21/07/2026
A few years ago, a health organisation deployed a machine translation tool to localise patient intake forms for a community speaking a low-resource African language. The forms came back looking correct. The QA team reviewed them. Nothing flagged. They shipped. Three months later, a field coordinator — a native speaker — read one of the forms and stopped.
A mistranslation in the consent section had quietly changed the meaning of what patients were agreeing to. Just enough to be wrong in a way that no benchmark system had been designed to catch. The tool had done its job but The subject matter had never been part of its education.
AI translates every language extremely well now, and it always has, but why does it do so well with English? Because English dominated the training data. The models learned from billions of English documents, decades of digitised English content with clean structure and reliable metadata.
Now consider what wasn't there. Hausa. Tigrinya. Uyghur. Kurdish dialects. Welsh. Yoruba. Low-resource languages. Not obscure, not fringe languages spoken by tens of millions of people who were simply not represented in the datasets that taught the machines to translate.
The result is an AI translation landscape that is quietly, systematically better for some populations than others. The harder question is what happens to the languages and communities that weren't included when the models were trained, and what happens when those models are now deployed to serve them anyway?
Underrepresented languages fed into undertrained models produce outputs that look like translations but carry errors that no quality system will catch, because the QA benchmarks were also built on the dominant languages. The error is invisible. The harm is real.
At Loc & More, our position is specific: AI accelerates translation. Human expertise determines whether that translation is accurate, culturally appropriate, subject matter related, and honest about its limits.
For low-resource language pairs, we are explicit with clients about where AI confidence ends and human judgment begins. Because the alternative is shipping translations that appear correct and aren't.
If your language strategy includes markets where low-resource languages are spoken, that distinction matters more than any efficiency metric.