphere, as well as lifestyle categories including porn, video games, Internet culture or covid. Each of the authors coded each of the 385 initial topics independently of each other based on the top 20 keywords for each topic and one representative document. After this first round, the Cohen’s Kappa score for all classifications was 0.3589. Several topics stood in the blurred space between two or more of our predefined categories. Subsequently, the co-authors went through each topic together, discussed differences, refined the definitions of the different labels and reached agreement on each topic. In addition to those topics that are part of our research design, we detected others (e.g. lifestyle categories) throughout our classification process. However, they are not part of this paper’s analysis.
Statistical tests
To investigate the general change in the content of the forum (RQ1) throughout the three time frames, we compared the prevalence of thematic scopes (extremism, black pill ideology, mental health and manosphere), toxic and misogynist threads as dependent variables between the three time periods T1, T2 and T3 as independent variables.
Our two sets of measurements (thematic scopes and language scores) have different characteristics. For the thematic scopes, we look at a binary variable per thread for each thematic scope. A thread counts as holding positive during a time frame if at least one post of that thread was made during that time frame. Threads can therefore appear in multiple time frames. We turned to non-parametric testing because the data is not randomly sampled from the forum and the dependent variables are not normally distributed. Since we are interested in the differences between each binary thematic scope across three independent groups (T1, T2 and T3) of different sizes, we used a chi-square test to test for an association between the time period and the prevalence of a respective theme. We used pairwise proportion z-tests as post hoc tests between the time frame pairs following each other: [T1, T2] and [T2, T3] to understand how the prevalence of extremist threads changed between the time periods. We used Bonferroni correction to control for multiple (3) comparisons. The misogyny and toxicity scores, on the other hand, were initially calculated per post and are continuous. We calculated the mean score per thread and chose the Kruskal–Wallis test with Mann–Whitney U post hoc tests since we can rely on data from the ratio scale as the dependent variables compared to the binary data of the thematic scopes but still lack normally distributed data. The Kruskal–Wallis test allowed us to detect whether the central tendencies of the samples across the three time frames differed in their central tendencies. The Mann–Whitney U post hoc test subsequently compares the central tendencies for pairwise comparisons.
For the user level (RQ2), we tested the fraction of posts each user posted to threads of the four thematic scopes under investigation, depending on the time frame and rank. For misogyny and toxicity scores, instead of focusing on participation in specific threads, we measured the level of toxicity and misogyny for the posts of each user throughout each time frame and rank. As the distribution of user activity is highly skewed, with most users posting only once or twice during the half-year of data, we excluded 20% of the least active users, resulting in users who posted a minimum of four posts or more. We also cut off the top 2% of active users, resulting in users who posted no more than 2002 times. This left 1481 (79.62%) users for analysis. In addition, we narrowed the sample to users who were active at least once in each of the three time periods (n = 421; 28.43%). In such a way, we address the question of whether the lockdown changed the radicalization trajectory of users on incels.is, focusing on the same set of users throughout all time frames. This allows to investigate effects that could be otherwise concealed behind the general development of the forum.
We used multivariable linear regression models to investigate the possible effects of the time frame and rank variables – one for each independent variable. The dependent variables are based on the time frame and rank variables: given the directional and time-dependent assumptions of this part of the analysis, we focused on the pairwise differences between T1 and T2 and T2 and T3. Therefore, we transformed the time frame variable into three binary variables (T1, T2 and T3). We only included the post-lockdown (T1) and pre-lockdown (T3) variables in our model. In such a way, the intercept represents the lockdown phase, and we can interpret the model results naturally against the lockdown condition.
Including a user’s rank as an independent variable controls for the assumption that users who spend more time and post more on the forum might be or might become more radical (RQ3). The incel.is forum has its own measure of user seniority based on the number of posts a user has posted, which we used for this analysis. We used a user’s current rank when posting the respective post for our model. Each rank level upwards demands the same condition (500 posts and one more week of activity since the previous rank level). Only the 12th rank (administrator) is not related to other ranks with the same condition. Therefore, we excluded these respective users from the analysis (3), leaving us with a final user sample of n = 418. We interpreted the rank variable as a ratio scale (origin is zero, equal distance between each scale level) ranging from 0 to 11. Since the rank variable is considered a possible confounder, we included additional interaction terms between the rank variable and the three time period variables.
