A health AI that performs impressively in English and poorly in everything else is not a partially solved problem. For most of the world it is an unsolved one, and the gap falls precisely on the people who have the least access to a clinician to ask instead.
Three compounding reasons
- The evaluation is in English. What gets measured gets optimised. Almost every public benchmark for clinical reasoning is English, so that is what systems are tuned to do well.
- The documents are mixed. In much of the world a report is printed in English while the conversation about it happens in another language entirely. The system must handle both at once, which is a harder problem than either alone.
- Translation is not the answer. Translating an English answer afterwards produces text that is grammatically fine and tonally wrong — register, idiom and the words people actually use for parts of the body do not survive the round trip.
What good looks like instead
Answering in the language of the question from the start, rather than answering in English and converting. Treating a clinical term that arrives in English inside another sentence as normal rather than as an error. And testing in the languages people actually use, which means accepting worse benchmark numbers in exchange for a product that works.
Why this is a commercial argument, not only a fairness one
The markets where health information is least accessible are also where the population is largest and growing fastest. A product that works properly in those languages is not a charitable extension of one built for English. It is a different product, and whoever builds it well has very little competition.
This article explains how our technology works. It is health information, not medical advice.