What if medical AI looks unstoppable right up until you change the language? One year into Code & Cure, we pull on an unsettling thread: a model can score around 90% on an English medical exam and then crash to about 55% on the same exam in French, with similar drops across other languages. That should give us pause—healthcare doesn't happen in a single language, and patient safety can't ride on English-only competence.

We dig into why this happens by putting human clinicians next to large language models. A doctor doesn't become "less medical" when they switch to Spanish or French; fluency shapes how smoothly they communicate, not what they know. LLMs work differently. They learn by predicting the next token from the data they see most, so when English dominates training, the patterns—and the medical "knowledge" riding inside them—are strongest in English. In lower-resource languages the patterns are thinner, and the model's apparent reasoning can fall apart even when the question contains everything it needs.

Then we take on the popular fix: machine translation. It sounds straightforward until you look at where it actually breaks—numbers, temporal qualifiers, negation, and culture-bound idioms. "Once a day" becoming "eleven times a day" is not a harmless glitch. We also unpack how common translation metrics can reward surface-level word overlap while missing exactly the meaning errors that matter most at the bedside.

For anyone building or using clinical AI, the takeaway is hard to dodge: if we want medical AI we can trust, multilingual competence can't be an afterthought. A system that's unsafe outside English shouldn't be called general medical intelligence.

References:

When medical AI fails outside English
Li et al.
BMJ Digital Health & AI (2026)

Credits:

Theme music: Nowhere Land, Kevin MacLeod (incompetech.com)
 Licensed under Creative Commons: By Attribution 4.0
 https://creativecommons.org/licenses/by/4.0/