Abstract
Online hostility poses a growing societal challenge, yet quantitative evidence on how social media users respond to different kinds of hostility targeting different identities is limited, even though insights into bystander perceptions are detrimental to combat the online hate endemic. This online experiment (N = 461) examines cognitive (perceived acceptability), affective (negative emotions), and behavioral (intervention intentions) responses to varyingly hostile comments (impolite vs uncivil vs intolerant) directed at a female science communicator with different ethnic (Black vs White) and LGBTQIA+ identity cues (heterosexual vs homosexual vs trans), thus shedding light on intersectional identities comprising social group affiliations with varying levels of marginalization. While intolerance triggered stronger emotional reactions than impoliteness and incivility (likely due to its discriminatory nature), participants were, somewhat paradoxically, more inclined to act (and advocated for more institutional action) against incivility. Furthermore, ethnic cues had a much stronger influence than LGBTQIA+ identity cues across response domains.
Women and other marginalized social groups like the LGBTQIA+ community or people of color (POC) are among the most frequent targets (Reichelmann et al., 2021) of online hostility. Intersectionality theory further suggests that people who are attributed with multiple marginalized statuses are even more vulnerable (Andersen and Hill Collins, 2016). Online hostility also has been documented for high-status public figures like scientists (Desmond, 2022), especially when they are read as female (Isbister et al., 2018).
On a similar note, hostile content tends to be nuanced. According to Rossini (2022), incivility and intolerance constitute distinguishable categories with distinct effects on targets and bystanders (i.e. unaffiliated passers-by who incidentally encounter hostility online; Latané and Darley, 1970). These two categories are not said to be exhaustive, neither upward (where threats have recently been coined; Pradel et al., 2024) nor downward. At the lower end of online hostility, we argue for impoliteness as a mildly aggressive, typically socially accepted form of voicing opposition (Culpeper, 2021) and, thus, the least severe form of hostility.
Within the societal battle against online hostility, unaffiliated users play an essential role either as a silently approving mass or supporting allies (Porten-Cheé et al., 2020). In hostile social media exchanges, every user potentially “acts” as a bystander, deciding whether to engage.
For this study, three research gaps are pivotal: First, it is unclear whether different categories of hostile content are evaluated differently across cognitive, affective, and behavioral measures by social media users. While sound theoretical reasoning is present that incivility and intolerance are differentiated concepts (Rossini, 2022), further empirical evidence on whether users’ responses differ when introducing impoliteness as a comparatively lighter shade is missing.
Second, it is widely recognized that members of marginalized communities, especially those with intersecting marginalized identities, are most frequently targeted and disproportionately impacted by online hostility (Andersen and Hill Collins, 2016), even if they simultaneously hold a privileged status, for example, due to their profession (Isbister et al., 2018). Despite that, there is still a lack of experimental work comparing whether different kinds of target identity cues trigger different responses from social media users and whether holding multiple marginalized statuses stands out as particularly vulnerable.
Third, causal mechanisms behind user responses remain poorly understood given the prevalence of content analyses and surveys. By confronting users with being bystanders and exposing them to different kinds of online hostilities against experimentally varied targets, our current causal understanding may be advanced meaningfully.
The present online experiment contributes to closing those gaps. Specifically, we examine how a quota-representative sample of adults from Germany responds to different kinds of online hostility (impolite vs uncivil vs intolerant) against a female-identifying science communicator whose group statuses were manipulated (Black- vs White-cued as well as heterosexual- vs homosexual- vs trans-identifying).
Many shades of hostility
While hate speech as an extreme variety of online hostility has gained most public attention, scholars agree that such a narrowly defined concept cannot capture the plurality of hostile social media content, which is why various categorizations of online hostility have been published over the years (e.g. Papacharissi, 2004; Rossini, 2022). Among these approaches, central (albeit inconsistently conceptualized) notions are incivility and intolerance, of which both concepts are said to be distinctive in their expressions and societal impact. Accordingly, Rossini (2022) proposed that incivility is primarily a violation of interpersonal norms, including highly aggressive language (e.g. vulgarities, name-calling, insults) and shouting (e.g. via all-caps, exclamation marks). While incivility has been found potentially harmful, it can sometimes also be beneficial, and in some contexts even considered appropriate for political participation and public discourse: For instance, it can facilitate debate if it functions as a marker of passionate engagement or a genuine display of emotion for politics, can help to overcome exclusionary power dynamics, or helps with interpersonal or even intergroup bonding (Masullo Chen et al., 2019).
Some hostile speech may violate deliberative norms by threatening out-groups’ basic human rights (Papacharissi, 2004). Such norm violations are what Rossini (2022) conceptualized as intolerance, which primarily refers to hostile speech against collectively accepted democratic norms in relation to marginalized social groups, including discriminatory, dehumanizing language, malicious stereotyping, and desires to exclude said groups from society (Gervais, 2021; Rossini, 2022). That is, in contrast to incivility, intolerance may constitute a direct threat to democratic values.
