Patrice Sterling

Position: Breaking the Dual Curse of Multilingual AI Requires Socio-Technical Guardrails, Not Post-Hoc Alignment

Large language models are deployed globally as universal systems, yet their safety mechanisms remain English-optimized. This creates a Dual Curse for speakers of low-resource languages: a Harmfulness Curse where harmful content generation rises from 1\% in English to 35\% in languages like Hausa, Ig…