Results
We start with reporting on the thread-level analysis (RQ1). We compared the prevalence of different thematic scopes, misogynistic threads and toxic threads (DV) between the three time periods (IV) – before, after and during the lockdown – to explore the short- and mid-term impact of the first global COVID-19 lockdown on the prevalence of extremist, black pill, mental health and manosphere content topics on incels.is.
The three time frames vary in total number of threads (see Table 1). Considering their length, the lockdown (∼1 month) and pre-lockdown (∼2 months and 3 weeks) phases were more active than the post-lockdown (∼2 months) phase. The four thematic scopes together made up over 42% of all forum threads for each time frame, with most threads being extremist in each time frame. All measures except manosphere showed significant differences between at least one pair of time frames (see Table 2). Based on the pairwise tests (see Table 3), we found that the fraction of extremist threads in the forum increased significantly by 1.252% in the post-lockdown period compared to the lockdown period (p = .04*). This is an increase from 1933 to 2212 active extremist threads in absolute numbers. This increase is especially noteworthy since the number of overall active threads decreased by 850 for the same period switch (see Table 1). The fraction of threads on the black pill theme increased significantly from the pre-lockdown to lockdown periods by 1.353% while the fraction of mental health threads decreased from pre-lockdown to lockdown by 1.614% (p < .001***) but bounced back from lockdown to post-lockdown period by 1.284% (p < .001***).
Absolute numbers (fractions compared to all active threads in the respective time frame in brackets) of active threads across themes and time frames are shown.
Thread level comparison. Chi-square test for the thematic scopes and Kruskal–Wallis test for the use of language scores.
Note: *p < .05; **p < .01; ***p < .001.
Thread-level pairwise comparison of thematic scopes, thread-level. Reporting the z-statistic, p-value and the difference in fractions between the respective time frames: pre-lockdown (T1), lockdown (T2) and post-lockdown (T3). Differences are reported as a fraction of either theme in T1 and T2 minus T2 and T3, respectively.
Note: *p < .05; **p < .01; ***p < .001.
The average level of misogyny in the threads increased during the lockdown significantly (p < 0.001***) and dropped significantly afterwards (p < 0.001***) (see Table 4). For toxicity, we found no change with the advent of the lockdown but a significant increase from the lockdown to the post-lockdown period (p < .001***) (see Table 4).
Pairwise comparison of misogyny and toxicity on the thread-level between the respective time frames: pre-lockdown (T1), lockdown (T2) and post-lockdown (T3). Reporting z-statistic and p-value.
Note: *p < .05; **p < .01; ***p < .001.
We now focus on the remaining research questions (RQ2 and RQ3) on the participation of incels.is users in extremist threads, threads on the black pill, threads on mental health and threads on the manosphere and how user seniority impacts their participation as well as the user-level toxicity and misogyny of the language used. Here, we focused only on the users who were active in all three time periods. What follows are the results of our multivariate linear regression models, one for each dependent variable: extremism, black pill, mental health, manosphere, misogyny and toxicity. The independent variables are pre-lockdown, lockdown, post-lockdown, rank and the respective interaction terms between rank and pre- or post-lockdown phase.
All dependent variables in our model show a positive baseline level during the lockdown period. All are statistically significant (p < .001***), implying that we found statistically significantly higher levels for all independent variables during the lockdown phase when not considering rank (see Table 5 and Table 6). For the numerical rank, we found significant negative associations with participation in extremist threads (−.0044; p =< .001***), the manosphere (−.0019; p = .025*) and the use of toxic language (−.0013; p = .004**). Those results indicate that seniority leads to less participation in such threads and a lower level of toxicity in user posts.
User-level comparison of the multivariable linear regression models with thematic scopes as independent variable. Coefficients represent percentage point changes in target variable (fraction of posts in respective thematic threads).
Note: *p < .05; **p < .01; ***p < .001.
