TY  - RPRT
        AU  - Dang, Hai-Anh H
        TI  - Count the Poor in AI’s Language Benchmarks
        PY  - 2026/Oct/
        PB  - Institute of Labor Economics (IZA)
        CY  - Bonn
        T2  - IZA Discussion Paper
        IS  - 19029
        UR  - https://www.iza.org/publications/dp19029
        AB  - A new coalition aims to let 3.4 billion speakers of underrepresented languages use AI in their own language and voice within five years, and it promises benchmarks to track progress. In an illustration with one large multilingual benchmark, weighting regions by where the world’s extreme poor live, rather than by population, lowers scores by 4 to 8 percentage points for leading closed models and by about 14 for the best evaluated open-weight model. Many African languages also still need more tokens for the same text, which raises the cost of serving them. This essay suggests that the coalition report whether people in poverty can complete useful tasks safely, in a language they use, at an affordable cost.
        KW  - generative AI
KW  - language technology
KW  - benchmarks
KW  - poverty
KW  - poorer countries
        ER  -