Swahili has upward of eighty million speakers. Hausa, Yoruba, Amharic and Oromo each have tens of millions. By any measure of human use these are major languages, and model performance on them remains substantially behind performance on languages with far fewer speakers but far more digitised text.
The reason is not that these languages are harder. It is that the training corpora are thin, and the reasons they are thin are historical and economic rather than linguistic.
Where the text is not
Web-scale training data is web-scale, which means it inherits whatever the web contains. Publishing in many African languages has been comparatively limited online; a great deal of formal writing happens in colonial languages, and a great deal of communication in local languages is spoken rather than written or happens in places crawlers do not reach.
The result is that a language with a hundred million speakers can contribute a vanishingly small share of a training corpus. No amount of architectural improvement fixes an absence of data.
Benchmarks that measure the wrong thing
Much evaluation for lower-resource languages uses translated versions of English benchmarks. This introduces a specific problem: the model is partly being assessed on how well it handles translationese, and questions rooted in one cultural context do not transfer.
A question about a legal procedure, a food, or a social relationship may have no clean equivalent, and a model that answers it in a way that reflects the source culture rather than the target one is scored as correct. Benchmarks built natively, by speakers, testing knowledge that matters to those speakers, produce a different and more useful ranking. There are far fewer of them because building them is slow work that funders rarely support.
The tokenisation tax
Tokenisers fitted mostly on high-resource languages fragment other languages into many more pieces per word. The consequences are direct: a longer sequence for the same content, so more compute per query, a smaller effective context window, and a higher price for anyone using a metered API.
Speakers of under-represented languages therefore pay more for worse output, which is a straightforwardly regressive property of how these systems are currently built. It is also one of the more tractable problems in the list.
What would actually change it
The interventions are known and none of them require a research breakthrough. Digitise and openly license existing text — archives, broadcast transcripts, government publications, educational material. Fund native benchmark construction by speakers. Train tokenisers on genuinely multilingual data. Support the regional research groups already doing this work rather than routing funding through institutions elsewhere.
All of this is corpus and evaluation work: labour-intensive, unglamorous, and difficult to fund because it produces infrastructure rather than papers. The bottleneck is not ideas. It is that nobody is reliably paid to do it.