When following Rossini’s (2022) conceptualizations of incivility and intolerance, many hostile ways of speaking, especially around the conceptual boundaries, may still be considered inadequately addressed. For instance, violations of politeness norms (e.g. mutual respect, interest in conversation; see Papacharissi, 2004) are hardly covered. We argue that this conceptual gap is consequential given that marginally (passive-)aggressive impoliteness (expressed, e.g. via shallow personal attacks, mildly negative tone, or some disregard for any genuine conversation) may more accurately represent hostile-yet-still-acceptable speech for most passionate public debates on social media (Culpeper, 2021), thus qualifying benefits previously ascribed to incivility in too generalized a way (Masullo Chen et al., 2019). It is theoretically distinguishable from incivility, as impoliteness encompasses subtle or even genuinely humorous remarks that, while less direct and offensive than blatant incivility, still convey underlying negativity and unfriendliness (i.e. light hostility). Due to its normative ambiguity, accounting for impoliteness may thus further open up the gray zone of somewhat questionable discourse practices: Oppositional impoliteness might serve as a suitable civil counterpoint to incivility and intolerance, which, if disregarded, may compromise understanding of either one when examining users’ cognitive, affective, and behavioral responses to each of them.
Social media users’ response across different forms of online hostility
When investigating people’s mental processes, it has a long tradition to discern between cognitive and affective processes (e.g. the cognitive-affective system theory; Mischel and Shoda, 1995). Volition theories (e.g. theory of planned behavior; Ajzen, 1985) further highlighted intentions as a related yet distinguishable concept informed by preceding cognitive and affective processes. Following these traditions, we differentiated social media users’ responses to online hostility into cognitive, affective, and behavioral components.
Users’ cognitive responses in hostile online environments can be understood through the lens of hostile attribution bias, which reflects how individuals interpret a comment’s intent by directly influencing whether they perceive it as socially acceptable (van den Berg and Lansu, 2020). However, the extent to which benevolent correction tendencies mitigate perceived hostility depends not only on the type of hostility but also on its target. In some cases, such reinterpretation of toxic traits can lead to increased acceptability of the interaction (Young Reusser et al., 2024). Perceived acceptability thus plays a crucial role in shaping both bystander interventions and target outcomes. The extent to which online hostility is deemed (un)acceptable varies significantly across individuals and situational contexts (Hurd et al., 2022). Moreover, research suggests that bystander apathy can signal implicit acceptability to others, thereby exacerbating harm for the targeted individuals (Obermaier et al., 2021).
We define users’ affective response via negative emotions given that research has revealed that many (empathic) people struggle to cope with encountering online hostility (van Noorden et al., 2015) and feel helpless about how to respond (Wachs et al., 2020). Further evidence documenting that witnessing online hostility can manifest uncomfortable experiences and can induce disgust (Gervais, 2021) mark the relevance of this connection.
Finally, users’ behavioral response in online environments can be conceptualized through intervention intentions, which for lack of an interactive scenario serves as a precursor of behavior. Findings emphasize the value of both indirect (e.g. flagging, disliking) and direct interventions (e.g. counterarguing; Porten-Cheé et al., 2020). Intervention intentions are known to be contingent upon both personal and contextual factors (e.g. prior victimization, social norms, number of bystanders; see Rudnicki et al., 2022). In summary, these findings indicate that explaining variation in perceived acceptability, negative emotions, and intervention intentions can be considered at the center of the debate about online hostility.
We assume that social media users respond considerably different to impoliteness, incivility, and intolerance. This is based on expectancy violations theory (Burgoon, 1993). This theory posits that users have developed context-specific expectancies concerning anticipated communicative behavior whose negative disconfirmation leads to dismissive evaluations and counteracting responses. Following its propositions, impoliteness may differ from both incivility and intolerance in at least two regards: First, users’ expectations concerning impolite speech on social media may likely not be as firm as for incivility nor intolerance, meaning that negative disconfirmation is less likely to occur (e.g. Vásquez, 2021). Second, given that it is characterized as marginally (passive-)aggressive in tone and substance, violating nothing more than common politeness but neither basic interpersonal nor fundamental deliberative norms, it is likely that impoliteness is considered mildly negative in valence, especially when compared to incivility and intolerance (Culpeper, 2021). Based on this theoretical reasoning, and irrespective of which path may apply more strongly, we therefore expected that:
Rossini (2022) reported that incivility is likely to emerge when there is substantial disagreement across commentators while intolerance flourishes in homogeneous discussions. She further noted that incivility may turn up more openly because it is perceived as less inappropriate compared to intolerance. In other words, incivility elicits weaker responses and can thus be expressed more freely, while instances of intolerant speech cause stronger responses and are therefore more likely to be moderated. These findings suggest that social media users’ responses to incivility and intolerance may differ as it can be expected due to their conceptualizations and due to the more ambiguous role of incivility in public discourse (Masullo Chen et al., 2019). Following these findings, we hypothesized that:
Online hostility toward different marginalized identities
Marginalized minorities are characterized by their systemically grounded assignment of an out-group status, which makes them regularly targeted with hostile behaviors (Matheson et al., 2019). According to social identity theory, individuals who are not assigned to one’s own social identity group are usually evaluated less favorable and more deserving of derogative remarks (Jardina, 2021). Relatedly, individuals who belong to marginalized groups often struggle with social disadvantages, negative stereotyping, and are more likely to suffer from physical and psychological harms (Matheson et al., 2019; Pascoe and Smart Richman, 2009).
