Language-Specific Gaps in AI Safety Training Datasets

Coverage claims don't match reality: Major AI companies claim their safety benchmarks cover a dozen or more languages, but when researchers audited 21 resources across three languages (Hausa, Swahili, and French), they found significant gaps that contradict these broad coverage claims.

African languages are systematically undertrained: The study found that self-harm and sexual-content safety categories have no native-language coverage in either Hausa or Swahili datasets—a complete absence rather than just reduced representation, meaning AI models may be particularly vulnerable to harmful requests in these languages.

Quality problems are fixable, not inevitable: When the same data pipeline was used for both Hausa and Swahili, Hausa fell below quality standards while Swahili passed—demonstrating that language-specific gaps are addressable through better practices, not inherent limitations of low-resource languages.

Safety gaps create real security vulnerabilities: The researchers connected their findings to documented evidence that AI jailbreaks (attempts to make models produce harmful content) remain effective in multi-turn conversations for non-English languages, suggesting that undertrained safety datasets leave non-English speakers at greater risk.

Transparency about limitations matters: Companies should stop making blanket claims about multilingual safety and instead be specific about which languages have robust safety training, allowing users to understand actual protection levels rather than assumed coverage.