Hate Speech Detection

Human disagreement is ubiquitous and well-known in labeling. However, variation in explanations, captured through token-level human rationales, remains far less explored. At the same time, it is unclear how to best evaluate human labels and rationales -- or even how to best aggregate rationales beyond majority vote -- in light of this variation. Yet, rationales may provide additional insights into the richness of human reasoning, that may differ in style, values and interpretations -- especially in subjective NLP tasks like hate speech detection. In this work, we unify diverse models, training strategies, loss functions, and existing evaluation metrics under a single protocol by systematically re-implementing them across different label and rationale representation spaces. Classification metrics are organized around two key properties -- predictive and distributional -- while explainability metrics through three complementary dimensions: plausibility, faithfulness, and complexity. In this unified supervision framework, we evaluate model behavior across classification and explainability metrics, as well as metric sensitivity to the choice of label (hard and soft) and rationale representation space (hard, intermediate and soft). Results show that both hard and soft metrics favor softer representations, highlighting their effectiveness in capturing variation and the need to rethink evaluation in subjective NLP.

Hateful meme detection remains a formidable challenge for vision-language models, as existing benchmarks are structurally observational - confounding rhetorical hate mechanisms with target community features and preventing causal evaluation of model vulnerabilities. To address this, we introduce FBHM, a systematically curated benchmark of Functionality Based Hateful Memes constructed along two orthogonal axes: 25 distinct rhetorical functionalities and 10 target communities (5,000 memes total). Benchmarking state-of-the-art VLMs reveals a severe generalization gap: models highly accurate on standard datasets catastrophically drop to near-random performance on FBHM, proving they exploit dataset-specific heuristics rather than robust multimodal reasoning. To efficiently close this gap, we propose LSV (learnable steering vectors), an ultra-low data regime strategy that applies a causal intervention objective on as few as 500 steering samples (50 unique base memes), boosting FBHM performance by ~30 Macro-F1 points while outperforming in-context learning and PEFT without degrading source-domain performance.

The proliferation of social media platforms and online communities has inadvertently catalyzed the spread of cyberbullying, hate speech, and other forms of online toxicity, making the effective governance of such harm a critical societal and computational challenge. While significant strides have been made in automating content moderation, existing research predominantly treats cyberbullying governance as passive, isolated detection at the post level. This reductionist view overlooks the continuous behavioral dynamics of users, the structural diffusion of toxic events, and the critical need for proactive mitigation. To bridge these gaps, this paper proposes a unified full-lifecycle governance framework that shifts the paradigm of cyberbullying governance from isolated static detection toward integrated, continuous, and proactive moderation. Drawing on cyberbullying research and adjacent fields, we systematically synthesize the state-of-the-art literature across four interconnected stages: (1) Content Identification, (2) User and Behavior Modeling, (3) Diffusion Dynamics and Early Warning, and (4) Intervention and Governance. Furthermore, we review available datasets and evaluation practices, and discuss emerging challenges including multimodality, explainability, algorithmic fairness, and the dual-use risks of generative AI, providing a roadmap for future research toward a safer and more resilient digital ecosystem.

Demographic information is often used to model annotator perspectives in subjective tasks such as hate speech detection, but its benefit is inconsistent: it improves performance in some settings and behaves as noise in others. This paper asks when demographic features help. We analyze demographic gain as a function of both data split properties and modeling frameworks. For data splits, we measure annotator disagreement, namely how often annotators assign different labels to the same example, along with training size and train-test demographic coverage. We find that demographic gains concentrate in regimes with low training disagreement, high test disagreement, fine-grained ambiguity measurement, sufficient training data, and greater demographic overlap. Motivated by these regimes, we introduce a gated demographic residual model that treats demographics as a selective adjustment to text-only predictions. Experiments on MHS and POPQUORN show that this design is effective, especially on high disagreement or low confidence examples. Overall, our results suggest that demographics should not be assumed useful by default; their value depends jointly on the data regime and the modeling framework.

Hate speech annotation is costly, subjective, and prone to annotator disagreement, making large-scale dataset construction challenging. We systematically analyze how well large language models (LLMs) align with human judgments across ten theoretically grounded subjective attributes, such as dehumanization, violence, and sentiment, evaluating both small and large variants of Llama 3.1 and Qwen 2.5. Our analysis reveals a consistent split across all models: behaviorally explicit dimensions (insult, humiliate, attack-defend) correlate strongly with human annotations, while evaluative dimensions (respect, sentiment, hate speech) are systematically inverted. Demographic persona conditioning reduces model confidence without improving alignment. Building on these insights, we propose combining attribute-level LLM predictions via a confidence-weighted Ridge regression to reconstruct continuous hate speech scores from the Measuring Hate Speech corpus, achieving $R^2$ of up to 0.71 and outperforming direct prompting baselines, demonstrating that structured attribute decomposition recovers a richer and more human-aligned signal than end-to-end label prediction alone.

