UbuntuGuard: A Culturally-Grounded Policy Benchmark for Equitable AI Safety in African Languages

Western-centric AI safety models fail African language speakers: Current safety guardrails are optimized for high-resource languages and Western cultural contexts, leaving speakers of African languages vulnerable to harms that existing systems don't recognize or prevent.

English-language safety benchmarks create a false sense of security: The paper found that models performing well on English safety tests significantly underperform when evaluated in African languages, meaning organizations relying on standard English benchmarks are likely overestimating their actual safety across global populations.

Cultural context matters for defining "harm": UbuntuGuard's 155 domain experts from sensitive fields like healthcare identified culturally specific risks and safety concerns that rigid, predefined safety categories miss—demonstrating that one-size-fits-all safety policies cannot adequately protect diverse communities.

Current multilingual AI models have significant gaps: Even models designed to handle multiple languages show only "partial but insufficient coverage" when transferred across languages, meaning deploying the same guardrail model globally without localization leaves gaps in protection.

Dynamic, policy-based approaches show promise but need improvement: While newer models that can adapt safety policies at runtime perform better than static ones, they still struggle to properly understand and apply culturally grounded safety standards in African language contexts, indicating this is an active area requiring continued development.