[arXiv]score: 0.12

Code-Mixed Text Causes 26.5% Decision Flip Rate in Hate Moderation Pipelines

June 5, 2026

Hate moderation systems tuned on clean English show a 26.5% decision flip rate when processing Tamil-English code-mixed inputs, with review burden nearly doubling (13.8% to 29.7%) and non-hate false-flag rate rising from 6.9% to 10.4%. A disagreement-based deferral rule partially mitigates automatic errors. Tamil-only inputs degrade further, pointing to language-coverage gaps rather than code-mixing alone.

HOW THIS AFFECTS YOU

●

builderIf you're running threshold-based content moderation, your clean-English-tuned thresholds will misfire significantly on multilingual or code-mixed inputs — recalibrate per language distribution.

●

policyWorth watching because moderation instability on non-English content creates systematic false-flagging and under-enforcement risks at scale, with direct compliance and fairness implications.

read original ↗arxiv.org

← back to feed