The language divide in AI
Low-resource languages are a fast-growing topic in the AI journals, and the digital divide is rising in economics. Language models work worse in languages with little training text, which are spoken mostly in poorer countries. Does that slow AI adoption, and its economic gains, where those languages are spoken?
What prompted this
In the AI journals, "low-resource languages" had 16 papers in July–September 2026 against about 10 expected; in economics, "digital divide" grew too (11 against 4); see the trends page. Since October 2025, all 46 papers with the low-resource-languages keyword are in AI journals, while "digital divide" appears on 70 papers (25 in economics, 23 in finance, 11 in management, 9 in AI). Only 8 of the digital-divide papers carry an AI keyword.
Questions
- How much worse do common language models perform in each language, and does that predict AI use by country once income and internet access are held fixed?
- When a model gets better at a language, do firms and workers who use that language adopt AI faster?
- Do translation layers close the gap, or do they cost accuracy where it matters most, such as in health, law and finance?
Data that could help
Multilingual benchmark scores by model and language, AI usage by country (app downloads, search interest), and data on the languages spoken in each country.
Next step
Match multilingual benchmark scores to countries by main language and correlate them with AI app adoption, controlling for income and internet use.