AI safety datasets fail non-English languages, study finds

A new arXiv paper (arXiv:2608.13695) titled 'Language-Specific Gaps in AI Safety Training Datasets' exposes critical deficiencies in how AI models are trained to handle non-English languages safely. Led by researchers Chialuka Prisca-Mary Onuoha, Bright Etornam Sunu, and Rashidat Sikiru, the study audited 21 safety datasets across 25 language slices, including low-resource Hausa, mid-resource Swahili, and high-resource French.

The audit revealed that safety claims often collapse at the language-specific level. For example, a Hausa-language slice failed its own translation-quality acceptance threshold, while Swahili passed comfortably in the same pipeline—proving these gaps are addressable but persistently overlooked. Critically, self-harm and sexual-content categories had *no native-language coverage* in either African language tier, a gap that defies resource-level explanations. The findings align with observed asymmetries in multilingual jailbreak robustness, where single-turn attacks are mitigated but multi-turn attacks remain effective, particularly in underrepresented languages.