CS Paper Radar

"You" Is Not Toxic: How LLM Labelers Rewrite the Web's Conversation Laws

Liang Yue, Eric A. Vance

cs.SI cs.CL ·submitted 2026-10-16 13:29 UTC · v1
theorycode/data

Abstract

What the Web community believes about toxic conversations, such as that toxicity rises in longer and deeper threads, was measured mostly with Google's Perspective API, which is shutting down; large language models (LLMs) will make the next measurements. We find that LLM labelers mistake second-person address for toxicity, and that this one bias manufactures laws that human labels do not show. On 1.25 million news comments with crowd labels, Claude Haiku flags 12% of the non-toxic comments that contain no "you" and 51% of those with four or more; a single crowd rater flags 6% and 10%. Deeper replies address someone more often, so under Haiku toxicity rises with reply depth (+0.93 percentage points per level), while under the crowd's labels it slightly declines (-0.12). Rewriting "you" in the third person removes two thirds of the false rise; flagging only as many comments as the crowd does not remove it, and showing Haiku the parent comment enlarges it. All seven LLMs we test over-flag "you", and at the crowd's flag rate, without the parent, any false rise of the open ones vanishes at a fixed number of second-person words; Haiku also over-flags "you" on Reddit, Gab and Wikipedia talk pages. Agreement with human labels does not reveal the bias: we prove that a fixed Cohen's kappa leaves a law nearly unconstrained. We therefore certify the law instead of the labeler. Threadsafe uses the labeler's scores to choose which comments human raters label and to predict the others, and corrects the predictions with the human labels; linear laws are estimated without bias whatever the labeler does. In simulated audits its intervals for the laws keep nominal coverage from 250 labels on, where uniform-sample methods such as PPI++ fall short, and at 1,000 to 8,000 labels are 0.75 times as wide as those of design-based supervised learning (DSL); a thousand human labels rule out Haiku's rising law.

Comment
43 pages, 4 figures, 44 tables. Labels, prompts and code: https://github.com/liyu2195/you-is-not-toxic

Closest papers

Nearest neighbours by meaning across the whole index.

Evaluating LLMs as Judges of Counterspeech Quality Against Online Hate Speech

cs.CL ·2026-11-03 ·Johnny Alali, Nicolás Benjamín Ocampo, Davide Ceolin
similarity 0.66
Abstract

Prolonged exposure to hate speech on social media can cause lasting psychological distress among users. Counterspeech has emerged as an effective, non-censorial mitigation strategy, but its deployment requires rigorous quality assessment. Manual assessment is unscalable due to the large volume of content on the Web and imposes severe emotional strain on human annotators. Large Language Models (LLMs) offer a scalable alternative, yet concerns remain regarding their reliability. In this paper, we investigate the extent to which LLMs can effectively perform this task. We test 12 open-weight LLMs of different sizes and families under zero-shot, few-shot, and chain-of-thought prompting, to approximate pointwise and pairwise human expert assessment of counterspeech quality across four dimensions: relevance, coherence, aggressiveness, and suitableness. We find that (i) LLMs show low agreement and moderate rank correlations with human experts: our best configuration, Qwen 2.5 72B on coherence, achieves up to a linear-weighted Cohen's $κ_l = 0.19$ and Spearman's $ρ= 0.42$ in pointwise evaluation and an unweighted Cohen's $κ= 0.25$ in pairwise evaluation. (ii) Converting pointwise scores into pairwise labels substantially improves human--LLM agreement for selected configurations. For same-instance coherence, Qwen 2.5 72B under zero-shot prompting achieves $κ= 0.35$ with converted labels, compared with $κ= 0.25$ when judging pairs directly, and Mixtral 8$\times$7B under few-shot prompting improves from $κ= 0.07$ to $κ= 0.34$. (iii) LLMs agree more with one another than with human experts, particularly in pairwise evaluation, where the strongest model pair, OLMo 2 32B and Llama 3.3 70B on coherence, reaches substantial agreement with $κ= 0.76$, whereas their agreement with the experts in the same configuration does not exceed $κ= 0.21$.

0.21

How User-AI Mistreatment Occurs and Matters in Conversational Systems?

cs.AI cs.CLcs.CYcs.HC ·2026-09-11 ·Fanqi Zeng, Sadid A. Hasan, Chaocheng He
similarity 0.62
Abstract · Accepted by the 10th Workshop on Online Abuse and Harms (WOAH 2026) at EMNLP 2026

