Checking text that is not in English

The interface is in English, and that is the whole of it. There was a period when it was offered in fifty other languages, all machine-translated, and enough of them came back wrong that showing them was worse than showing nothing at all. They have been taken out rather than left sitting there half-working, and we would rather say so on this page than let somebody find it out a paragraph into their own language.

The distinction that actually matters is a different one, and it survives the change intact: the language the app is written in and the language of the text you are checking were never the same question. You can read every button here in English and hand us an essay in Urdu, and the second of those is what this page is about.

The honest position

We built and calibrated the detection modules mainly on English, so English is where you can lean on them hardest, and in any other language a mark is better treated as a prompt to go and look than as a finding — the further that language sits from English structurally, the more that applies. And translating a piece into English before you check it will not rescue you either, because translation flattens exactly the irregularities the modules read as a person.

There is a second thing we would rather say plainly, since it affects a lot of the people reading this page. Writing carefully in a language that is not your first tends to be more regular, more formal and more evenly constructed than writing in your first, and every heuristic detector in existence — ours included — reads regularity as a signal. That is a real bias, we know about it, and it is the reason two frontier models now argue over the text rather than the heuristics deciding on their own. It is also why anyone reviewing somebody else's work should be reading the reasoning attached to a mark and not counting the marks. There is more on this at False positives.

What comes back

Marked passages and the reasoning attached to them stay in the language of the text you submitted. We are not going to show you a translation of your own sentence and then ask whether it sounds like you, which is the only other option available and a worse one.

The verdicts themselves, the two labels and the explanations around them, are in English along with the rest of the interface.

Questions

Will the interface be translated again?

Not on any plan we would want to announce. The people who use this have asked for English, and a translation nobody has read is a liability rather than a feature — the last set shipped a German tutorial step that read "Showing Noch lange nicht vorbeik" where it meant "Showing somebody".

Does the detection work on my language at all?

It runs on anything you give it, and the caveat above is the honest one: treat a mark as somewhere to look rather than as a verdict, and read the reasoning rather than the count.

Does writing in English as a second language count against me?

It counts against you in every heuristic detector, ours included, which is exactly why the verdict you see comes from two models arguing rather than from the heuristics on their own.