arXiv cs.CL· Liang Yue, Eric A. Vance·· 12 小时前AI 评分49
研究揭示 LLM 标注器将第二人称误判为毒性，Claude Haiku 等七款模型均过度标记 "you"
"You" Is Not Toxic: How LLM Labelers Rewrite the Web's Conversation Laws
AI 导读
研究发现 LLM 标注器会把第二人称 "you" 误判为毒性，从而制造出人类标注中不存在的规律。在 125 万条新闻评论上，Claude Haiku 对不含 "you" 的无毒评论标记率为 12%，含四个及以上 "you" 的达 51%，而单个人工标注者分别为 6% 和 10%。
正文 · 原文
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.
| Comments: | 43 pages, 4 figures, 44 tables. Labels, prompts and code: this https URL |
| Subjects: | Social and Information Networks (cs.SI); Computation and Language (cs.CL) |
| Cite as: | arXiv:2610.23446 [cs.SI] |
| (or arXiv:2610.23446v1 [cs.SI] for this version) | |
| https://doi.org/10.48550/arXiv.2610.23446 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Liang Yue [view email]
[v1]
Fri, 16 Oct 2026 13:29:30 UTC (185 KB)
来源：arXiv cs.CL · arxiv.org