Fluent isn't safe.
Bantu languages aren't built like English. Meaning is carried by syllables, noun classes, and verb grammar that flat training text quietly throws away. A model can sound fluent and still be unable to do the basic clinical things — and it doesn't know it can't. It answers confidently. In a consultation, confidently wrong is dangerous.
It misnames the body
One English part maps to different terms, plurals, and noun classes per language. Without the inventory, a model guesses — and guesses wrong on the parts a clinician actually names.
It breaks the grammar of a symptom
“It hurts” isn't one phrase — concord changes with the body part's noun class, with number, with negation. Get the agreement wrong and the sentence is no longer the symptom the patient reported.
It misses the code-switch
Real patients start in Bantu and switch to English mid-sentence. Models trained on clean monolingual text don't expect it — and drop exactly the clinical detail that switched.
Standard spelling collapses distinctions the language depends on. The same flat string can carry several different meanings; the syllable — the tone-bearing unit — is exactly what orthography discards. The fix isn't more text. It's the closed, enumerable substrate underneath the text: the legal parts, the agreement rules, the attested ways people actually say things.
This isn't a long-tail edge case. It's hundreds of millions of speakers — ~235M across these 21 languages alone — and the exact place your health AI most needs to be right.