Research on harmful meme detection has garnered significant attention, resulting in the development of numerous datasets and methods. However, progress in detecting Chinese harmful memes lags considerably, primarily due to two challenges: first, accurately assessing a meme's harmfulness depends heavily on understanding deep cultural context; second, many memes are semantically ambiguous, making harmfulness highly subjective. To address these issues, we focus on the interpretable detection of Chinese harmful memes by constructing the first Chinese harmful meme explanation dataset, Ex-ToxiCN-MM. This dataset offers opposing interpretations, categorized as "harmful" and "non-harmful", for each meme, aiming to rigorously evaluate a model's ability to discern and comprehend ambiguous, culturally grounded content. We built a specialized knowledge base of Chinese cultural concepts and offensive vocabulary to supply models with essential prior knowledge (C-HarmKB). To address the ambiguity and lack of background knowledge in meme attribution, we have developed a comprehensive attribution analysis framework, RIKE, which includes an Attribution Knowledge Enhancement module (AKE) and a Relative Intent Reasoning module (RIR). Extensive quantitative and qualitative experiments demonstrate that our method outperforms mainstream baseline models across multiple metrics in the task of attributing harmful memes in Chinese. The code, Ex-ToxiCN-MM dataset, and Chinese Harmful Semantic Knowledge Base (C-HarmKB) involved in this study have been open-sourced at https://github.com/wimiw123/Ex-ToxiCN-MM

In recent years social media has become an increasingly popular tool for communication. People use it to share their ideas, exchange information, and discuss thoughts. Given its prevalence and widespread reach, social media must remain a safe space for people. Content generated on social media can be abusive and it has become increasingly important to detect such content. In this paper, we use a language-based preprocessing and an ensemble of several models and analyze their performance of abusive comment detection. Through extensive experimentation, we propose a pipeline that minimizes the false-positive rate (marking non-abusive as abusive) so that these systems can detect abusive comments without undermining the freedom of expression.

Toxic speech detection has become a crucial challenge in maintaining safe online communication environments. However, existing approaches to toxic speech detection often neglect the contribution of paralinguistic cues, such as emotion, intonation, and speech rate, which are key to detecting speech toxicity. Moreover, current toxic speech datasets are predominantly text-based, limiting the development of models that can capture paralinguistic cues.To address these challenges, we present ToxiAlert-Bench, a large-scale audio dataset comprising over 30,000 audio clips annotated with seven major toxic categories and twenty fine-grained toxic labels. Uniquely, our dataset annotates toxicity sources -- distinguishing between textual content and paralinguistic origins -- for comprehensive toxic speech analysis.Furthermore, we propose a dual-head neural network with a multi-stage training strategy tailored for toxic speech detection. This architecture features two task-specific classification headers: one for identifying the source of sensitivity (textual or paralinguistic), and the other for categorizing the specific toxic type. The training process involves independent head training followed by joint fine-tuning to reduce task interference. To mitigate data class imbalance, we incorporate class-balanced sampling and weighted loss functions.Our experimental results show that leveraging paralinguistic features significantly improves detection performance. Our method consistently outperforms existing baselines across multiple evaluation metrics, with a 21.1% relative improvement in Macro-F1 score and a 13.0% relative gain in accuracy over the strongest baseline, highlighting its enhanced effectiveness and practical applicability.

This paper presents a multi-stage framework for detecting reclaimed slurs in multilingual social media discourse. It addresses the challenge of identifying reclamatory versus non-reclamatory usage of LGBTQ+-related slurs across English, Spanish, and Italian tweets. The framework handles three intertwined methodological challenges like data scarcity, class imbalance, and cross-linguistic variation in sentiment expression. It integrates data-driven model selection via cross-validation, semantic-preserving augmentation through back-translation, inductive transfer learning with dynamic epoch-level undersampling, and domain-specific knowledge injection via masked language modeling. Eight multilingual embedding models were evaluated systematically, with XLM-RoBERTa selected as the foundation model based on macro-averaged F1 score. Data augmentation via GPT-4o-mini back-translation to alternate languages effectively tripled the training corpus while preserving semantic content and class distribution ratios. The framework produces four final runs for the evaluation purposes where RUN 1 is inductive transfer learning with augmentation and undersampling, RUN 2 with masked language modeling pre-training, RUN 3 and RUN 4 are previous predictions refined via language-specific decision thresholds optimized via ROC analysis. Language-specific threshold refinement reveals that optimal decision boundaries vary significantly across languages. This reflects distributional differences in model confidence scores and linguistic variation in reclamatory language usage. The threshold-based optimization yields 2-5% absolute F1 improvement without requiring model retraining. The methodology is fully reproducible, with all code and experimental setup available at https://github.com/rbg-research/MultiPRIDE-Evalita-2026.