Within online platforms, members of marginalized minorities, especially when read as female (Isbister et al., 2018), receive more hostile comments (Reichelmann et al., 2021)—even when they are members of high-status groups (Desmond, 2022). Existing research has revealed that bystander exposure to online hostility can lead to desensitization (Soral et al., 2018), less perceived solidarity (Schäfer et al., 2022), and anti-minority radicalization (Bilewicz and Soral, 2020). Firmer out-group perceptions have also been related to less empathy (van Noorden et al., 2015).
Concerning ethnic minorities in particular, extant studies have documented that subtle cues are often enough to trigger stereotyping, stigma, and out-group thinking (Carlson, 2020) and that racism does not create much distress for the out-group (Karmali et al., 2017)—aligning with expectancy violation theory where it might be more expected, thus producing lesser responses. Furthermore, Hurd et al. (2022) indicated that out-group discrimination may stay unanswered because of lacking awareness and responsibility. Following these findings, we therefore hypothesized in a predominantly White sample that:
Despite the prevalence of anti-LGBTQIA+ online hostility (Reichelmann et al., 2021), evidence on how users respond to it is scarce. For instance, Meglich et al. (2020), again in line with both expectancy violation theory and social identity theory, reported that hostility against lesbians (but not bisexuals) was less likely seen as mistreatment or harmful nor intervention-worthy compared to heterosexual targets. Notably, none of these studies addressed online hostilities. Here, LGBTQIA+ youths often feel that they have to engage in rigorous self-censorship (Buss et al., 2022). However, comparative evidence on whether users’ responses are contingent upon targets’ sexual or gender identity is largely absent. Given this limited comparative evidence, we referred to social identity theory for the following hypothesis in a predominantly cis-heteronormative sample:
Although it has been highlighted that a thorough differentiation of sexual orientations and gender identities is necessary as hostility may unfold differently (Suen et al., 2020), reviews indicate that research has thus far been unsuccessful in this regard (Bauer et al., 2017). Still, given that research is unclear about how users’ responses may differ between non-heterosexual and non-cisgender identities, we opted for merely asking:
It is broadly accepted that people who are attributed with multiple marginalized identities are more vulnerable to (online) hostility (Andersen and Hill Collins, 2016). Being assigned more than one marginalized identity implies that one is more likely subject to resentments, which may enhance rejection (Jardina, 2021). However, there is also a greater chance that other people may sympathize with at least one of the marginalized identities, leading to increased engagement (Amira et al., 2021). Whether these explanations are both causally relevant for social media users’ responses, however, is unclear. Accordingly, we again decided to ask:
Method
A 3 (comment: impolite vs uncivil vs intolerant; within-subject, randomly ordered) × 2 (ethnic cues: Black vs White; between-subject) × 3 (LGBTQIA+ self-identification: heterosexual vs homosexual vs trans; between-subject) online experiment was conducted in April, 2022. Study materials, data, and scripts (and more detailed methods) are available at OSF: https://osf.io/6utpb/. The study was screened and approved to be of minimal ethical risk by the IRB of the Faculty of Social Sciences (subunit: Department of Communication) at University of Vienna (ID: 20220311_013).
Participants
The sample was recruited by a market research company based on representative quotas for age (limited from 18 to 65 years), gender, and education in Germany. The sample consisted of N = 461 participants (age: M = 42.72, SD = 13.61, range: 18-65 years), including n = 234 self-identified women and n = 227 men. Concerning participants’ highest educational degree, we coded n = 69 into a lower-education category (i.e. no educational degree or lower-secondary education), n = 264 into a medium-level education category (i.e. secondary education), and n = 128 into a high-level education category (i.e. completed university education). Participants were predominantly heterosexual (n = 390) and identified as White/Caucasian (n = 401).
Stimulus material
Online hostility target—between-subjects factor
We created six Twitter profiles of a female-identifying moderately popular (i.e. 1.221 Tweets, 379 Followers) science communicator named Dr. Johanna Meser (@sciencejo). Reflecting the experimental conditions as between-subjects factor, the six profiles differed in (a) who was presented in the profile photo (White- vs Black-cued) and (b) how the communicator described herself in her bio (heterosexual vs homosexual vs trans; see Figure 1).
Stimulus material concerning the online hostility target.
