paperAugust 2026Unreviewed

Language-Specific Gaps in AI Safety Training Datasets

Chialuka Prisca-Mary Onuoha, Bright Etornam Sunu, Rashidat Sikiru

Abstract

Large language model providers routinely cite multilingual safety benchmarks spanning a dozen or more languages as evidence that their models are safe for non-English-speaking users. We show that these collection-level coverage claims frequently do not survive inspection at the level of an individual language. Auditing 21 resources across 25 language slices, of which 20 count as datasets under our counting rules, spanning three languages chosen to represent low- (Hausa), mid- (Swahili), and high

Cite

@misc{onuoha2026languagespecific,
  title = {{Language-Specific Gaps in AI Safety Training Datasets}},
  author = {Chialuka Prisca-Mary Onuoha and Bright Etornam Sunu and Rashidat Sikiru},
  year = {2026},
  month = aug,
  eprint = {2608.13695},
  archivePrefix = {arXiv},
  url = {https://www.semanticscholar.org/paper/e49b7db925cfbc19b9d81d2927d28c9670abb89b}
}