Profiling Digital Hate: A Multidimensional Measurement Approach Based on the Perspective API

Promises and Pitfalls of the Perspective API

To aid automatic classification of potentially “toxic” (i.e., hateful) content at large and global scales, the Perspective API, developed by Google Jigsaw, is frequently utilized, particularly in social scientific research (e.g., Blumer & Kleinberg, 2025; Gervais et al., 2025) and for content moderation tasks due to its cost-free and relatively easy-to-use implementations in Python and R (e.g., Votta, 2019), as well as its ability to perform beyond English across multiple languages (e.g., Nogara et al., 2024). Its latest version further employs a multilingual, token-free architecture that is robust to code-switching, character obfuscation, and domain shifts, supposedly enabling the accurate detection of hateful content across contexts (Lees et al., 2022).

Currently, the Perspective API provides probability scores ranging from 0 to 1 for six attributes, with partially overlapping descriptions: (a) Toxicity ¹ (i.e., rude, disrespectful, or unreasonable comment that is likely to make people leave a discussion), (b) Severe Toxicity (i.e., very hateful, aggressive, disrespectful comments, high likelihood to make a user leave a discussion, less sensitive to more mild forms of toxicity), (c) Identity Attack (i.e., negative or hateful comments targeting someone because of their identity), (d) Insult (i.e., inflammatory or negative comments), (e) Profanity (i.e., swear or curse words, other obscene or profane language), and (f) Threat (i.e., intention to inflict pain, injury, or violence; Google Jigsaw [2025a]). Additionally, the model also provides a variety of lesser-tested attribute scores (e.g., Sexually Explicit for references to sexual acts, body parts, or other lewd content), which are still considered preliminary. In practice, different thresholds have been used for classification tasks (e.g., Mihaljević & Steffen, 2023; Oh & Downey, 2025; Weber et al., 2025).

Whilst not much is known about the concrete model training, it is documented that Perspective API models were trained on millions of comments from diverse online sources and languages, including Wikipedia and The New York Times. According to Google Jigsaw (2025b), each comment is annotated by three to ten native-language raters who assess the attributes using standardized guidelines, with final labels derived from a relative agreement among raters concerning a given attribute. Benchmark evaluations have shown great promise for the API. For instance, Muralikumar et al. (2023) found high alignment between Toxicity scores and human participants’ ratings of toxicity. Other benchmarks have shown that it is able to classify toxic language (Hede et al., 2021) and explicitly antisemitic content, while struggling with more implicit utterances (Mihaljević & Steffen, 2023).

Indeed, prior evaluations focused predominantly on the Toxicity score (e.g., Gervais et al., 2025; Hede et al., 2021; Kim et al., 2021; Oh & Downey, 2025; Schmidt et al., 2024; TeBlunthuis et al., 2024), thereby treating digital hate as a unidimensional construct while also capturing only a narrow aspect of the API’s full scope. These evaluations have also highlighted several limitations, including a lack of transparency and reproducibility (Pozzobon et al., 2023) and various relevant cognitive biases (i.e., a tendency to assign higher Toxicity scores to German comments compared to other languages; Nogara et al. [2024]). The Toxicity score has also been criticized in terms of performance (e.g., Rosenblatt et al., 2022), especially for certain subtypes of digital hate (e.g., political incivility), so much so that researchers concluded that, in line with what Google Jigsaw (2025c) themselves stated as intended, API scores should be used to aid human judgment in form of a socio-technical information system instead of using them on their own (Gervais et al., 2025). Moreover, Rieder and Skop (2021) argue that the API’s design, especially its reliance on binary classification and decontextualized inputs, can obscure complexities of human communication and reinforce platform governance logics that prioritize automation over deliberation. Accordingly, what is needed from a (computational) social science, but also from a practical perspective, is an approach that leverages multiple lower-level indicators and anchors them in a theoretically meaningful distinction between different higher-order forms of harmful communication (e.g., Matthes et al., 2025; Oh & Downey, 2025). In this study, we propose such a multi-indicator measurement approach by using lower-order constructs (i.e., Perspective API attributes) to predict higher-order constructs of digital hate, rather than relying on single-score proxies. This helps us move beyond treating “hate” as a single construct by drawing on existing theorization to distinguish between two key forms (i.e., incivility and intolerance; Rossini, 2022) that provide the conceptual foundation for structuring and evaluating these indicators.

