Large language models perform markedly worse in most African languages than in English, French or Mandarin. The reason is not that Yoruba, Amharic, Swahili or Wolof are intrinsically harder to model. It is that these systems learn from text collected at scale from the internet, and the internet contains comparatively little text in those languages.
How the shortage compounds
A language with limited digital text produces a weak model. A weak model makes poor tools. Poor tools discourage people from writing in the language digitally, and often push them towards writing in a language the tools handle well. That produces less text, which produces a weaker model. The loop closes.
Breaking it requires deliberately created data rather than more scraping, because the scraping has already been done and this is the result.
Tokenisation is a quieter tax
Before a model sees text it is broken into tokens, and the vocabulary used for that is fitted mostly to high-resource languages. Text in an under-represented language tends to fragment into many more tokens than an equivalent English sentence. That means a longer sequence for the same meaning, which means more compute, a smaller effective context window, and — where the service charges per token — a higher price for the same task.
Speakers of under-represented languages therefore pay more for worse results. It is a structural cost that no amount of prompting fixes.
What actually helps
Open, permissively licensed corpora built by people who speak the languages; evaluation benchmarks that measure whether a system is genuinely useful for real tasks rather than translating a test set; and tokenisers fitted with these languages in mind. Several research communities on the continent have been doing exactly this work for years, largely without the attention given to model releases.
The gap will close where communities build the data. It will not close on its own.