Safety research often focuses on model-generated harms, but users may also direct hostility, coercion, and adversarial pressure at models. Understanding how and when that occurs is essential for accurately interpreting model behaviour, alignment drift, and real-world deployment risks. In this paper, we audit 777K English LMSYS-Chat-1M conversations with two independent detectors: an eight-category lexicon for hostility directed at the model, and the dataset's moderation signal; and show that they capture different, weakly overlapping phenomena. The lexicon identifies insults, threats, and jailbreak coercion aimed at the assistant, while moderation flags are dominated by toxic-content solicitation rather than hostility at the model. Together, they mark about 5% of user turns; adjusting the narrower lexicon-harassment union for measured precision puts mistreatment aimed at the assistant at 0.90%. These absolute rates describe arena-style evaluation traffic and should not be read as deployment-wide base rates. We find that user hostility varies 13-fold across models, driven largely by who each model attracts rather than by model behaviour: first-turn hostility spreads far wider than post-response hostility, and more than fifteenfold separates the extremes even after deduplicating opening prompts. Within conversations, assistant apologies are consistently associated with higher odds of next-turn hostility under both detectors; the effect survives restricting to non-refused prior turns and to jailbreak-free conversations, and is positive in 20 of 23 models. Yet across models, more apologetic models receive less hostility overall. Finally, hostility also shows temporal structure, with coercive openings front-loading the first turn while affective hostility accumulates over a session. We release the lexicon, the detector cross-validation pipeline, and all derived tables.

0.13

Using LLMs to Mimic the Conversational Dynamics of Reddit Communities

cs.HC ·2026-08-29 ·Vedaant Jain, Yoshee Jain, Ishq Gupta, Aditi Shrivastava, Koustuv Saha +1 more
similarity 0.61
Abstract · 5 pages, 2 tables, 1 figure

Online communities face a constant battle against toxicity and misinformation. While human moderators struggle to keep pace with the volume of content, LLMs offer a promising solution for automatically generating constructive responses and shaping online interactions. This paper preliminarily investigates if LLMs can mimic the communication styles of Reddit users using their comment history as context. We evaluate two prompting approaches: predicting a target comment and filling in masked comments. We find that LLMs outperform expectations at replicating comment structure and formality, but struggle to accurately capture nuanced emotions, e.g. understating joy and overstating anger. These findings highlight a promising direction for LLMs in guiding online conversations towards prosociality influencing emergent communication patterns and norms within the community. The results of our study inspire future work with more rigorous methods of evaluation to explore the LLMs' effectiveness across diverse online communities to better understand their broader societal impact.

0.25

"Shut Up and Let Me Enjoy My Otome": Understanding and Measuring the Toxicity in Otome Game Communities

cs.CR cs.SI ·2026-09-07 ·Yage Zhang, Xinyue Shen, Yukun Jiang, Michael Backes, Yang Zhang
top venue similarity 0.60
Abstract · Accepted by the ACM Conference on Computer and Communications Security (CCS) 2026

Otome games, a romance simulation genre primarily targeting female, have emerged as a major force in the global gaming market, attracting hundreds of millions of players and billions in revenue. Despite their popularity, otome game communities face pervasive online toxicity, which has been largely unexplored. In this work, we present the first large-scale measurement of toxicity in otome game communities across social platforms. We introduce OtomeSCAN, a framework for collecting, evaluating, and analyzing 620,045 posts from Weibo and Reddit spanning 18 months. To support robust analysis, we manually annotated a ground-truth dataset of 4,308 posts, identifying eight target groups such as players and game developers. We evaluate seven toxicity detectors on the dataset, including general-purpose models and our proposed LLM-based detectors, with our best model achieving F1-scores of 0.82 (Weibo) and 0.78 (Reddit). Our analysis reveals significant platform-based differences in toxicity: 22.20% of otome-related posts on Weibo are toxic, compared to 3.71% on Reddit. Besides, real-world events like in-community conflicts can rapidly escalate toxicity, with toxicity ratios increasing to 37.09% in just 72 hours during an external attack on Weibo. We also flag 191 potential-coordination clusters in otome game communities, 64.40% of which target game developers, with several accounts participating repeatedly across multiple clusters. We hope our work inspires further research on community-specific toxicity and contributes to building healthier online spaces for marginalized gaming communities.

0.41

Where Does Rhetorical Distortion Live? A Multi-Platform, Multi-Language Measurement of Influencer Content

cs.SI cs.CLcs.CY ·2026-10-19 ·Xucheng Yu, Jiangping Chen, Haohan Wang
similarity 0.59
Abstract · 12 pages

Much of the manipulation in influencer content is not false but disproportionate, framing accurate facts as historic, urgent, or catastrophic. Existing approaches either verify factual claims or detect clickbait and propaganda in single-platform, single-language text; however, they overlook where such framing originates, because they treat a post and the replies beneath it as one unit. To address this gap, we measure rhetorical distortion with one instrument applied identically across six platforms and two languages, classifying creator posts and their replies separately. Our three-tier pipeline detects five literature-grounded dimensions (Significance Inflation, Anxiety Manufacturing, Novelty Claims, Loaded Language, and Temporal Distortion): bilingual lexical rules propose candidates, negative filters cancel known false positives, and an LLM verifies or relabels the candidates. This design keeps labels transparent and comparable across platforms, calls the LLM on only 14% of posts, and supports validation against human annotators and a second model. On 13,060 posts from 330 accounts on RSS/newsletters, YouTube, Bluesky, Reddit, Weibo, and X/Twitter, platform is far more strongly associated with account-level distortion than content category (partial $η^2=0.49$ vs. $0.09$), and X/Twitter scores highest (DI 7.74 vs. 0.78 for RSS). On X/Twitter, the top-5 replies under a post carry 3.5x the distortion of the tweet they answer (1.5x at equal length), while creator tweets alone are indistinguishable from YouTube and Bluesky; two human annotators and GPT-4o confirm this gap. Measured distortion on threaded platforms thus lives largely in the audience layer, which calls for measuring creator and reply text separately.