Digital Hate in Theory: Incivility vs. Intolerance

An overwhelmingly broad range of often conceptually overlapping hateful online phenomena has been observed, including (but not limited to) manifestations of prejudice and discrimination (e.g., Kopytowska & Baider, 2017), aggressive behaviors (DeMarsico et al., 2022), disinformation and other manipulation tactics (Weimann & Pack, 2023), and individual- or group-oriented harassment (Valenzuela-Garcia et al., 2023). Developing training datasets that accurately reflect this diversity presents a great methodological challenge that requires striking a delicate balance between preserving linguistic and contextual variability and generating abstracted representations that are both conceptually coherent and analytically usable. However, while capturing subtle nuances is important, an excessive focus on details can hinder the extraction of broader and, more importantly, generalizable patterns. Said differently: An overemphasis on preserving every nuance can obstruct sufficiently transferable insights, underscoring the need for balancing complexity with theoretical and operational parsimony (Healy, 2017).

Following a long-standing typological discourse about hostile online expressions (see, e.g., Meddaugh & Kay, 2009), this study draws on the distinction between online incivility and intolerance as a theoretically framework to structure different forms of digital hate, two related but distinct key concepts used to meaningfully structure linguistically versatile hostile expressions originating from conceptually overlapping phenomena via different kinds of norm violations. Incivility refers to an offensive communication style (e.g., marked by impoliteness and harsh tone, including profanities, insults, character assassination, and outrage; Bianchi et al. [2022]) that in most contexts can be regarded as violations of common interpersonal communication norms, whereas intolerance refers to content that violates egalitarian democracy norms, expressing harmful, discriminatory, or exclusionary intent toward groups or viewpoints (e.g., marked by prejudice, hatred, threats; Rossini, [2022]; hate speech, dehumanization, serious threats, personal abuse and harassment, democratic threats; Bianchi et al. [2022]). This distinction between incivility and intolerance is not only important on a conceptual but also on a perceptual level, as intolerance is often perceived as more severe compared to incivility (e.g., Kümpel & Unkel, 2023).

In essence, incivility tends to be equated to simple impoliteness: an attack on an individual’s “face” in interaction, breaching conversational etiquette (Culpeper, 1996; Papacharissi, 2004). While it has been shown that incivility is associated with lower perceived discussion quality (Wang, 2020) and lower perceived rationality (Popan et al., 2019), it could not be shown that incivility directly leads to reduced quality of deliberative discourse (Schroll & Huber, 2022); however, it has been shown that it can lead to silencing effects (Min & Shen, 2023). Contrary to this, uncivil content can also serve as a form of emotional or moral expression, especially for those with limited institutional or societal power (Chen et al., 2019; Stryker et al., 2016), and has even been shown to elicit positive emotions (Kosmidis & Theocharis, 2020). Given that incivility may be best considered as ambiguous with notable constructive effects on public discourse, particularly within certain communities, there is growing consensus that such instances should not necessarily be subject to moderation (Oh & Downey, 2025).

To examine whether multiple Perspective API scores can be meaningfully combined to capture these higher-order constructs, we derive hypotheses linking specific attributes to online incivility and intolerance. Using the Perspective API’s scoring system, we hypothesize that Toxicity, Insults, and Profanity will positively predict incivility as these scores most closely capture forms of expression that disrupt norms of respectful interpersonal discourse but do not involve anti-democratic discrimination or dehumanization of targets (Rossini, 2022). Previously, they have also been investigated under the incivility label (e.g., toxicity in Anderson et al. [2018]; insults and profanities in Bianchi et al. [2022]). The Toxicity score has consistently been used as a proxy for incivility (e.g., Hopp et al., 2018; Schmidt et al., 2024). Importantly, Gervais et al. (2025) explicitly examine whether Perspective API’s Toxicity score functions as a valid indicator of political incivility, demonstrating that this measure aligns strongly with interpersonal disrespect and norm-violating tone. Furthermore, Coe et al. (2014), in a foundational content-analytical study, identify name-calling, insults, and profanity as central and defining indicators of incivility in mediated discourse. By isolating these features, we capture incivility as a stylistic violation of discourse norms rather than as a substantive rejection of others’ rights or identities. Thus, we assumed:

H1

Comments scoring higher on (a) Toxicity, (b) Insult, and (c) Profanity are more likely to be categorized as uncivil by human annotators.

