The stack is English-first, all the way down

Everyone knows the training data leans English. It is the first thing anyone says about language coverage, and it is true. But it is also the least interesting layer of the problem, because it is the one that gets budget. The layers underneath it rarely do.

Five places the assumption hides

• The tokenizer, fitted to Latin script, so other writing systems pay more tokens for the same meaning.
• The evaluation suites, which decide when a model is good enough to ship and are overwhelmingly English.
• The documentation and prompt guides, which teach a style of asking that only exists in English.
• The error messages, which switch to English exactly when a user is most confused.
• The folklore — the tips people trade about what phrasing works — which does not transfer.

Any one of these is a papercut. Together they decide who finds the tool usable. A product can be technically multilingual and still feel like a translated shell over an English application, and users can tell the difference within about thirty seconds.

What we do about it

We treat language as a product surface, not a locale file. That means the interface, the model behaviour and the support all follow the language you arrive in. It means our own evaluations run in the languages we claim to support, not only in English with a translation pass afterwards. And it means the failure cases we chase first are the ones that only show up outside Latin script.

This is slower than shipping an English product with a language switcher. It is also the only version worth building, because the switcher approach fails precisely where it matters — at the moment someone is trying to describe something they care about.

A tool that works best for the users it was tested on will always call everyone else a long tail.

We will get parts of this wrong in languages nobody here speaks. The plan is to be specific in public about what we have tested, so that being wrong is legible rather than quietly absorbed by the people affected.