The language map of frontier AI

The public conversation about frontier AI treats the models as globally competent. They are not, yet. Every major frontier model was trained on a corpus that is heavily weighted toward English, and the resulting distortion is now the field's most persistent open problem. A model that is world-class in English can be conspicuously worse in Arabic. It can be conspicuously worse in Bengali, in Vietnamese, in Bahasa Indonesia, in Urdu, in Swahili. The gap is not evenly distributed, and it does not close on its own.

That distortion is what the labs are now actively working to close, and the professionals doing the closing are multilingual specialists in emerging markets.

Why English-weighted training is the problem

Training a language model on the internet gives you a model that is best at whatever the internet is best at, and the internet is disproportionately English. This is true in raw volume. It is more true when you filter for high-quality text: technical writing, professional documentation, and scholarly literature. A physicist in Iran, an accountant in Egypt, a physician in Bangladesh, and a lawyer in Kenya are all reading English-language training data as their reference corpus, but the frontier's coverage of their specific working language is much thinner.

For deployment, this becomes a real problem. A hospital in Cairo cannot use a medical AI that reasons well in English but produces unusable Arabic. A law firm in Karachi cannot rely on a model whose Urdu comprehension of a statute drops off at the second clause. The commercial ceiling on frontier AI is limited by every non-English market it cannot serve at native quality.

The correction is happening in the field

The response, across every major lab, has been the same. Contract native-speaking domain experts to evaluate model outputs in the target language, to author high-quality training data in that language, and to adversarially test the model against realistic prompts. The programs are running now, at scale, in Arabic, Hindi, Urdu, Bengali, Vietnamese, Indonesian, Tagalog, Swahili, Turkish, Persian, Portuguese, Russian, and Ukrainian, among others.

The specialists doing this work are not translators in the traditional sense. They are working professionals in medicine, law, engineering, research, and adjacent fields who happen to be native in a target language. They know the difference between a technically correct translation and one a working professional would actually write. Frontier labs are paying for that difference.

What this means if you are one of those specialists

Two things.

First, the compensation for language-specific expert work is often higher than the equivalent English-only work, because the supply is thinner. This is not always advertised. It shows up when you apply for a role that is nominally in one language and turns out to be in your working pair.

Second, the current cohort of specialists doing this work has an unusual leverage moment. The models being trained today will define the deployed AI landscape of the next decade. If Arabic, Bengali, or Bahasa Indonesia AI is going to work well at scale, the training data being written now is what will determine that. The people writing it get to shape what "correct" means in their language.

Where to start

If you are a working professional whose native or working language is not English, and you have credentialed expertise in medicine, law, engineering, research, teaching, or a similar field, this market is open to you.

Sign up, upload a CV that makes both your field and your working languages visible, and see what the current catalog surfaces. Country-specific guides live at the blog.