Human label variation has been established as a central phenomenon in NLP: the perspectives different annotators have on the same item need to be embraced. Data collection practices thus shifted towards increasing the annotator numbers and releasing disaggregated datasets, harmful language being most resourced due to its high subjectivity. While this resulted in rich information about \textit{who} annotated (sociodemographics, attitudes, etc.), the \textit{what} (e.g., linguistic properties of items), and their interplay has received little attention. We present the first large-scale analysis of four reference datasets for harmful language detection, bringing together annotator characteristics, linguistic properties of the items, and their interactions in a statistically informed picture. We find that interactions are crucial, revealing intersectional effects ignored in previous work, and that a strong role is played by lexical cues and annotator attitudes. Effect patterns, however, vary considerably across datasets. This urges caution about generalization and transferability.

This paper compares a PyCaret AutoML branch and a CNN-BiLSTM branch for binary hate speech detection on Indonesian Twitter using the HS label from the corpus of Ibrohim and Budi. Both branches share the same preprocessing pipeline so that the comparison reflects modelling differences rather than inconsistent data preparation. The conventional branch uses TF-IDF with a lexicon-based abusive-word count, whereas the neural branch learns dense token representations and captures both local phrase patterns and bidirectional context. The benchmark is built from the released 13,130-row annotation table, whose HS label yields a 58:42 class ratio. On the held-out split, CNN-BiLSTM achieves the best result with 83.8% accuracy, 79.8% precision, 82.7% recall, and 81.2% F1-score. Within the PyCaret branch, Random Forest is the strongest conventional model with 77.2% accuracy and 77.0% F1-score. The neural branch therefore improves accuracy by 6.6 points and F1-score by 4.2 points. Exploratory corpus analysis, learning curves, and confusion matrices show that the dataset is short-text, moderately imbalanced, and still difficult because many decisions depend on local lexical cues plus short contextual composition. The study concludes that PyCaret AutoML is an effective conventional benchmarking framework, whereas CNN-BiLSTM is the stronger end model for the reported benchmark setting.

Information disorder is a challenging phenomenon that affects society at large. This phenomenon entails the diffusion of misleading, misinforming, and hateful content online. In different contexts, one aspect of the problem may prevail, but overall, this is a broad problem that requires comprehensive solutions. While each dimension of the problem (hate speech, disinformation, misinformation, etc.) requires in-depth analysis, in this paper, we look into the possibility of argument structure to provide relevant information to link these different areas of the problem. In particular, we focus on the WSF-ARG+ dataset, which consists of white supremacy forum messages annotated in terms of argument structure (premises and conclusion). There, we leverage the checkworthiness and hatefulness annotations of the argument components to obtain insights into the hatefulness of the whole message. Our results show promising insights (up to 96% F1), indicating the possibility of extending this direction in the future to tackle hateful content identification and information disorder countering.

Detecting hate speech in memes is challenging due to their multimodal nature and subtle, culturally grounded cues such as sarcasm and context. While recent vision-language models (VLMs) enable joint reasoning over text and images, end-to-end prompting can be brittle, as a single prediction must resolve target, stance, implicitness, and irony. These challenges are amplified in multilingual settings. We propose a prompted weak supervision (PWS) approach that decomposes meme understanding into targeted, question-based labeling functions with constrained answer options for homophobia and transphobia detection in the LT-EDI 2026 shared task. Using a quantized Qwen3-VLM to extract features by answering targeted questions, our method outperforms direct VLM classification, with substantial gains for Chinese and Hindi, ranking 1st in English, 2nd in Chinese, and 3rd in Hindi. Iterative refinement via error-driven LF expansion and feature pruning reduces redundancy and improves generalization. Our results highlight the effectiveness of prompted weak supervision for multilingual multimodal hate speech detection.