User-level comparison of multivariable linear regression models with language scores as the independent variable. Coefficients cannot be interpreted as percentage points since the target variable refers to the average score across a user’s posts.
Note: *p < .05; **p < .01; ***p < .001.
Looking at the time frames T1 and T3, we found a significant negative association for participation in extremist threads compared to the lockdown period for both (−.0317, p < .001**; −.0288, p = .001**). Users posted before lockdown 3.17% and after lockdown 2.88% less in extremist threads. For the black pill threads, we found the same phenomenon: less participation during pre-lockdown (−2.13%; p < .001***) and post-lockdown (−2.32%; p < .001***) periods compared to the lockdown period. For threads on mental health, we found reversed correlations. Here, we observed positive effects during the pre-lockdown (1.27%; p = .013*) and post-lockdown (1.2%; p = .033*) times. Although mental health threads also played a crucial role in the forum during the lockdown (see constant term; 7.35%, p = .001***), participation in such threads was greater, both before and after the initial lockdown period. For manosphere-themed threads and the misogyny level of posts, we did not find significant effects. The toxicity level of user posts showed a positive effect for the post-lockdown phase (.0068, p = .037**).
The effect sizes presented are small because they represent the change in the fraction of participation in a specific thematic thread compared to participation in any other thread. Their significance is shown when considering the specific context of, for example, extremist thread participation. Specifically, the reported decrease of 2.88% in the fraction of participation in extremist threads, while seemingly minor in the broader context of all online engagement, represents a 12.60% reduction in participation in extremist threads when only considering such threads. This highlights the importance of contextualizing effect sizes, as seemingly minor shifts in overall participation can translate into substantial changes within a specific thematic thread. This argument holds for all reported effects. Therefore, the statistically significant effects always mirror relevant real-world changes in user behaviour.
Finally, we look at the interaction terms to assess whether observed effects on the time frame side can be at least partially explained by user seniority. The first interaction term (T1 * rank) showed only non-significant effects, suggesting that the effect of the pre-lockdown period on the dependent variables did not change meaningfully through an increase in the user’s rank. The interaction term T3 * rank was statistically significant (p = .001**) and positive for extremism-themed threads. This suggests that the negative effect of the post-lockdown period on users’ participation in extremist threads was moderated by the user’s rank and vice versa, with each rank increasing the fraction of posting in extremist threads by 0.56%.
To contextualize this interaction, we plotted the distribution of participation in extremist threads for each time frame and rank combination (see Figure 2). We can see a lower level of participation in extremist threads for the last four ranks across all time periods. Higher ranks generally led to reduced participation in extremist threads, regardless of the time frame. However, by plotting the level of participation in extremist threads per rank between time frames, we can make the observation suggested by our regression results (see Figure 3): While the highest ranks held the lowest levels of participation in extremist content overall, those were the ones that increased their activity in extremist threads throughout the time periods (ranks 8, 10, 11 and 12); in other words, whereas the lower rank users did not continue to radicalize beyond the lockdown, the more senior users did.
Distribution of participation in extremist threads across user ranks for each time frame. The figure contains three subplots, one for each time frame (pre-lockdown, lockdown and post-lockdown). Each subplot shows a box-plot for the subset of users of a specific rank during a specific time period. Lower ranks generally correspond to higher participation in extremist threads, regardless of the time period. Outliers are not shown in this figure.
Distribution of participation in extremist threads across time frames for each user rank. The figure contains three subplots, one for each rank (one to twelve). Each subplot shows the distribution of participation in extremist threads for all users of the same rank grouped by time period. Outliers are not shown in this figure.
Discussion
This study investigated the impact of the first COVID 19 lockdown on misogynist extremism within the incel community, operationalized in four key dimensions: toxicity, misogyny, general extremism and the incel-specific black pill ideology.
