A model can pass frontier safety tests and still fail local users

Researchers interviewed by Rest of World argue that AI safety standards are disproportionately shaped around wealthy countries. Tests may assume reliable connectivity, functioning courts, strong data protection and institutions capable of documenting harm.

Those assumptions leave gaps when models operate in other languages and public-service systems. The article cites research on medical translation in Africa, including Tigrinya translations that confused distinct diseases and rendered antibiotics as insecticides. It also describes the risks of automated identification systems excluding people from essential services.

Elizabeth Orembo of Research ICT Africa distinguishes frontier model risks—such as deception or autonomous cyber capabilities—from deployment harms such as exclusion, surveillance and the inability to seek correction. A system can improve on the first while remaining dangerous on the second.

The argument is not that frontier risks are imaginary. It is that deciding which failures count, whose languages are tested and who can contest an outcome is already a consequential allocation of safety resources.