By contrast, intolerance entails expressing harmful, discriminatory, or exclusionary intent toward groups or viewpoints, often by denigrating someone on the basis of their identity (e.g., race, religion, or gender), by rejecting participatory rights of others (Lugosi-Schimpf & Thorlakson, 2021; Rossini, 2022), or through threatening language potentially posing a risk to the safety, dignity, or participation of the target in public discourse (Bianchi et al., 2022; Rossini, 2022). Recent research on toxic speech further distinguishes these forms of discourse from incivility by conceptualizing identity-based attacks and threats as manifestations of intolerance that go beyond tone-based norm violations (Pradel et al., 2024). Intolerance undermines democratic discourse not only by silencing targeted others but also by constraining expression among the general public by creating an atmosphere where self-censorship is common, conformity is encouraged, and the discursive space for pluralism and disagreement shrinks (Gibson, 1992). Compared to incivility, intolerance typically yields more harmful effects on audience responses and discourse quality (Kümpel & Unkel, 2023; Lu & Liang, 2024) and is more often found in homogeneous discussions about minorities and civil society (Rossini, 2022). Complementing this perspective, computational studies show that identity-targeted attacks and threat-related language constitute a distinct class of harmful discourse oriented toward group-based exclusion rather than general interpersonal disrespect (ElSherief et al., 2018). We hypothesize that the Perspective’s Severe Toxicity (which may partially overlap with incivility but increasingly reflects hostility that borders on intolerance), Identity Attack, and Threat scores most directly capture this discriminatory intent:

H2

Comments scoring higher on (a) Severe Toxicity, (b) Identity Attack, and (c) Threat are more likely to be categorized as intolerant by human annotators.

Making Digital Hate Classification More Meaningful

Lastly, the automated classification of digital hate has become a major focus in computational social sciences and linguistics, serving both analytical and content moderation purposes. Although complex machine- and deep learning models used still lack interpretability and reliability (Sahin et al., 2025), extensive model evaluations for text-as-data applications are often neglected, especially in the computational social sciences (Bernhard-Harrer et al., 2025).

Previous research has not only shown mixed results concerning the performance of automated systems for digital hate classification (e.g., Sharif et al., 2024; Weber et al., 2025) but also high dependency on the dataset, even when using the same model workflow, suggesting that there is not the one model that fits best for all cases (Malik et al., 2023). This is particularly relevant for latent constructs such as incivility and intolerance, where meaning varies across communicative settings and requires theoretically grounded and context-sensitive approaches to text analysis (Muddiman et al., 2019). Another problem arises through overfitting (i.e., an algorithm fits too closely to its training data), making reported evaluation metrics questionable, and leading to poor generalizability (Arango et al., 2022). Classification models are usually evaluated using standard performance metrics from computational sciences (i.e., precision, recall, accuracy, and F1 score; Gordon et al., 2021), while standard social science content analysis metrics for reliability (i.e., extent of agreement beyond what would be expected by chance among coders or between coders and a standard), such as the more conservative Krippendorff’s Alpha (Hayes & Krippendorff, 2007), are often neglected, even though such evaluations are important for robust applications (e.g., Kralj Novak et al., 2022; Oh et al., 2023). This also holds true for classifications based on the Perspective API scores (e.g., Gervais et al., 2025; Rosenblatt et al., 2022). Additionally, there is only a very limited number of automatic classifiers available that distinguish between incivility and intolerance (Oh & Downey, 2025). Besides not publicly available deep learning approaches (e.g., F1 scores ∼.70; Bianchi et al. [2022]) and niche domain-specific lexicon-based approaches (Oh & Downey, 2025; Oh et al., 2023) with arguably little generalizability, to the best of our knowledge at the time of writing, only a single deep learning-based classifier is available that showcases improvable performance (i.e., F1_incivility: .70, F1_intolerance: .59; Pendzel et al., 2024). This lack of potentially broadly applicable and interpretable tools underscores the need to examine whether existing linguistic and toxic-content indicators, such as those predicted by the Perspective API, can be systematically used to both predict nuanced forms of digital hate and evaluate the performance of other, less interpretable classification systems. Rather than aiming to replace high-performing deep learning models, this study proposes a complementary approach that structures existing model outputs in a theoretically meaningful way. By combining multiple Perspective API attributes, we move from treating Toxicity as a single proxy toward modeling higher-order constructs as configurations of lower-level signals. While these scores are themselves generated by opaque systems, their combination enables a more differentiated and theory-guided approximation of how distinct linguistic dimensions contribute to classification outcomes. Building on these limitations, we examine whether multiple Perspective API attributes can be combined to predict incivility and intolerance as nuanced forms of digital hate and to provide a structured basis for evaluating classification outcomes. We ask:

RQ1: How well can the linguistic and toxic-content variables predict whether a message is classified as (a) non-hateful, (b) uncivil, or (c) intolerant?