The lockdown did not lead to an increase in extremist–themed threads on the forum level. However, this effect was found for other dimensions of misogynist extremism, such as the fraction of threads on the core (black pill) ideology of the incel community and the usage of misogynistic slurs. Threads themed with extremism, as well as toxic language, surged only after the lockdown ended. These observations point to the long-term radicalization of general forum communication previously explored (Baele et al., 2023) with temporary increases in ideological threads, which are, after all, the foundation of extremist beliefs within the community, and misogynist language as first steps towards misogynist extremism (Green et al., 2023) and a subsequent increase in extremist topics and toxic language as advanced steps of radicalization. Consumption of ideological content (e.g. black pill) led to the radicalization of a user, even if the original content consumed was not extremist in itself (Ribeiro, Ottoni, et al., 2021). For the impact of COVID-19 specifically, a similar radicalization of far-right groups was found on Telegram (Schulze et al., 2022).
Going beyond studying online radicalization through aggregations at the forum level, we also looked at the user-level effects. Focusing exclusively on users who have been active in every time period yielded different results than the forum-level investigation: users posted more on extremist and black pill–themed threads and less in mental health–themed threads during the lockdown phase than before or after. The seniority and time periods yielded negative effects regarding participation in extremist threads for the pre- and post-lockdown periods. However, their interaction effect was positive for the post-lockdown condition. In other words, more senior users tended to participate more actively in extremist threads after the lockdown than before and during the lockdown (compared to less senior users). Toxicity rose in the post-lockdown phase, whereas misogyny did not change at all – posts during the lockdown were not more misogynist than before or after. However, we found a continuous presence of such slurs, independent of the time period.
Our user-level analysis revealed the role of seniority, which would otherwise have been concealed in a forum level analysis. We found radicalization effects seemingly correlated to the timing of the first lockdown and its aftermath. However, they were not driven by new members as suggested in previous research (Davies et al., 2021) but by senior forum members. Research has shown that members of the incel community participated in a selective (re)interpretation of science to justify their worldviews (Cannito and Ferrero Camoletto, 2022; Daly and Reed, 2022; Johansson Wilen et al., 2024). A simplistic conclusion could see an alignment with the unprecedented doubts about health guidelines and scientific knowledge sparked by the COVID-19 infodemic – the radicalization of senior incels being due to the infodemic coinciding with their practices. However, a key differentiation must be made: the infodemic discourse is focused on anti-intellectualism (i.e. the generalized distrust of experts and intellectuals) (Merkley and Loewen, 2021; Rathore and Farooq, 2020). In contrast to that, incels do not engage in anti-intellectualism in general but are rather fond of science – and believe that scientific findings instead support their perspective (Johansson Wilen et al., 2024). Hence, there is no need to deny it. Therefore, we consider a different explanation to be more likely to have sparked the role of senior members in the observed radicalization phenomena. More senior members of the forum who found ways to cope with real-world issues such as loneliness relapsed back to old habits during the first lockdown and after. The combination of lacking real-world contact (support net) and their incel membership perhaps led to a re-radicalization of these users throughout the times of isolation.
However, the significant drop in mental health–related threads during lockdown also supports the assumption that incels perceived the lockdown time as calming (Vu, 2020) and thus had a positive impact on mental health issues. Therefore, our analysis reveals a striking dichotomy: while the lockdown period saw a general reduction in mental health–related discussions, suggesting a calming effect on newer members, it appears to have triggered a re-radicalization among senior members. This divergence likely stems from these long-standing users having something to lose in the first place (established real-world coping mechanisms) during isolation. In contrast, newer members, lacking these established coping strategies, may have found the lockdown’s isolation less distressing. This diametric effect, where the same event leads to opposite outcomes based on user seniority, warrants further investigation in the future.
Our findings, based on digital trace data, reinforce evidence that the discussion of mental health issues is a core aspect of the incel community and points to participation in such discussions as a coping mechanism. However, mental health issues do not excuse the defamatory and extremist tendencies prevalent in the forum, regardless of the time frame (Kelly, Rothermel and Sugiura, 2024). We showed that the prevalence of mental health topics in the whole forum, as well as the participation of active users in these topics, was lower during the pandemic compared to before and after. Thus, we provided novel insights based on digital trace data for academic discussion on the topic of mental health within the incel community (Kelly et al., 2024; Marwick and Caplan, 2018; Marwick et al., 2022; Moskalenko, Gonzalez, et al., 2022). We interpret these results about the reduced prevalence and participation in mental health topics during lockdown as a proxy for decreased mental health issues (less discussion = fewer problems). This implies that the lockdown phase improved the mental health of incels on incels.is.
