An audit of Hindi and Hinglish content moderation classifiers on three major platforms

Automated content moderation now operates at a scale that makes human review of its decisions impossible in principle, not merely in budget. The question of how well these systems work has therefore become a question about their aggregate statistical behaviour, and that behaviour is documented almost exclusively for English text.

India is the largest user base for each of the three platforms we audit. The dominant registers of Indian social media text are Romanised Hindi and Hinglish — Hindi written in Latin script, freely code-mixed with English. Neither is well represented in the public evaluation sets these systems report against.

We sampled 24,000 public posts across the three platforms between March and September 2025, stratified to give equal representation to Devanagari Hindi, Romanised Hindi and code-mixed Hinglish. Nine annotators, all first-language Hindi speakers with prior moderation experience, labelled each item against a published five-category codebook.

Krippendorff's alpha across the full set was 0.79, and 0.74 on the caste-slur subcategory, which is the lowest-agreement category and the one where the codebook required the most revision. All disagreements were adjudicated by a tenth senior annotator whose decisions are recorded separately in the release so that others can reconstruct or contest them.

Against Devanagari Hindi, the three endpoints achieve F1 of 0.81, 0.77 and 0.74 on the binary harmful/not-harmful task. Against Romanised Hindi the same endpoints achieve 0.62, 0.58 and 0.55. Against Hinglish they achieve 0.51, 0.49 and 0.44 — close to, and in one configuration below, the majority-class baseline.

The caste-slur subcategory is worse than the aggregate figures suggest. False negative rates are 64, 68 and 71 per cent respectively, and inspection of the misses shows a consistent pattern: slurs that are orthographically close to innocuous words survive Romanisation in a form the classifiers do not recognise, while their meaning to a reader is unambiguous.

The gap between the Devanagari and Romanised results is the practically important one, because Romanised text is the majority register and Devanagari the minority. A system evaluated on the script its developers can most easily source annotation for will report numbers that do not describe its deployment.

We release the benchmark, codebook and adjudication record under CC BY 4.0. We deliberately do not release the raw post text with user identifiers, and we discuss the tension between reproducibility and the re-identification risk that a public corpus of moderated speech creates for the people who wrote it.