We crafted a Twitter post in which our target argues against a continuation of “disputed” rent-cap policies and for alternatives to reduce housing prices (see Figure 2). The topic and was intended to create a mild conflict due to its counterproposal to fairly popular rent-cap policies. See OSF for a more detailed reasoning on the stimulus design.
Stimulus material concerning the targeted communicator’s post.
Hostile user comments—within-subjects-factor
To balance out comment-level idiosyncrasies, participants were presented with a fix-ordered set of three allegedly public Twitter comments in each experimental condition. Visible meta-content (i.e. usernames, Twitter handles, profile pictures) was designed to be high in both external and internal validity (see OSF for detailed descriptions). The “like” counts were randomly assigned between 25 and 70, reflecting plausible engagement numbers for similar content and profiles. In practice, they ranged from 29 to 61. As this variation occurred within the manipulation, we considered its impact negligible while ensuring differences within each condition remained relatively consistent without appearing artificial.
Each set of comments was manipulated in their content (see Figure 3). Following prior work (Masullo Chen et al., 2019), we designed: (a) impolite comments including passionate, mildly (passive-)aggressive language; (b) uncivil comments comprising vulgarities, pejoratives, name-calling, all caps, heavy punctuation; and (c) intolerant comments which include cynical, discriminatory, dehumanizing language, malicious stereotyping, and a desire to remove “them” from society. Aside from this, other stimulus features (i.e. emoji use, colloquial language, spelling errors, characters, see OSF for details) were kept equal across conditions. All participants were exposed to all three kinds of hostility consecutively as a within-subjects factor. The presented order was randomized.
Stimulus material concerning impolite, uncivil, and intolerant user comments.
Procedure
This study was part of a data collection that included another thematically unrelated project. First, information material was presented including a trigger warning and self-help resources for digital violence. Participants specified their sexuality and ethnic background and then completed items concerning their social media experience, Twitter experience, and attitude toward rent-cap policies. After they were instructed to “carefully read” the assigned Twitter profile and were presented with manipulation checks.
Next, participants were instructed to read Dr. Meser’s tweet about rent-cap policies and then saw hostile user commentary. After each set of comments, participants completed another manipulation check and answered items about perceived acceptability, negative emotions, and intervention intentions. Finally, we asked them about their attitudes toward LGBTQIA+ people as well as POC. Afterward, participants were debriefed, thanked, received complete information about the study, and given the opportunity to provide feedback.
Measures
Since the survey was conducted in German and due to a lack of validated translations, all measures were translated by us. Most items were slightly adapted to suit the stimulus material; the exact item wordings can be found on OSF. Scales were presented in a fixed order while item order was randomized.
Dependent variables
Cognitive response: perceived acceptability
We measured participants’ perceived acceptability with a self-constructed four-item scale inspired by Oliveira and Levine (2008). The scale ranged from 1 = I do not agree at all, to 5 = I agree completely (ωs = .92–.95).
Affective response: negative emotions
Based on Clark and Watson (1994; PANAS-X), participants were asked to state how strongly (1 = not at all, 5 = very much) they felt the following emotions during reading the user comments: “anger,” “revulsion,” “aversion,” and “disgust” (ωs = .93–.95).
Behavioral response: intervention intentions
Participants were instructed to indicate how much they agree (1 = I do not agree at all, 5 = I agree completely) with seven statements about intervention intentions. Exploratory factor analyses (EFAs) suggested a three-factor solution, including (a) inaction, (b) direct intervention, and (c) institutional intervention calls (see OSF for details). Reliability coefficients varied between acceptable and excellent across subscales and conditions: Spearman–Brown’s ρs = .78 –.81 for inaction, ρs = .91–.94 for direct intervention, and ρs = .76–.90 for institutional intervention calls.
Covariates
Concerning varied experience levels with the stimulus material, we assessed participants’ experience (1 = very low, 5 = very high) with social media in general (M = 3.27, SD = 1.04; ω = .79) and Twitter in particular each via three self-constructed items (M = 1.69, SD = 1.60; ω = .96). We further accounted for participants’ attitude toward rent-cap policies (1 = strongly opposed, 5 = strongly in favor) via a single item. As expected, participants on average maintained a favorable position, M = 4.01, SD = 1.00.
Finally, we measured attitudes toward LGBTQIA+ individuals and POC. For the former, we used five items from Woodford et al. (2012; ω = .87); for the latter, we employed four items from Henry and Sears (2002; ω = .80). For all item statements, participants were asked for agreement on 5-point Likert-type scales (1 = not at all, 5 = entirely). Overall, participants demonstrated a favorable attitude toward LGBTQIA+ individuals, M = 3.90, SD = 1.01, and low average levels of (symbolically) racist attitudes against POC, M = 2.33, SD = 0.95.
Statistical analysis
Following recent methodological calls (e.g. McNeish and Wolf, 2020) to reconsider the standard “alpha-and-sum” approach in favor of factor scores, we entered congenerically weighted Bartlett scores (allowing for each item to contribute differently) instead of unit-weighted sum scores (where each item contributes identically) into further statistical models. All analyses were performed with R. Robust fit indices were applied where possible.