Memes convey meaning through the interaction of visual and textual signals, often combining humor, irony, and offense in subtle ways. Detecting harmful or sensitive content in memes requires accurate modeling of these multimodal cues. Existing CLIP-based approaches rely on static fusion, which struggles to capture fine grained dependencies between modalities. We propose DARC-CLIP, a CLIP-based framework for adaptive multimodal fusion with a hierarchical refinement stack. DARC-CLIP introduces Adaptive Cross-Attention Refiners to for bidirectional information alignment and Dynamic Feature Adapters for task-sensitive signal adaptation. We evaluate DARC-CLIP on the PrideMM benchmark, which includes hate, target, stance, and humor classification, and further test generalization on the CrisisHateMM dataset. DARC-CLIP achieves highly competitive classification accuracy across tasks, with significant gains of +4.18 AUROC and +6.84 F1 in hate detection over the strongest baseline. Ablation studies confirm that ACAR and DFA are the main contributors to these gains. These results show that adaptive cross-signal refinement is an effective strategy for multimodal content analysis in socially sensitive classification.

Reclaimed slur usage is a common and meaningful practice online for many marginalized communities. It serves as a source of solidarity, identity, and shared experience. However, contemporary automated and AI-based moderation tools for online content largely fail to distinguish between reclaimed and hateful uses of slurs, resulting in the suppression of marginalized voices. In this work, we use quantitative and qualitative methods to examine the attitudes of social media users in LGBTQIA+, Black, and women communities around reclaimed slurs targeting our focus groups including the f-word, n-word, and b-word. With social media users from these communities, we collect and analyze an annotated online slur usage corpus. The corpus includes annotators' perceptions of whether an online text containing a slur should be flagged as hate speech, as well as contextual features of the slur usage. Across all communities and annotation questions, we observe low inter-annotator agreement, indicating substantial disagreement among in-group annotators. This is compounded by the fact that, absent clear contextual signals of identity and intent, even in-group members may disagree on how to interpret reclaimed slur usage online. Semi-structured interviews with annotators suggest that differences in lived experience and personal history contribute to this variation as well. We find poor alignment between annotator judgments and automated hate speech assessments produced by Perspective API. We further observe that certain features of a text such as whether the slur usage was derogatory and if the slur was targeted at oneself are more associated with whether annotators report the text as hate speech. Together, these findings highlight the inherent subjectivity and contextual nature of how marginalized communities interpret slurs online.

Hate, derogatory, and offensive speech remains a persistent challenge in online platforms and public discourse. While automated detection systems are widely used, most focus on censorship or removal, raising concerns for transparency and freedom of expression, and limiting opportunities to explain why content is harmful. To address these issues, explanatory approaches have emerged as a promising solution, aiming to make hate speech detection more transparent, accountable, and informative. In this paper, we present a hybrid approach that combines Large Language Models (LLMs) with three newly created and curated vocabularies to detect and explain hate speech in English, French, and Greek. Our system captures both inherently derogatory expressions tied to identity characteristics and direct group-targeted content through two complementary pipelines: one that detects and disambiguates problematic terms using the curated vocabularies, and one that leverages LLMs as context-aware evaluators of group-targeting content. The outputs are fused into grounded explanations that clarify why content is flagged. Human evaluation shows that our hybrid approach is accurate, with high-quality explanations, outperforming LLM-only baselines.

Online hate speech and abusive language pose a growing challenge for content moderation, especially in multilingual settings and for low-resource languages such as Lithuanian. This paper investigates to what extent modern multilingual sentence embedding models can support accurate hate speech detection in Lithuanian, Russian, and English, and how their performance depends on downstream modeling choices and feature dimensionality. We introduce LtHate, a new Lithuanian hate speech corpus derived from news portals and social networks, and benchmark six modern multilingual encoders (potion, gemma, bge, snow, jina, e5) on LtHate, RuToxic, and EnSuperset using a unified Python pipeline. For each embedding, we train both a one class HBOS anomaly detector and a two class CatBoost classifier, with and without principal component analysis (PCA) compression to 64-dimensional feature vectors. Across all datasets, two class supervised models consistently and substantially outperform one class anomaly detection, with the best configurations achieving up to 80.96% accuracy and AUC ROC of 0.887 in Lithuanian (jina), 92.19% accuracy and AUC ROC of 0.978 in Russian (e5), and 77.21% accuracy and AUC ROC of 0.859 in English (e5 with PCA). PCA compression preserves almost all discriminative power in the supervised setting, while showing some negative impact for the unsupervised anomaly detection case. These results demonstrate how modern multilingual sentence embeddings combined with gradient boosted decision trees provide robust soft-computing solutions for multilingual hate speech detection applications.

