Fourteen languages, one AI — and why humans still get the last word

One of my websites, its-all-right.org, is a deliberately small, quiet place — and it speaks fourteen languages: German, English, Spanish, French, Polish, Italian, Japanese, Ukrainian, Marshallese, two varieties of Portuguese, Chinese, Hindi and Urdu. Urdu runs right to left, with its own embedded typeface. For a one-person project, that would have been out of reach a few years ago: fourteen professional translations cost more than a site like this will ever earn.

With AI translation, it worked — and worked well. The models hit register and tone with surprising confidence, translate consistently, and keep constraints in mind if you state them, such as: this site consoles, it does not gloss over. Still, this text doesn't end in praise. It ends in a division of labour.

What the native speakers found

For Polish, my friend P.'s mother went through the site page by page, on paper. Her feedback taught me two things. First: some AI sentences were simply off — grammatically awkward, dictionary-assembled. That much I had half expected.

Second, and this I had not expected: her more beautiful version sometimes drifted away from the original. Where the German says "a smile", she suggested "a moment of clarity" — better Polish, different statement. The AI had translated faithfully; the human translated freely. Suddenly the decision was back on my desk: fidelity to the original, or native elegance? That's not a question you can delegate to a machine, because it isn't a translation question at all. It's a question about what the site is supposed to say.

For Ukrainian, an acquaintance — a native speaker — reviewed the full version and praised it outright. That belongs in an honest account too: sometimes the review simply confirms the AI. But I only knew that once a human had checked.

Staying honest where nobody can check

For Marshallese — the language of the Marshall Islands, a small island nation existentially threatened by global warming — I have no one who could review it. The choice was: leave it out, or publish with the cards on the table. So those pages carry a small note, in English: a careful first translation, not yet checked by a native speaker; anyone who can help is warmly invited. Japanese, Chinese and Hindi carry the same note.

Publishing that note feels more exposed than quietly hoping everything is fine. But it's the only stance that fits a site meant to build trust: say what has been verified — and what hasn't.

The system behind it

To keep verified translations verified, there's a technical guardrail: every human-approved version is captured as a snapshot. If a later AI session touches an approved text, an automated comparison flags it — nothing that a person signed off on gets silently overwritten. The AI may add new content at any time; it may not steamroll the reviewed parts.

That pattern generalises far beyond translation: AI covers the surface area, humans check the substance, and a system makes sure human judgement doesn't evaporate in the next automated pass. For how these guardrails look in our workshop more broadly, see Agents on a long leash.

What to take away

If you have AI translate something — a website, a letter, a menu: expect the result to sound good. That is precisely why anything public or important needs a pair of native-speaker eyes. Not because the AI translates badly, but because only humans notice where "correct" and "right" part ways — and because in the end, someone has to decide which of the two the text should serve.