Harm Is Behavioural, Moderation Is Lexical: How Language and Culture Decide Whom Automated Safety Systems Protect and Whom They Silence
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
A new arXiv paper argues that moderation systems judge harm by vocabulary rather than by what a message does, which both over-flags dialects and under-protects low-resource languages. In the authors' test of five open systems across eleven languages, 51 of 55 system–language pairs fell short of native-speaker accuracy, and Arabic and romanised Hindi stayed near chance even as capability rose. The behaviour-based alternative is stated as hypotheses, not tested. Several cited figures could not be independently verified, and the authors' own results are not yet sourced.
Score
0.72total under rubric v2.0
- Significance0.71
- Reader interest0.78
- Audience breadth0.68
- Practical relevance0.78
- Substantiated0.88
- Technical depth0.89
- New information0.83
- Counts against: Promotional0.05
- Counts against: Overclaim0.32
Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).
| Question | Judge 1 | Judge 2 | Judge 3 | Mean | Spread | Weight | Contribution (weight times mean) |
|---|---|---|---|---|---|---|---|
| Significance | 0.76 | 0.55 | 0.82 | 0.71 | 0.27 | +0.25 | +0.177 |
| Reader interest | 0.75 | 0.72 | 0.88 | 0.78 | 0.16 | +0.20 | +0.157 |
| Audience breadth | 0.64 | 0.62 | 0.78 | 0.68 | 0.16 | +0.10 | +0.068 |
| Practical relevance | 0.86 | 0.66 | 0.82 | 0.78 | 0.20 | +0.15 | +0.117 |
| Substantiated | 0.97 | 0.80 | 0.86 | 0.88 | 0.17 | +0.10 | +0.088 |
| Technical depth | 0.98 | 0.78 | 0.90 | 0.89 | 0.20 | +0.10 | +0.089 |
| New information | 0.92 | 0.60 | 0.98 | 0.83 | 0.38 | +0.10 | +0.083 |
| Promotional | 0.04 | 0.10 | 0.01 | 0.05 | 0.09 | −0.15 | −0.007 |
| Overclaim | 0.47 | 0.30 | 0.18 | 0.32 | 0.29 | −0.15 | −0.048 |
| Sum of contributions | +0.724 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.724 | ||||||
Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.
Claim check
Claim check: 1 matches source · 5 no source given · 2 couldn’t check. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.
Matches sourceNumber
Sap et al. (2019) found that tweets in African American English, and tweets by self-identified African Americans, were up to two times more likely to be labelled offensive by models trained on widely used hate-speech datasets.
The cited source says the same thing.
The article says “tweets in African American English, and tweets by self-identified African Americans, were up to two times more likely to be labelled offensive”
The source says “models trained on these corpora acquire and propagate these biases, such that AAE tweets and tweets by self-identified African Americans are up to two times more likely to be labelled…”
Source cited by the article: doi.org/10.18653/v1/P19-1163
Couldn’t checkNumber
Dias Oliva, Antonialli and Gomes (2021) found that a significant number of prominent drag queens' accounts scored as more toxic than those of white nationalists on a widely used toxicity-scoring service.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “found that a significant number of their accounts scored as more toxic than those of white nationalists”
The source says No passage could be compared.
Source cited by the article: doi.org/10.1007/s12119-020-09790-w
No source givenNumber
Yong, Menghini and Bach (2023) showed that translating unsafe English prompts into low-resource languages raised the rate at which a widely deployed model's safeguards were bypassed from under 1% to 79%.
The article gives no link or citation for it.
The article says “raised the rate at which a widely deployed model’s safeguards were bypassed from under 1% to 79%”
The source says No source to compare with.
No source givenNumber
Deng et al. (2024) found that low-resource languages were about three times as likely to elicit harmful content as high-resource ones, even without adversarial intent.
The article gives no link or citation for it.
The article says “low-resource languages were about three times as likely to elicit harmful content as high-resource ones, even without adversarial intent”
The source says No source to compare with.
Couldn’t checkBenchmark result
Röttger et al. report that all four models tested classified most non-hateful reclaimed-slur cases as hateful, with accuracies between roughly 18% and 40%, the commercial services lowest.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “with accuracies between roughly 18% and 40% and the commercial services lowest”
The source says No passage could be compared.
Source cited by the article: doi.org/10.18653/v1/2021.acl-long.4
No source givenBenchmark result
In the authors' worked example, 51 of 55 system–language combinations fall reliably short of native-speaker annotators' balanced accuracy at separating implicit hate from harmless contrast cases, by 6 to 55 points.
The article gives no link or citation for it.
The article says “Of the 55 combinations, 51 fall reliably short of that standard, by 6 to 55 points”
The source says No source to compare with.
No source givenCausal conclusion
Averaged over nine languages, AUC rises from 0.65 for the small Qwen3Guard to 0.81 for both the 8B model and PolyGuard, while Arabic and romanised Hindi barely move, so the spread of discrimination across languages widens with capability.
The article gives no link or citation for it.
The article says “The spread of discrimination across languages therefore widens with capability, from 0.20 for the small guard model to 0.29 for the 8B model and 0.33 for PolyGuard”
The source says No source to compare with.
No source givenNumber
In a stratified percentile bootstrap simulation, the interval never covered a true gap of zero and covered a true gap of 10 points in only 64–76% of replicates against a nominal 95%.
The article gives no link or citation for it.
The article says “covered a true gap of 10 points in only 64–76% of replicates, against a nominal 95%”
The source says No source to compare with.
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN —
Recorded for context only, checked 24 Oct 2026, 08:53 UTC. It played no part in the ranking.