0.24

Reliable but Design-Sensitive: Instrument Uncertainty in LLM Annotation

cs.CL stat.ME ·2026-09-22 ·Thomas Reiter, Christoph Kern, Fedor Miasnikov, Sofiia Nikolenko, Rob Chew +2 more
similarity 0.58
Abstract · Accepted to "3rd Workshop on Uncertainty-Aware NLP" @ EMNLP 2026 (archival)

Large language models (LLMs) can give reliable labels under one setup yet change those labels when researchers make other reasonable design choices. We tested seven LLMs, 12 task designs, three independent runs, and 3,000 tweets labeled for offensive language and hate speech. Repeating the same model and task design produced high agreement (median Fleiss' $κ= 0.91$). Agreement fell when we changed the task design for the same tweets (median Cohen's $κ= 0.76$). Task design and model choice increased the variance of estimated prevalence by factors of 76.7 for offensive language and 110.6 for hate speech compared with sampling variance alone. Variation across LLM task designs reached 560-572 basis points, compared with 270-331 basis points across five human instrument versions. Confidence scores did not solve this problem. They tracked repeated model outputs more closely than agreement with human labels, and grouping six tweets in one prompt lowered mean offensive-language confidence by 660 basis points. We call the variation caused by task design and model choice instrument uncertainty. Researchers can measure it only by comparing reasonable task designs. Repeating one setup or relying on confidence scores cannot replace that test.

0.23

VARM-Bench: Benchmarking Verifiable Structured Reasoning in Chinese Abusive Speech Moderation

cs.AI ·2026-08-16 ·Mingyu Yuan, Shengtao Wen, Lingbing Guo, Zhen Bi, Xiang Chen
benchmark similarity 0.58
Abstract

The widespread circulation of abusive online content has increased the need for reliable moderation of Chinese social-media text. Existing Chinese benchmarks support label classification, fine-grained toxicity categorization, and target-aware extraction, but do not provide a unified representation for deterministically verifying the stated basis of a moderation decision. We introduce VARM-Bench, a benchmark for field-anchored chain-of-thought rationales in Chinese abusive-speech moderation. Each instance contains a concise natural-language rationale with explicit anchors for six decisions: target, target type, target explicitness, author stance, harmfulness label, and fine-grained category. Our deterministic protocol evaluates field correctness, target alignment, output validity, complete-record agreement, and hidden record errors conditioned on correct final decisions, without relying on an LLM judge. Under a common structured-output protocol, we evaluate language models across multiple model families using zero-shot prompting, taxonomy guidance, and structured CoT supervision, and analyze lexical-cue sensitivity and field-level errors. Results show that strong label-level performance can conceal substantial errors in complete moderation records. VARM-Bench provides an auditable and reproducible benchmark for evaluating verifiable moderation rationales in Chinese abusive-speech moderation.

0.04

Do LLMs Change Their Minds Like Humans? Diagnosing Human--LLM Divergence in Single-Turn Persuasion Judgments

cs.CY cs.CL ·2026-08-30 ·Lin Chen, Yitong Chen, Yong Li
top venuecode/data similarity 0.57
Abstract

Large language models (LLMs) are increasingly deployed as proxies for human participants in social simulations, yet whether they update their beliefs in response to persuasive arguments, as humans do, remains poorly understood. We conduct a systematic comparison using a naturally occurring online persuasion corpus in which original posters explicitly verify whether a reply changed their view. Our results show that LLMs achieve only slight agreement with humans (Cohen's kappa ranging from 0.079 to 0.178). Content-level analyses show that humans and LLMs agree on the strongest persuasion cues but diverge on finer ones: humans are more swayed by novel content and assertive language, whereas LLMs favor topical similarity and surface-level formatting. At the level of persuasion strategy, LLMs underweight emotional appeals and overweight credibility signals relative to humans, while the type of proposition under debate exerts no measurable effect on the degree of divergence. Furthermore, switching from first-person role-playing to third-person observation shifts all models toward greater resistance to persuasion, with the effect varying across persuasion strategies and textual features. These findings highlight the risk of treating LLM judgments as faithful proxies for human belief updating and point to structural differences in how LLMs and humans process persuasive discourse. Our code is available at https://github.com/tsinghua-fib-lab/LLM-belief-update-cmv.

0.51