Results
Tables were created to facilitate an overview of the data. Table 1a presents descriptive information on continuous variables and Table 1b on categorical variables. Zero-order correlations between all variables can be found on OSF.
Descriptive information for all continuous parameters.
This table is the property of the authors.
Descriptive information for all categorical parameters.
This table is the property of the authors.
Randomization check
Randomization checks were conducted for both between-subject conditions (i.e. ethnic cues and LGBTQIA+ self-identification). Although we found a small yet significant imbalance of participant gender across the LGBTQIA+ identity conditions, we accepted that our randomization procedure was valid. Detailed results are presented on OSF.
Manipulation check
All three experimental manipulations were subject to manipulation checks (see OSF for detail results). Results demonstrated that (a) 93.85% White- and 99.54% Black-read communicators were in the White- and Black-cued condition, respectively; (b) 53.70% heterosexually read communicators were in the heterosexual condition, 70.78% homosexual-read communicators were in the homosexual condition, and 91.30% trans-read communicators were in the trans condition (although it must be noted that the homosexual- and the trans-identified condition was misread considerably to 25.97% and 54.49%, respectively); and (c) a strong majority of participants matched the stimulus comments with the intended descriptions: 85.25% in the impolite, 80.48% in the uncivil, and 56.24% in the intolerant condition.
Main analysis
Mixed analyses of covariance (ANCOVAs) were conducted using comment type (impolite vs uncivil vs intolerant) as a within-subjects factor as well as ethnic cues (Black vs White) and LGBTQIA+ identity (heterosexual vs homosexual vs trans) as between-subject factors. Age, gender, education, attitudes toward LGBTQIA+ individuals, POC, rent-cap policies, and social media and Twitter experience were covariates. In addition, we included within-level interactions between comment type and all other variables and a between-level interaction between ethnic cues and LGBTQIA+ identity. Descriptives are presented in Table 1a.
Impoliteness versus incivility versus intolerance (H1ac and H2ac)
Perceived acceptability
Results showed a main effect, F(1.64,730.28) = 327.35, p < .001, part. η2 = .423. Post hoc test revealed significant differences between impoliteness and both incivility, t = 28.03, p < .001, Cohen’s d = 1.511, and intolerance, t = 28.89, p < .001, d = 1.450; ratings for uncivil and intolerant comments did not differ, t = −1.14, p = .256, d = −.061. Thus, H1a was supported by our data, while H2a was not.
Negative emotions
Again, a main effect emerged, F(1.93,859.29) = 351.59, p < .001, part. η2 = .441. Post hoc testing demonstrated significant differences across all conditions, that is, between impoliteness and both incivility, t = −25.91, p < .001, d = −1.258, and intolerance, t = −28.47, p < .001, d = −1.383, and between incivility and intolerance, t = −2.57, p = .010, d = −.125. Both H1b and H2b were supported.
Intervention intentions
Significant main effects were found for inaction, F(1.97, 876.33) = 36.16, p < .001, part. η2 = .075, direction intervention, F(1.78,793.55) = 15.93, p < .001, part. η2 = .034, and call for institutional action, F(1.96,873.12) = 370.79, p < .001, part. η2 = .454. With one exception (incivility vs intolerance regarding direct intervention), post hoc testing demonstrated significant differences across all conditions across all three dependent variables (see OSF for detailed statistics). While H1c was supported across all three operationalizations, both significant differences between incivility and intolerance turned out as the inverse of H2c as participants scored higher for the former than the latter.
White versus Black communicator (H3ac)
Perceived acceptability
We found neither a main effect, F(1,446) = 1.33, p = .250, part. η2 = .003, nor an interaction effect with comment type, F(1.64,730.28) = 2.35, p = .107, part. η2 = .005. Accordingly, H3a was not supported.
Negative emotions
Analysis documented a main effect, F(1,446) = 6.88, p = .009, part. η2 = .015, but no significant interaction effects with comment type, F(1.93, 859.29) = 1.70, p = .184, part. η2 = .004. The main effect was mainly driven by significant differences in the impolite and the intolerant comment conditions. Unexpectedly, descriptives revealed that participants experienced stronger negative emotions when a Black-cued communicator was targeted, such that H3b was not supported.
Intervention intentions
For inaction, neither a main effect, F(1,446) = 0.77, p = .381, part. η2 = .002, nor an interaction effect with comment type was detected, F(1.97, 876.33) = 0.37, p = .688, part. η2 < .001. For direction interaction, however, a main effect did emerge, F(1,446) = 7.82, p = .005, part. η2 = .017, that was not further qualified by an interaction effect, F(1.78, 793.55) = 1.03, p = .352, part. η2 = .002. Notably, said main effect was primarily driven by the incivility and intolerance conditions and, given that participants scored higher in the Black-cued conditions, ran counter to our assumption. Similar effects can be reported regarding calls for institutional intervention where we had a main effect, F(1,446) = 11.38, p < .001, part. η2 = .025, but no significant interaction effect, F(1.96, 873.12) = 0.21, p = .807, part. η2 < .001. Across all comment types, participants’ calls for institutional interventions were stronger when they had seen comments against a Black-cued communicator. In summary, our data provided counterevidence to H3c.