Data Collection

Data collection took place each day between September 23 and November 31, 2024, following a positive ethical screening from the Institutional Review Board of the Department of Communication, University of Vienna (Approval No. 1159). Using YouTube’s Developer API and a hashtag-based search strategy, three key hashtags (i.e., “#Israel,” “#Gaza,” and “#Hamas”) were selected to broadly capture videos related to the Israel-Hamas war, which was selected as a prototypically polarized discourse context, as such environments are more likely to generate substantial amounts of hateful content. For each day in the sampling period, up to 90 topically relevant videos from the DACH region (i.e., Germany, Austria, and Switzerland) were retrieved, including solely videos posted within the last 5 days of the collection day to minimize moderation impacts. From each video, up to the first 100 comments posted within the preceding 3 days were collected. To avoid duplication, videos and comments were sampled only once. Comments were automatically language-filtered (Honnibal & Montani, 2017) to retain only English and German-language content. ² Restricting the sample to comments posted within a narrow time frame was intended to reduce the potential influence of content moderation practices on data validity.

The English-language comments were then pre-processed by removing metadata, duplicates, extra white spaces, line breaks, user mentions, and URLs, resulting in 38,311 comments from 950 videos. User mentions (e.g., @username) and URLs were automatically removed using Regular Expressions to reduce potential identifiability. However, we did not apply full named-entity recognition or automated removal of all potentially identifying information within the text to retain linguistic information (e.g., see Franzke et al., 2020). Subsequently, a random sample of 4,000 comments was drawn from this corpus, which served as the basis for our analyses. All data collection and analysis scripts are available at: https://osf.io/svn98.

Data Annotation

For this random sample, each comment was manually classified as intolerant, uncivil, or non-hateful (Rossini, 2022) by a single trained researcher with prior expertise in digital hate. Comments containing discriminatory, dehumanizing, or threatening content were coded as intolerant, if no intolerance was present, comments containing insults, profanity, or violations of discourse norms were coded as uncivil (Rossini, 2022). All remaining comments were coded as non-hateful. The unit of analysis was the individual comment. Coding was conducted following a carefully constructed codebook which included definitions, instructions, and examples of each of the three categories (see OSF Appendix; categorical coding). For an initial validation of the codebook, an automatic annotation approach was conducted using a pre-trained RoBERTa-based model (Table 1; Pendzel et al., 2024), yielding moderate agreement with the manual classification (Krippendorff’s α = .394). Cases without initial agreement between manual and automated coding were subsequently reviewed by the primary annotator and adjudicated using the codebook definitions and examples to assign a final baseline classification. Based on this adjudicated dataset, classification performance was as follows: non-hate (precision _non = .783, recall _non = .865, F1 _non = .822), incivility (precision _inc = .721, recall _inc = .551, F1 _inc = .625), intolerance (precision _int = .518, recall _int = .543, F1 _int = .530), with an overall accuracy of .728 (overall α = .512). The final classification frequencies were non-hate (n = 2,271), incivility (n = 1,132), and intolerance (n = 597; see Table 1).

OPEN IN VIEWER

Table 1.

Corpus Classification

Comments
Type	N manual (%)	N classifier; Pendzel et al. (2024) (%)
Non-hate	2,271 (56.8%)	2,509 (62.7%)
Incivility	1,132 (28.3%)	865 (21.6%)
Intolerance	597 (14.9%)	626 (14.9%)

To assess intercoder reliability, a second manual trained annotator with expertise on the topic of digital hate, independently coded a subsample of the data. An intial pilot sample (n = 50) resulted in low agreement (Krippendorff’s α = .292; Hayes & Krippendorff, 2007), reflecting conceptual difficulty of distinguishing between categories, as constructs such as incivility are inherently context-dependent and require theoretically grounded coding procedures (Muddiman et al., 2019). Following refinement of the codebook, reliability was reassessed on a stratified random subsample to achieve a more balanced representation of the three classification categories (n = 119; n _non-hate = 56 [47.1%], n _incivility = 31 [26.1%], n _intolerance = 32 [26.9%]), resulting in substantial agreement (α = .774). The subsample size was selected as a feasible yet sufficiently heterogeneous subset of the data to capture variation across categories. All annotation procedures followed established principles of content analysis, including codebook development, iterative reliability assessment, and transparent documentation of coding decisions (Neuendorf, 2017).

