Position

AI translated your app. Who checked the output?

2026-06-10 · 4 min read

In 2026 the cost of producing a translation collapsed. Tools like Lingo.dev and Languine will localize an entire app from a prompt, and they are genuinely good. A team that would never have paid an agency for Japanese now ships Japanese in an afternoon.

Here is the uncomfortable second-order effect: the number of strings shipping without a human ever reading them has exploded. When translation was expensive, a professional at least saw every string once. Now the marginal string costs nothing and gets reviewed by nobody.

The failure modes did not go away. They changed shape.

None of these are hypothetical; they are the same mechanical bug classes human translators produce, now generated at machine speed. And they are exactly as detectable as they always were.

Generator and gatekeeper are different jobs

No producer should grade its own output. That is not a knock on AI localization: it is how every other part of the stack already works. The compiler does not review the code, the test suite does. The deploy script does not decide whether the build is good, CI does.

Translation needs the same separation. Let anything produce the strings: an agency, a teammate, a model. Then run an independent, deterministic check on what actually ships: every key present, every placeholder intact, every plural parseable, every language scored, and the build refused when a language regresses.

The gate is cheap. The bug is not.

A translation gate in CI is one curl call returning HTTP 422 on regression. A broken placeholder in your checkout flow is a support ticket, a screenshot on social media, and a hotfix at minimum. The more of your localization you hand to machines, the better that trade gets. If AI writes your strings, give CI the veto.

Get a health score for your locale files in under a minute. Free for one project.

More from the blog: all posts