When looking at an extremist forum, a sound theoretical grasp of the ideology in question is necessary to understand the potential radicalization effect of a globally disruptive crisis event such as the COVID-19 pandemic. From a methodological perspective, our research showed that topic modelling is helpful for greater insight into extremist content and allows for capturing topical trends over time. Perhaps more importantly, we demonstrated that a multidimensional approach, distinguishing between ideological content, extremist content and different use of language scores, produces much more in-depth and intricate details on the radicalization mechanics of an online group. Future empirical research should explore the definitory and operational separation of a community’s ideological and extremist markers and core topics.
Recognizing that our study focuses only on correlations and temporal associations is crucial. We cannot definitively establish lockdown periods as the causal factors behind the observed changes. Other external events or internal forum dynamics might have been contributing factors. However, the significance of the global lockdown and the lack of external factors of sufficient significance during that time allows us to make our assumptions with confidence. The second crucial limitation is that our findings on the post-lockdown period only account for a limited time period after the lockdown ended. We cannot claim with certainty that the effects found during this period hold up beyond the observed time periods. Suppose the found effects do not hold over the long term. In that case, our findings contribute to an understanding of delayed effects – not everything an event causes must be observable within the exact time of the event. Future research on online radicalization should focus on a multifaceted longitudinal perspective to challenge such limitations.
Conclusion
Our research combines a theoretically sound definition of the specific kind of extremism present within the incel community with the addition of other non-extremist topics using data-driven measurements to explore the impact of the first COVID-19 lockdown on users of the online forum incels.is. We further provide a platform-level assessment and a user-level analysis. This contributes to the existing literature by operationalizing the theoretical soundness of incel-specific extremism as misogynist extremism and by implementing this definition in an empirical study. Our approach, through a topic model and the support of two distinct language scores, allows for a novel perspective, not just on key extremist dimensions of online radicalization but also adjacent themes, such as toxicity, mental health issues and non-extremist ideological discourse (manosphere). This novel, multi-perspective approach allows us to draw nuanced insight on the forum- and user-level beyond the scope of previous research on the topic.
Our findings suggest a complex interaction between the pandemic, radicalization and mental health within the incel community. Although the lockdown period itself did not lead to an immediate surge in extremist content, it coincided with an increase in engagement with the core incel ideology and misogynistic language. Interestingly, the most pronounced increase in extremist content and toxic language occurred after the lockdown, suggesting a delayed radicalization effect, particularly among senior forum members. This highlights the importance of longitudinal studies in understanding the long-term impacts of external events on online radicalization.
Furthermore, the observed decrease in mental health discussions during the lockdown challenges the simplistic notion that social isolation uniformly exacerbates mental health issues. This finding warrants further investigation of the potential coping mechanisms and support structures within the incel community and how they may interact with radicalization processes.
Future research should build on these findings by exploring the long-term effects of the pandemic and the short-term effects of other external events on incel radicalization, examining the specific mechanisms through which seniority contributes to radicalization and further investigating the relationship between mental health and extremism within this community. This research contributes to a deeper understanding of online radicalization and provides valuable data for developing effective interventions and counter-narratives. We provided evidence for short- and long-term radicalization effects that depend on user seniority. These findings hold implications for counter-radicalization efforts, which should focus on the long-term ideological resurgence of senior members rather than solely on the temporary radicalization of new members.
Footnotes
Funding
The author(s) disclosed receipt of the following financial support for the research, authorship and/or publication of this article: We would like to acknowledge funding by the Federal Ministry of Education and Research of Germany (BMBF) under grant no. 16DII131 (Weizenbaum-Institut fur die vernetzte Gesellschaft – Das Deutsche Internet-Institut) and by the European Regional Development Fund project ‘ Beyond Security: Role of Conflict in Resilience-Building’ (reg. no.: CZ.02.01.01/00/22_008/0004595).