Heterosexual versus homosexual versus trans communicator (H4ac and RQ1ac)
Perceived acceptability
Results showed no main effect across LGBTQIA+ identity conditions, F(2,446) = 2.82, p = .060, part. η2 = .013; however, an interaction effect with comment type was found, F(3.27,730.28) = 2.97, p = .027, part. η2 = .013. Post hoc testing revealed a significant difference between homosexual and trans communicators in the impoliteness condition only, t = 3.38, p = .013, d = .389, suggesting that impolite comments were more acceptable when addressing a homosexual than a trans communicator (not supporting H4a but answering RQ1a).
Negative emotions
Neither a main effect, F(2,446) = 1.26, p = .285, part. η2 = .006, nor a significant interaction with comment type were detected, F(3.85, 859.29) = 0.25, p = .905, part. η2 = .001. This means that our data provided no support for H4b and answered RQ1b by stating no meaningful differences.
Intervention intentions
Similar to negative emotions, we found no significant main effects nor interaction effects with comment type across inaction, direct intervention, or institutional intervention calls (see OSF for detailed statistics). Accordingly, our data did not support H4c and answered RQ1c without meaningful differences.
Vulnerability of individuals with multiple marginalized identities (RQ2)
Across acceptability, negative emotions, and all three indicators for intervention intentions, we found not a single significant interaction effect between ethnic cues and LGBTQIA+ identity or significant three-way interaction with comment type (see OSF for detailed statistics). Therefore, we answer RQ2 by stating that our data provided no evidence for the impact of holding multiple marginalized identities.
Influence of covariates
Perceived acceptability
Covariate analysis showed higher acceptability scores for younger and male participants, as well for those with worse LGBTQIA+ and POC attitudes. Interactions with comment type were found for gender, education, LGBTQIA+ attitudes, and Twitter experience.
Negative emotions
Stronger emotional responses for older participants and women, as well as for those with more positive POC attitudes and greater social media experience were detectable. Interactions with comment type were found for education, LGBTQIA+ attitudes, and POC attitudes.
Intervention intentions
The analysis showed lower inaction scores for older participants and those with more positive POC attitudes. Stronger direct intervention intentions were found for older participants, those with lower education levels, and those with greater social media experience. Interactions with comment type were found for LGBTQIA+ attitudes and POC attitudes.
Misreading of target
No significant predictors for misreadings of a homosexual communicator were found in relation to covariates. In contrast, higher misreading rates of a trans communicator were found for older participants, those with lower education levels, and those with less positive LGBTQIA+ attitudes.
More details regarding the influence of covariates can be found on OSF.
Discussion
Witnessing online hostility is commonplace for many social media users (Reichelmann et al., 2021). Female-read individuals, members of marginalized communities, and, as of late more often, scientists are preferred targets (Desmond, 2022). Within these communication processes, social media users quickly become bystanders who can apathetically allow hostility to unfold or act as a target’s ally against it. However, for these bystanders to be better understood, more empirical insights are needed into how users respond to different online hostilities across targets with varying characteristics.
To contribute to this mission, we presented findings from an online experiment where participants were confronted with varyingly hostile content (impolite vs uncivil vs intolerant) targeting a female-identifying science communicator whose stimulus Twitter profile was experimentally manipulated with regard to ethnic cues (Black vs White) and LGBTQIA+ identity (heterosexual vs homosexual vs trans). In summary, while hostility nuances show complex patterns, ethnic cues were generally more influential than markers of LGBTQIA+ identity (possibly because many of our participants appeared to lack understanding for the latter), and membership to multiple marginalized communities did not stand out as more vulnerable.
Content nuances: obvious minor hostilities, intricate major ones
Concerning the weakly evidenced empirical distinction between impoliteness, incivility, and intolerance, we found that incivility and intolerance may be considered hostilities of another level in comparison with impoliteness. They were both perceived as much less acceptable and produced much stronger negative emotions and intervention intentions than impoliteness, which suggests that most users consistently distinguish between minor and major norm violations within social media.
For both major violations, the picture appears more complicated. According to literature (e.g. Rossini, 2022), intolerance is conceptualized as a more problematic offense than incivility (which may even have desirable outcomes; Masullo Chen et al., 2019). This assumption was supported only for negative emotions (in line with Culpeper, 2021, but in contrast to Gervais, 2021). An important avenue for further research may be to declare incivility (but not intolerance) as an appropriate emotional expression that signifies passion rather than hostile intent may reduce superficial reactance or negative attributions and thus help with establishing positively controversial discourse without self-censorship. Exploring the spectrum of positive emotional experiences could further enrich current research.