Lastly, Perspective API scores (i.e., Toxicity, Severe Toxicity, Identity Attack, Insult, Profanity, Threat; probability score ranges from 0–1) were computed for each comment (Table 2; Lees et al., 2022).

OPEN IN VIEWER

Table 2.

Perspective API Scores

Type	ToxicityM (SD)	Severe ToxicityM (SD)	Identity AttackM (SD)	InsultM (SD)	ProfanityM (SD)	ThreatM (SD)
Non-hate	.123 (.114)	.006 (.012)	.073 (.106)	.047 (.079)	.033 (.045)	.034 (.065)
Incivility	.338 (.201)	.031 (.068)	.146 (.153)	.354 (.230)	.131 (.149)	.058 (.100)
Intolerance	.394 (.179)	.062 (.101)	.353 (.191)	.271 (.199)	.155 (.138)	.084 (.120)

Statistical Analysis

For data analysis, we adapted multiple methods, including generalized linear (GLM) and additive models (GAM). We report average marginal effects (i.e., average change in the predicted outcome probability for a one-unit increase in a predictor [AME]; Howell-Moroney [2024]) and pseudo-R ² values for GLMs (Veall & Zimmermann, 1996). Raw beta coefficients are reported in the Appendix. For GAMs, we report the pseudo-R ² values as described in Wood (2017). To compute a threshold for converting obtained probabilities to class labels, ROC curve analysis (Hand & Till, 2001) was used. For evaluating classification performance, we applied accuracy, precision, recall, F1 score, and Krippendorff’s alpha coefficients. All model outputs are available on OSF.

Why Toxicity is Not Enough

Our findings provide nuanced evidence for links between the Perspective API scores and incivility and intolerance in the context of the Israel-Hamas war. Consistent with H1, Insult and Profanity predicted incivility, which is in line with the notion that such language features primarily disrupt norms of respectful interpersonal discourse without necessarily targeting specific groups or posing a threat to democracy (Bianchi et al., 2022; Culpeper, 1996; Papacharissi, 2004; Rossini, 2022). Contrary to our assumption, Threat was also more strongly tied to incivility than intolerance (in contrast to Rossini [2022] and Bianchi et al. [2022] who understood it as a particularly severe form of intolerance and also Pradel et al. [2024] who even introduced it as a distinct category of digital hate), possibly indicating that the Perspective API may be detecting personal or episodic aggression rather than intense individual- or group-directed violations against physical integrity norms here. Alternatively, this pattern could also reflect data limitations, such as class imbalances or artifacts of content moderation practices that reduce the visibility of overt threats, which is why this unexpected finding should not be over-interpreted prematurely (see Table 2). Then again, Identity Attack emerged as a strong positive predictor of intolerance, supporting H2 and underscoring that intolerance is primarily characterized by discourse targeting groups due to their respectively assigned social identities (Bianchi et al., 2022; Lugosi-Schimpf & Thorlakson, 2021; Rossini, 2022). By comparison, the lack of a significant effect for Severe Toxicity may stem from its emphasis on intensity rather than targeted hostility, combined with its relative rarity in the data (see Table 2). Taken together, these results demonstrate that incivility and intolerance can be broadly differentiated based on multiple probability scores of linguistic features provided by the Perspective API in this discursive context, with incivility linked to insults, profanity, and possibly threats, and intolerance distinguished most clearly by identity-based attacks.

Importantly, the prior, virtually sole, focus on Toxicity in the literature (Gervais et al., 2025; Hede et al., 2021; Kim et al., 2021; Oh & Downey, 2025; Schmidt et al., 2024; TeBlunthuis et al., 2024) may thus have obscured important conceptual variation across different forms of harmful discourse (Rossini, 2022). This simplification is not merely technical but conceptual. As incivility and intolerance are not equivalent, conflating them can over-penalize abrasive but non-prejudicial expressions and under-detect polite yet discriminatory messages. Accordingly, models should differentiate between these two types of discourse. Practically, this enables targeted responses (i.e., de-escalation for incivility, protective measures for intolerance) and clearer accountability in research and policy by reporting per-class performance. Future research should therefore refine models that integrate various indicators to enable more accurate and theoretically grounded classifications of online hate (e.g., Bianchi et al., 2022; Matthes et al., 2025; Oh & Downey, 2025).

