Translation quality varies enormously by language pair. Assume less reliability the further you get from widely published languages.
Do now: Translate one short message, then translate it back and compare the meaning.
Back-translation catches meaning errors. Only a fluent human catches tone.Most of the world works in more than one language, and machine translation is genuinely transformative for access. It is also uneven — strong between well-resourced languages, much weaker for many African languages, and consistently poor at register, which is what makes a message respectful or rude.
Give it what a human translator would need
Supply audience, purpose, register and any terms that must not change. A message to elders, a message to customers and a message to teenagers are three different translations of the same content, and a model given no audience will pick a neutral default that fits none of them.
Provide a small glossary for terms with agreed local renderings — product names, official titles, technical words your organisation uses consistently. Otherwise the same term will come out three ways across three documents, which quietly erodes trust.
Say what must be preserved exactly: numbers, dates, names, legal phrases, safety instructions. And say what may be adapted: idioms, examples, greetings. Good translation is not word-for-word, but you must decide where it may move.
Audience and register are not decoration. They are the translation.
Back-translate, then have a human read it
Translate back into the original language, ideally with a separate pass, and compare against your source. Meaning drift shows up quickly this way: negations flipped, conditions dropped, a 'may' becoming a 'must'.
Back-translation will not reveal tone problems, because a rude sentence translates back into a polite one. Register, formality and respect need a fluent reader. For anything public, contractual, medical or safety-related, that reader is not optional.
Be especially careful with low-resource languages. Output can be fluent and confidently wrong, because there was less material to learn from and the fluency comes from the same process either way. Fluency is not a reliability signal in any language, and least of all in these.
Multilingual by design, not by translation
For material you will publish in several languages, write the source simply: short sentences, one idea each, few idioms, explicit rather than implied logic. This improves the original and makes every translation better.
Consider code-switching, which is how a great many people actually speak. A message that mixes languages the way your audience does may communicate better than a formally correct version in one language, and a model will not do this unless you ask.
Keep a bilingual glossary as an organisational asset, reviewed by a fluent speaker and updated when a term changes. It is a small file that prevents a large amount of inconsistency.
Write the source to be translated, and every version improves.