Surprisingly, participants evaluated incivility and intolerance as similarly acceptable, and even called for more decisive institutional action against incivility. What is most intriguing are the inverse patterns in affective and behavioral responses. Although incivility includes interpersonal aggression, it was intolerance with its discriminatory and anti-democratic contents that led to stronger affective reactions, suggesting better awareness and amplified sensitivity among participants. Despite awareness and sensitivity, participants nevertheless appeared to be more willing to go after incivility than intolerance despite more upset feelings when confronted with the latter. Given that expectancy violations theory posits that stronger deviations from normative expectations should elicit more pronounced responses (Burgoon, 1993), our results indicate that incivility and intolerance may be perceived as similarly anti-normative, at least when it comes to cognitive and behavioral measures (for which additional considerations might need to be acknowledged; see Meerson et al., 2025). Together, this suggests that cognitive, affective, and behavioral outcomes of witnessing norm-violating speech might not necessarily draw on the same underlying mechanisms, possibly due to normalization of such language in online spaces. However, to put this into perspective, they were both perceived as much less acceptable and produced much stronger negative emotions and intervention intentions than impoliteness, which suggests that most users distinguish between minor and major norm violations within social media.
Several explanations for these patterns are plausible. It is possible that users want to prioritize self-protection and consider intolerant discussants as potentially more harmful (Wong-Lo and Bullock, 2014). Akin to findings on out-group harassment (Obermaier et al., 2021), another potential explanation could involve stronger insecurities about how to respond to intolerant comments (Wachs et al., 2020). Possibly, participants thought it even probable to be targeted by incivility themselves in the future. To shed light on those complex matters, future studies with fine-grained methods (e.g. think-aloud protocols) are needed.
Ethnic minority status: out-group favoritism or internalized attitudes
Concerning participants’ responses to the experimental manipulation of ethnic cues, results turned out opposite to what we expected based on social identity theory for a primarily White sample (Jardina, 2021). Faced with a Black-cued female science communicator receiving any kind of hostile comments, participants experienced a stronger negative emotional response, stated a higher likelihood for intervening directly, and also called for stronger institutional intervention. Given the sample’s favorable attitude toward POC, it appears also likely that participants generally condemn (online) attacks against ethnic minorities. Since these patterns did not occur equally across hostile content conditions but selectively in the case of both negative emotions and direction intervention intentions, or, in the case of institutional intervention calls, with varying effect size, our data further highlights that participants appear to be receptive for nuances.
Limited effects for sexuality, gender identity, and multi-marginalization
In contrast to ethnic cues, we found only limited evidence that LGBTQIA+ self-identification or membership to multiple marginalized communities affected participants’ responses. It could be argued that visible indicators of someone’s sexuality and gender identity are typically relatively subtle (here operationalized via bio self-descriptions and pride flags), thus reducing salience.
We noticed 25.97% of our sample misread the homosexual communicator as heterosexual, and 54.49% misread the trans communicator. Although not intended, this reflects a broader reality, namely that even explicitly stated or visually signaled queerness is often overlooked. Being a finding in itself, this can serve as a protective advantage in avoiding anti-LGBTQIA+ hostility but also hinders normalization, visibility, and community-building efforts. Despite very few participants in our sample exhibiting an unfavorable attitude toward LGBTQIA+ individuals, recognition gaps persist. Heteronormativity in digital spaces as well as methodological refinements in strengthening identity-based manipulations warrants further study.
This may also have given attributional primacy to ethnic cues over varying sexuality and gender identity (or multi-marginalizations), which might also explain the lack of interaction effects regarding multiple marginalized identities. Irrespective of whether these explanations are valid, the non-impact of LGBTQIA+ status may also explain consistently non-significant findings for membership to multiple marginalized communities. However, instead of disregarding the well-established assumption that multiply marginalized people are particularly vulnerable to bystander apathy or worse responses (e.g. Suen et al., 2020), we understand these findings as yet another call to further explore and test how people process identity-relevant information of others in greater breadth, for instance, across the whole LGBTQIA+ continuum (Bauer et al., 2017), and also greater depth, including several layers of marginalization. For instance, based on our study, this follow-up research question should be tested in a qualitative context including exposure to digital profiles with (multiple) marginalized identity cues: How do bystanders’ (a) heteronormative beliefs, and (b) LGBTQIA+ knowledge, attitudes, and affiliation influence their recognition of LGBTQIA+ identities in digital contexts and their subsequent intervention intentions in response to online hostility targeting LGBTQIA+ individuals?
Demographics, attitudes, and online experience may matter
We found patterns that older participants viewed hostile comments as less acceptable, experienced stronger negative emotions, and were more likely to intervene themselves (but, notably, without calling institutions to do so) compared to younger participants. Unsurprisingly, similarly consistent trends were identified for LGBTQIA+ and POC attitudes where more positive stances were related to stronger awareness, empathy, and intervention intentions. People with lower education may be less sensitive to content nuances (similar to their earlier documented insensitivity toward gender identities). Together, these results provide another strong argument for further investigating why certain characteristics may (not) be predictive of social media users’ immediate response to online hostility.