Annotating data remains a costly bottleneck for supervised NLP. Active learning (AL) reduces the number of human labels needed by selecting only the most informative instances, while instruction-tuned LLMs attack the same bottleneck from the other side, making labels cheap enough to annotate entire corpora. This raises two questions: can LLM labels replace human labels within the AL loop, and does AL remain necessary when entire corpora can be cheaply labeled? We investigate both by training supervised hate speech classifiers on a new dataset of 278K German political TikTok comments, comparing human and LLM annotation under matched conditions. LLM annotation at scale outperforms human-supervised classifiers at roughly one-tenth the cost, for both a closed-source (GPT-5.2) and an open-weight (Qwen3.5-122B-A10B) LLM, and the advantage is robust under soft-label evaluation. It hinges on the annotation interface: only a two-question decomposition mirroring the human annotation task unlocks it. AL provides no reliable advantage over random sampling in our prefiltered pool. Error structure depends on the LLM: only GPT-5.2 matches the human FP/FN balance, while other variants over-flag border-control and economic-competition discourse. Humans remain essential as evaluators; for training labels, the question shifts to which LLM, which interface, and what shape of pool.

Online hate speech is associated with harms ranging from deteriorating mental health to violence, yet how consistently platforms moderate hate, and whether enforcement is feasible at scale, remain poorly understood. We audit hate speech moderation on Twitter (now X) using 540,000 tweets annotated by trained native speakers, representative of a full day on the platform. Five months after posting, 80% of hateful tweets, including violent ones, remained online. Removal was only marginally more likely than for non-hateful tweets, far below scams or adult content, and insensitive to severity and reach. Automated detection could not reliably classify hate but ranked it highly, enabling human triage. Simulating this workflow, current staffing curbed little exposure, yet substantial reductions proved financially feasible, far below applicable regulatory fines. Persistent hate reflects resource allocation, not technical limits.

Hateful content online is often expressed using fact-like, not necessarily correct information, especially in coordinated online harassment campaigns and extremist propaganda. Failing to jointly address hate speech (HS) and misinformation can deepen prejudice, reinforce harmful stereotypes, and expose bystanders to psychological distress, while polluting public debate. Moreover, these messages require more effort from content moderators because they must assess both harmfulness and veracity, i.e., fact-check them. To address this challenge, we release WSF-ARG+, the first dataset which combines hate speech with check-worthiness information. We also introduce a novel LLM-in-the-loop framework to facilitate the annotation of check-worthy claims. We run our framework, testing it with 12 open-weight LLMs of different sizes and architectures. We validate it through extensive human evaluation, and show that our LLM-in-the-loop framework reduces human effort without compromising the annotation quality of the data. Finally, we show that HS messages with check-worthy claims show significantly higher harassment and hate, and that incorporating check-worthiness labels improves LLM-based HS detection up to 0.213 macro-F1 and to 0.154 macro-F1 on average for large models.

In this work, we examine hateful memes from three complementary angles - how to detect them, how to explain their content and how to intervene them before being posted - by applying a range of strategies built on top of generative AI models. To the best of our knowledge, explanation and intervention have typically been studied separately from detection, which does not reflect real-world conditions. Further, since curating large annotated datasets for meme moderation is prohibitively expensive, we propose a novel framework - PEST - that leverages task-specific generative VLMs and the few-shot adaptability of large VLMs to cater to different types of memes. We believe this is the first work focused on generalizable hateful meme moderation under limited data conditions, and has strong potential for deployment in real-world production scenarios. Warning: Contains potentially toxic contents.

The development of robust hate speech detection systems remains limited by the lack of large-scale, fine-grained training data, especially for languages beyond English. Existing corpora typically rely on simplistic toxic and non-toxic labels, and the few that capture hate directed at specific minority groups lack the positive counterexamples required to distinguish genuine hate from mere discussion. In this work, we introduce ToxSyn-PT, the first Portuguese large-scale corpus explicitly designed for multi-label hate speech detection across nine protected minority groups, including the non-toxic counterexamples absent in all other public datasets. Generated via a controllable four-stage pipeline, ToxSyn contains discourse-type annotations to capture rhetorical strategies of toxic/non-toxic language, such as sarcasm, dehumanization, and cultural appreciation. Our experiments reveal a catastrophic, mutual generalization failure compared to existing datasets from social-media domains: models trained on social media struggle to generalize to minority-specific contexts, and vice-versa. This finding indicates they are distinct tasks and exposes summary metrics like Macro F1 can be unreliable indicators of true model behavior, as they completely mask model failure. We publicly release ToxSyn on HuggingFace to support reproducible research on synthetic data generation and benchmark progress in hate-speech detection for low- and mid-resource languages.

