PaperfootJoin free
arXiv

AI labelers mistook “you” for toxicity in online comments

Claude Haiku flagged 12% of non-toxic news comments without “you” and 51% with four or more, versus a single crowd rater’s 6% and 10%.

Paper by Liang Yue, Eric A. Vance
Large open light blue speech bubbles sit beside smaller gray bubbles covered by dark blue-green squares, with tiny figures and loose squares below.
Preprint: method

Edited by Boris Djordjevic · Paperfoot

Paper

The study tested seven large language models as toxicity labelers, using 1.25 million news comments with crowd labels to examine apparent conversation patterns. All seven over-flagged “you.” Under Claude Haiku’s labels, toxicity rose by 0.93 percentage points for each level of replies; under crowd labels, it declined by 0.12 percentage points. Deeper replies addressed someone more often. Rewriting “you” in the third person removed two thirds of the false rise.

The researchers also developed Threadsafe, a method that used model scores to choose comments for human review and predict labels for the rest, then corrected predictions with human labels. They proved that it estimated straight-line trends without bias regardless of the labeler. In simulated audits, its uncertainty ranges captured the true trends at the intended rate from 250 human labels onward.

Why it matters

Automated labels can create apparent rules about conversations that human labels do not support. Threadsafe’s proof shows how human review can support unbiased estimates of straight-line trends despite a biased labeler.

Caveats

Threadsafe’s audit results came from simulations, while the comment-labeling comparisons used crowd judgments as their reference. The work is an arXiv preprint and has not been peer-reviewed.

In this paper

measuring toxic conversations

Methods
Threadsafe (this paper)PPI++ (baseline)design-based supervised learning (DSL) (baseline)Claude Haiku (evaluated)Perspective API
Datasets
Reddit (evaluated)Gab (evaluated)Wikipedia talk pages (evaluated)

Results

Claude Haiku, flag rate, non-toxic comments that contain no "you": 12%; Claude Haiku, flag rate, non-toxic comments with four or more "you": 51%

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%.

Abstract

Claude Haiku, toxicity rise with reply depth, toxicity change per level of reply depth: +0.93 percentage points per level

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).

Abstract

What it shows

LLM labelers mistake second-person address for toxicity, manufacturing conversation laws that human labels do not show.

We find that LLM labelers mistake second-person address for toxicity, and that this one bias manufactures laws that human labels do not show.

Abstract

A fixed Cohen's kappa leaves a conversational law nearly unconstrained, failing to reveal labeler bias.

Agreement with human labels does not reveal the bias: we prove that a fixed Cohen's kappa leaves a law nearly unconstrained.

Abstract

Threadsafe estimates linear laws without bias regardless of what the labeler does by combining scored human labels with model predictions.

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.

Abstract

A thousand human labels rule out Haiku's rising toxicity law under Threadsafe audits.

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.

Abstract

Quoted word for word from the abstract on arXiv.

The paper

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

Liang Yue, Eric A. Vance

arXiv, Oct 16, 2026, Open access, Preprint, not peer-reviewed

doi.org/10.48550/arxiv.2610.23446

arXiv:2610.23446v1 in cs.SI, also cs.CL

AbstractPDFHTML

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

Cite as BibTeX
@misc{yue2026toxic,
  title = {"You" Is Not Toxic: How LLM Labelers Rewrite the Web's Conversation Laws},
  author = {Yue, Liang and Vance, Eric A.},
  year = {2026},
  eprint = {2610.23446},
  archivePrefix = {arXiv},
  primaryClass = {cs.SI},
  doi = {10.48550/arxiv.2610.23446},
  url = {https://arxiv.org/abs/2610.23446}
}

Related