Future research should examine how bystander affiliation—such as allyship with LGBTQIA+ individuals or racial minorities—shapes responses to online hostility. Power dynamics and in-group/out-group relations significantly influence perceived hostility and intervention (e.g. Obermaier et al., 2023).
Generally, the overall findings have clear practical implications: Our findings highlight the need for tailored counter strategies for different forms of online hostility. While impoliteness may be managed without many issues through community norms, incivility and intolerance appear to demand stricter measures, for instance, by means of content moderation, and distinct nudging strategies for facilitating bystander intervention to overcome considerations against stepping up (e.g. fear to be targeted oneself, hopelessness if engagement might make a difference; see Meerson et al., 2025). Specifically, given its scale, it might be promising to implement AI-driven moderation tools to help users with addressing situational uncertainties related to the bystander intervention model (i.e. recognition, urgency perception, personal responsibility, knowledge and skills evaluation, and intervention enaction; Latané and Darley, 1970).
Limitations
In addition to what was already mentioned above, this study is subject to limitations. Apart from some usual suspects of online experiments (i.e. self-report measures), stimulus material can be criticized. We fully acknowledge that our discrete operationalizations of ethnicity, sexuality, and gender identity cannot fully cover how they typically emerge in real (virtual) environments; however, we implemented simplistically distinct conditions to maximize effects as well as internal validity.
Similar criticism can be voiced for content manipulation where we also noticed somewhat considerable variance in how participants (mis)understood our examples of impoliteness, incivility, and intolerance. Being in line with previous research (e.g. Culpeper, 2021), this variance can be considered a meaningful empirical insight in itself and again calls for more testing.
While we did not measure participants’ expectations of impolite, uncivil, or intolerant comments to avoid priming effects and confirmation bias, future research could explore this through other methods, such as think-aloud protocols, or by employing clever multi-phase designs.
The cultural context of our German sample may limit generalizability. Given Germany’s relatively positive attitudes toward LGBTQIA+ individuals and POC, bystanders’ responses to online hostility may differ in societies with stronger anti-LGBTQIA+ or racial biases. Future research should explore how cultural norms shape perceptions of and interventions against anti-LGBTQIA+ hostilities in less accepting or non-Western settings, where legal, social, and ideological factors may alter hostility prevalence and bystander intervention. This also applies to a more diverse sampling of subpopulations. While this was neither feasible nor aimed at within the present study, we acknowledge it as an important avenue for future research.
Finally, while we stand by our argument that active social media users are most likely bystanders on regular basis, we still acknowledge that they can only be referred to as potential bystanders in this study given the non-interactive stimulus scenario. Actual behavior may be tested with more sophisticated methods using interactive stimulus environments.
Conclusion
Online hostility is a pervasive modern-day misery, particularly for members of marginalized communities. The present study reports findings from a stimulus-based online experiment investigating immediate responses from social media users after being exposed to different kinds of hostile commentary (i.e. impoliteness, incivility, and intolerance) against a (fictional) female-identifying science communicator for whom we manipulated both ethnic cues and LGBTQIA+ self-identification. Findings shed light on how social media users differentiate between online hostilities and how target characteristics influence their responses, advancing the current literature both empirically (by testing for causal effects of content nuances and target characteristics) and conceptually (by empirically substantiating recent taxonomic categorizations). On a brighter note, our results indicated some promising responses from users to hostile comments against marginalized ethnic groups that may be a first step into conquering perpetrators with intergroup solidarity.
Footnotes
Acknowledgements
The authors want to thank Iara Noronha dos Reis for assisting in drafting parts of the stimulus material.
Author contributions
Melanie Saumer conceptualized the study, contributed to data collection and analysis, and drafted the manuscript. Kevin Koban contributed to conceptualizing the study, contributed to data collection and analysis, and assisted in drafting the manuscript. Jörg Matthes assisted with data collection, interpretation, and final manuscript review.
Funding
The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This research was supported by the European Research Council (ERC) under the project Digital Hate: Perpetrators, Audiences, and (Dis)Empowered Targets (Grant Agreement No. 101055073). Melanie Saumer is additionally funded by the Austrian Academy of Sciences (ÖAW) through a DOC fellowship.
Ethical considerations
An ethical screening of this study was conducted by the Institutional Review Board (IRB) of the Faculty of Social Sciences (subunit: Department of Communication) at University of Vienna under the ID 20220311_013. No ethical concerns were reported.
Consent to participate
Informed consent was secured from all participants in an opt-in manner, where participants had to click “I consent to participating in this study” after having read an extensive informed consent briefing and before being forwarded to the actual study.
Consent for publication
The notion that the results of this completely anonymized quantitative online study will be published has been included in the informed consent as well. Participants agreed to it before proceeding to the study.