This paper investigates the use of an LLM to generate auxiliary background context for social media posts, and explores four methods to incorporate this context into the input of an SBERT-based Hate Speech Detection (HSD) classifier. These are: text concatenation, embedding concatenation, a hierarchical transformer-based fusion, and LLM-driven text enhancement. We evaluate the impact of our context generation and incorporation strategies in a textual setting on the Latent Hatred dataset of implicitly hateful tweets and a multimodal setting on the MAMI dataset of misogynous internet memes. Results are evaluated against a zero-context baseline, two previous approaches based on entity linking, and a zero-shot LLM classifier. Findings indicate that incorporating generated context improves HSD performance by up to 3 and 6 F1 points on textual and multimodal settings respectively, from a zero-context baseline to the highest-performing system, based on embedding concatenation.

It is challenging to detect hate speech in Low Resource Languages (LRLs) because of the absence of annotated data, the informality of its language structure, and the lack of standardized grammar. A good example of such a challenge is Roman Urdu which is broadly used by South Asians on social media and has a high variation while lacking contextually consistent spellings. The objective of this paper is to conduct a comprehensive assessment of Large Language Models (LLMs) for Hate Speech Detection (HSD) in Roman Urdu script and fine-tune these models using the Parameter-Efficient Fine-Tuning (PEFT) method called Low-Rank Adaptation (LoRA). To evaluate zero-shot inference, we benchmarked it against PEFT on different transformer models, including Mistral, LLaMA, Falcon, and multilingual BERT. Experiments are conducted on the PURUTT (Parallel Urdu and Roman Urdu Corpus for Toxic Comments and Transliteration) dataset with over 72,000 annotated comments. The results suggest that zero shot models perform moderately (F1 = 0.56), but updating a small fraction of the model trainable parameters improves the classification performance significantly (F1 > 0.93). Our results have shown that PEFT delivers outstanding performance alongside excellent computational efficiency, making it highly suitable for low-resource language processing tasks.

Large language models (LLMs) are increasingly used for hate speech moderation, often within human--AI workflows in which reviewers provide feedback before a final decision. Such feedback introduces two manipulation directions: whitewashing hateful content as normal and smearing normal content as hateful. This study examines the susceptibility of initially correct model judgments to annotator-style rebuttals and analyzes whether attack effectiveness differs across manipulation directions. We introduce a rejudge protocol that extends direct contradiction with decision-boundary perturbations and adversarial rationales. Experiments with multiple LLMs on two hate speech datasets show that annotator-style rebuttals substantially degrade moderation performance, with stronger effects in multi-turn settings. The results further reveal stable, model-specific asymmetries between whitewashing and smearing across attack configurations, indicating distinct directional vulnerability patterns. Explicit reasoning prompts and defensive instructions reduce these effects but do not eliminate them. These findings highlight the need for direction-aware safeguards and dedicated feedback-robustness evaluation in human--AI moderation workflows.

Urdu, the world's tenth most spoken language with 246 million speakers, remains almost entirely absent from mainstream LLM safety evaluation and nine years of WOAH proceedings. To investigate whether this absence has measurable consequences for content moderation reliability, five large language models, GPT-4o, Claude Sonnet 4.5, Gemini 2.5 Flash, Qwen-2.5, and Llama-3.1, were tested across six datasets spanning Nastaliq Urdu, Roman Urdu, English, and code-switched Urdu-English. Across the five Urdu-script datasets, label instability between original-script and English-translation classification ranged from 15.9% (Gemini 2.5 Flash) to 31.6% (Qwen-2.5), with a 'Missed-in-Urdu' rate, content flagged as harmful in English translation but passed as normal in the original script, ranging from 2.4% to 9.9% (median 4.3%). A complete enumeration of all 205 papers across nine ALW/WOAH editions via the ACL Anthology API confirms zero dedicated Urdu papers across the entire period. Results indicate that current LLMs provide uneven safety assurance across Urdu's script varieties, with smaller open-weight models showing substantially higher instability and missed-harm rates than frontier closed models.