hate-speech

Researchers introduce a theory-grounded benchmark for automatically detecting mental health stigma in online news and social media, evaluating LLMs and toxicity/sentiment classifiers. Results show neighboring constructs poorly capture stigma and that LLMs overpredict stigma unless given explicit operational rules.

The paper proposes a novel scope-conditioned generation framework that integrates structured stereotype characteristics into Large Language Model prompts for effective multilingual counterspeech, validated on a human-curated dataset with significant improvements in factuality and effectiveness.

A multi-stage agentic framework for generating, refining, and evaluating counter-narratives to combat hate speech and misinformation, with human validation showing improved persuasiveness and safety over expert-written counterspeech.

A US immigration policy targeting noncitizen technology researchers monitoring online hate speech is facing legal challenges, with court rulings sparking appeals and debates over free speech and deportation threats.

Elon Musk criticizes a misleading post by AOC, referencing Grok's analysis that most 2026 cases were officially ruled suicides with no evidence of foul play.

This paper evaluates retrieval-based in-context learning approaches for detecting criminally relevant hate speech in German social media posts, finding that few-shot prompting outperforms zero-shot but retrieval methods offer marginal gains, with models better suited for triage than autonomous moderation.

The paper introduces CHARM, a moral foundation detection framework that integrates psychological constructs for robust performance, showing improved cross-domain generalization and application to COVID-19 discourse.

This paper examines the susceptibility of LLM-based hate speech moderation to annotator-style rebuttals, showing that such attacks degrade performance and reveal directional asymmetries between whitewashing and smearing manipulations.

This paper presents a comparative study of parameter efficient fine-tuning and prompt engineering techniques for hate speech classification in Roman Urdu.

This paper introduces AHA-Memes, the first large-scale Arabic hateful meme benchmark with fine-grained multi-label annotations, covering 5K manually annotated and ~66K silver-labeled memes, and benchmarks various multimodal models for culturally grounded hate detection.

This paper introduces a hierarchical framework for multimodal sexism detection in memes that models annotator disagreement using frozen Gemini Embedding 2 representations and a lightweight gated MLP, achieving 1st place on fine-grained sexism categorization at EXIST 2026.

Introduces a comprehensive hate speech dataset for Turkish and Arabic, and develops state-of-the-art BERT-based models for hate speech analysis including classification, intensity prediction, target identification, and span detection.

This paper finds that 42.6% of annotator disagreement in HateXplain concentrates at the hate/offensive boundary, demonstrating that majority vote silences minority values and leads to models being wrong but highly confident on contested inputs.

This paper studies hate speech cascades on Bluesky and uses multi-LLM agents to simulate them, finding that such simulations reproduce key patterns like stance monoculture and toxicity-delta direction, and that amplifier targeting on dense networks yields 7.5–12.9% reduction in hateful content with low benign collateral.

A new report from the Center for Countering Digital Hate (CCDH) reveals that racist comments targeting politicians tripled after Meta relaxed its content moderation rules, with violent threats and hate speech quadrupling and bullying doubling.

New research from the Center for Countering Digital Hate shows that abusive comments, violent threats, and hate speech against US lawmakers on Facebook tripled or quadrupled in the six months after Meta relaxed its speech rules in early 2025.

Elon Musk highlights Grok's response to a user who copied Gemini's analysis of a Belgian hate speech conviction and asked Grok to reply.

This paper studies the use of large language models to assist expert counterspeech writing when hate speech and misinformation co-occur, testing knowledge-driven strategies with human evaluation. The mixed strategy combining fact-checkers' and NGOs' guidelines proved most effective.

This replication study evaluates DExperts for mitigating toxicity in LLMs, finding near-perfect safety against explicit toxicity but reduced effectiveness against implicit hate speech and a significant latency trade-off.

Researchers from UCLA examine how automated content moderation tools, including Perspective API, fail to distinguish between reclaimed and hateful uses of slurs for LGBTQIA+, Black, and women communities. The study finds low inter-annotator agreement even among in-group members and poor alignment between community judgments and AI moderation tools, highlighting the need for context-sensitive approaches.