LLM content moderation fails on reclaimed slurs, audit finds
Source published ·Modelwire updated
Original coverage: arXiv cs.CL ↗·How Modelwire adds context
The development
As platforms migrate from Perspective API to LLM-based content moderation, a new audit exposes a critical blind spot: systems struggle to distinguish reclaimed slurs from slurs used as attacks, often silencing the marginalized communities they're meant to protect. Researchers tested eight LLMs and legacy baselines on IYKYK, a corpus annotated by community members, finding that purpose-built moderation systems perform no better than general models at this nuanced task. The finding signals that encoding policy as natural language prompts, rather than structured rules, creates systematic failure modes in high-stakes moderation decisions where context and speaker identity matter most.
Modelwire’s AI-generated summary of coverage from arXiv cs.CL.
Modelwire analysis
ExplainerOur AI-generated reading of the wider context and the next developments to watch.
The real finding isn't that LLMs struggle with reclaimed slurs (that's somewhat expected). It's that purpose-built moderation APIs perform no better than general models, suggesting the problem isn't model capability but how policy gets translated into prompts rather than rules.
This connects directly to the BMMS framework from earlier this week, which argues that compliance measurement needs to shift from generic safety dimensions toward policy-specific verification. Both papers identify the same underlying problem: encoding rules as natural language instructions creates systematic blind spots that fixed benchmarks don't catch. The content moderation audit also echoes the linguistic profiling work from November 6th, which showed how identical factual content produces different outcomes based on framing. Here, identical speaker intent (reclamation vs. attack) produces different moderation decisions depending on context the model can't reliably extract from a prompt.
If Perspective API or a major platform releases a structured rule-based moderation layer (not a prompt-based one) in the next six months and reports measurably better performance on the IYKYK corpus, that validates the paper's diagnosis. If adoption of LLM-based moderation continues unchanged, it signals the industry is treating this as a research finding rather than a deployment problem.
This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error
Coverage behind this analysis
These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.
·arXiv cs.CL
New framework measures language model compliance against specific policies
Researchers have developed a framework that shifts LLM evaluation from fixed behavioral dimensions toward policy-specific compliance measurement. Rather than assessing hallucination or toxicity in isolation, BMMS takes a rulebook as input and returns a compliance profile, enabling regulated industries to verify whether deployed systems actually follow their operational constraints. The suite introduces severity-weighted scoring and…
MentionsPerspective API · IYKYK · arXiv
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “Policy Over Purpose: Auditing LLM Content Moderation of Reclaimed Slur Usage”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.