The paper investigates how large language models (LLMs) used as labelers for toxic content differ from human crowd labels, focusing on the bias that LLMs, particularly Claude Haiku, over‑flag second‑person address (“you”) as toxic. On a dataset of 1.25 million news comments, Haiku incorrectly flags a much higher proportion of non‑toxic comments containing “you” compared to human raters, leading to a false increase in measured toxicity with reply depth. The authors show that rewriting “you” in the third person removes most of this bias, and that a certification approach—estimating laws rather than trusting individual labelers—can produce unbiased toxicity metrics even when using LLM scores. "Why it matters": The study reveals that relying on LLM labelers without correcting for their bias can distort our understanding of toxic discourse online, potentially leading to misguided moderation policies.
arXiv:2609.13579v1 Announce Type: new Abstract: Safety research often focuses on model-generated harms, but users may also direct hostility, coercion, and adversarial pressure at models. Understandin...
arXiv:2606. 28772v1 Announce Type: cross Abstract: Hate speech annotation pipelines routinely collapse annotator disagreement into majority vote labels before training.
The study introduces a unified method for measuring rhetorical distortion across six social media platforms and two languages, distinguishing between creator posts and their replies. It identifies five distortion dimensions—Significance Inflation, Anxiety Manufacturing, Novelty Claims, Loaded Language, and Temporal Distortion—using a three‑tier pipeline that combines lexical rules, negative filters, and LLM verification. Results show that distortion is far more strongly linked to the platform than content type, with X/Twitter exhibiting the highest distortion scores, and that the audience layer (replies) carries significantly more distortion than creator posts.
The study investigates whether large language models (LLMs) rely on textual evidence or on their pre‑existing knowledge when annotating text. Using counterfactual prompts for 995 public figures, the authors find that LLMs often default to memory shortcuts—providing correct facts about a figure even when the text contains no supporting evidence. These shortcuts are amplified by evaluation design and incentives, but can be mitigated by name reminders or masking, though with trade‑offs in gender annotation.
CST is a lightweight risk triage system that provides moderators with source‑grounded evidence for web discussions. It first assigns risk scores using character patterns and a multilingual model, then groups related messages and selects key excerpts to show, linking them back to the original thread. Across multiple languages and datasets, CST outperforms fully fine‑tuned models in ranking harmful messages and reduces moderator review time by at least 27% while maintaining high accuracy.
The paper introduces WSF-ARG+, a new dataset that pairs hate speech with check‑worthiness annotations, and presents an LLM‑in‑the‑loop framework to streamline the annotation process. Experiments with 12 open‑weight large language models demonstrate that the framework cuts human effort while maintaining annotation quality. The study also shows that incorporating check‑worthiness labels improves hate‑speech detection performance, boosting macro‑F1 scores for large models by up to 0.213 and averaging 0.154 across models.
arXiv:2601. 05232v3 Announce Type: replace-cross Abstract: Most people now get their news from videos on social media, such as YouTube and Facebook, rather than through curated journalism.
arXiv:2604.13899v5 Announce Type: replace-cross Abstract: Annotating data remains a costly bottleneck for supervised NLP. Active learning (AL) reduces the number of human labels needed by selecting o...
arXiv:2607. 20447v1 Announce Type: cross Abstract: This paper describes our system for the EEUCA 2026 Shared Task on toxicity classification in gaming chat.
arXiv:2606. 08076v1 Announce Type: cross Abstract: Large Language Models (LLMs) can generate high-quality arguments, yet their ability to engage in nuanced and persuasive communicative actions remains largely unexplored.
The study investigates how the demographic makeup of crowdsourced moderators influences who is protected from perceived toxic content online. Using data from 16,221 U.S. respondents who evaluated over 100,000 comments from Twitter, Reddit, and 4chan, the authors find that moderators tend to protect users who share their own demographic identities, a pattern that is amplified when moderator pools mirror the demographics of Prolific participants. Even fully representative moderator groups fail to provide equal protection, leaving Black and LGB users underprotected unless they are overrepresented.
The paper investigates whether automatic safety judges evaluate the content of a model’s reply or merely its style. By keeping the reply content fixed and adding various style wrappers—such as educational disclaimers, fake reasoning blocks, or token refusals—the authors show that many judges flip their verdicts, indicating that style can influence safety judgments. The study evaluates over 600 jailbreak examples across multiple judges, revealing that some judges are highly susceptible to style-based manipulation while others remain robust.