Perspective API as a Low-Threshold Alternative

Our hierarchical classification approach showed both potentials and limitations when using indicator-based computational models for distinguishing non-hateful expressions from incivility and intolerance in the context of the Israel-Hamas war. While GAMs slightly outperformed GLMs (i.e., likely due to GAMs’ ability to capture non-linear relationships between API scores and classification outcomes, whereas GLMs assume linear effects on the log-odds), overall performance remained somewhat below that of a pre-trained state-of-the-art deep learning classifier (Pendzel et al., 2024). At the same time, our models yielded comparable reliability together with clear advantages in terms of accessibility, interpretability, and domain independence. Unlike training resource-intensive, often language-specific, and opaque deep learning systems by oneself (Sahin et al., 2025), generalized linear and additive models can be readily deployed without specialized machine learning expertise or additional training. This makes them, after being further evaluated, potentially valuable practical tools for multilingual, topic-independent research on digital hate, even if they cannot fully substitute more complex model architectures.

Importantly, our results also highlight the continued necessity of human involvement. As argued by Google Jigsaw (2025c) and Gervais et al. (2025), automatic systems cannot (yet) replace careful human annotation. A potentially promising avenue here lies in integrating these models into human-in-the-loop workflows, for instance, using Perspective API scores to (partially) pre-filter large (possibly even multilingual) datasets before human annotation. Such a socio-technical approach would minimize annotators’ exposure to harmful content (particularly given well-known challenges to their well-being and health; Matthes et al., 2025), reduce costs, and still preserve the nuanced judgments required for building variation-rich digital hate corpora (ElSherief et al., 2021; Kirchmair et al., 2024).

New Evaluation Methods Are Necessary

Pendzel et al.’s (2024) pre-trained classifier served in this study as the state-of-the-art reference for how multiple Perspective API probability scores can serve as lower-level constructs for assessing more complex black-box models. In the hate vs. non-hate comparison, the negative impact of Severe Toxicity on intolerance suggests that the underlying training data in the present study may contain only a few highly severe cases (likely reflecting content moderation practices that make such content difficult to collect; e.g., ElSherief et al. [2021]). On the contrary, when distinguishing incivility from intolerance, the results more or less aligned with our theoretically derived profiles: Insult and Profanity (as well as, unexpectedly Threat) were characteristic of incivility, while Identity Attack consistently predicted intolerance. This match suggests a potential for indicator-based Perspective API probability scores to not only be able to approximate but also offer a structured, theory-guided perspective on how different lower-level indicators relate to these outcomes. Importantly, such evaluation moves beyond conventional metrics of model performance (e.g., accuracy, F1 scores; Gordon et al. [2021]) by uncovering which lower-level linguistic features drive higher-order classifications and why.

As emphasized in prior work (e.g., Bernhard-Harrer et al., 2025; Grimmer & Stewart, 2013), continuous methodological innovation is needed to ensure robust and interpretable model evaluation. However, our findings also caution that evaluation through API scores is not without limitations given that they may reproduce biases present in the original data (e.g., Nogara et al., 2024) and require further validation before they can be considered reliable standards, especially across languages and topics. Future work should therefore refine these approaches, triangulating API-based evaluations with human annotation and alternative computational methods to more fully assess the validity of higher-order black-box classifiers in capturing multiple forms of digital hate.

Limitations

First, our analysis is based exclusively on English-language data within a single, highly polarized context (i.e., the Israel-Hamas war), which raises concerns about topic and language bias. Specifically, linguistic markers of incivility and intolerance may vary substantially across cultural and linguistic contexts (i.e., particularly in less identity-based or less conflictual domains), potentially limiting the generalizability of our findings (e.g., Arango et al., 2022; Nogara et al., 2024). As a result, researchers seeking to apply this approach to other domains should account for context-specific linguistic and discursive patterns and are encouraged to validate and, if necessary, recalibrate the model within their respective settings. Second, while we focused on key Perspective API scores, not all available attributes were examined due to issues of statistical collinearity and conceptual redundancies. Additionally, potential interaction effects were not modeled, obscuring more complex data patterns. Third, our operationalization of incivility and intolerance treats them as distinct categories, yet in practice these forms of harmful discourse can overlap, meaning that some conceptually ambivalent cases may not be cleanly captured by our typology. Fourth, the full corpus of 4,000 comments was coded by a single primary annotator, whereas a second annotator independently coded only a small reliability subsample to assess interrater reliability. Finally, our sample size inherently restricts statistical power and model robustness, a deliberate trade-off to prioritize expert annotation. To mitigate this, we use out-of-sample validation and report robust average marginal effects.