AI translation has quietly become good enough that the interesting question is no longer "is it accurate" but "when is it accurate enough, and how would I know?"
Where it is genuinely reliable
Understanding something in another language. Reading a foreign email, article or document to know what it says. Near-solved for major languages, and the risk of a small error is usually low.
Internal communication. Team messages, notes, informal exchanges where both sides know it is machine-translated and can ask.
First drafts for a human translator. Faster to edit good machine output than to start blank — this is how much of the professional industry now works.
Getting the gist at volume. Support tickets, reviews, survey responses in many languages.
Where it still goes wrong
Idiom and register. It will translate the words and lose the tone. Formal where you meant warm, or the reverse — and in languages with grammatical formality (German Sie/du, Japanese keigo, French vous/tu) getting this wrong is not a nuance, it is rude.
Marketing copy. Almost never survives translation. Wordplay, rhythm and cultural reference do not transfer, and a literal translation of a good headline is usually a bad headline.
Names, places and products. Frequently translated when they should not be. Always check that your product name survived.
Anything legal, medical or contractual. The consequences of a subtle error are severe and the error is invisible to you.
Less-resourced languages. Quality falls off sharply outside the major ones. The model is equally confident either way, which is the dangerous part.
Ambiguous source text. Machine translation cannot ask what you meant. A human translator would.
Getting better results
Give context, not just text. "This is a support reply to a customer who is frustrated. Keep it warm and apologetic." Context changes register, and register is where machine translation most often fails.
Specify formality explicitly for languages that mark it. "Use formal address."
Say what not to translate. Product names, brand terms, technical vocabulary you have standardised.
Provide a glossary for anything you use repeatedly. Consistency across a document matters more than any individual choice.
Translate whole paragraphs, not sentences. Sentence-by-sentence loses the connective logic and produces subtly disjointed text.
Back-translate to check. Translate your output back to the source language and read it. This catches inversions and dropped negations — the errors that matter most and are hardest to spot.
Dedicated translators vs general models
Dedicated translation services are fast, cheap at volume, and consistent. Best when you need a lot of text moved quickly.
General models (Claude, GPT, Gemini) are better at localisation — adapting tone and register rather than converting words. They will also explain their choices and offer alternatives, which is genuinely useful when you cannot read the output.
For anything a customer reads, the general models usually produce better results because the task is more editorial than mechanical.
The rule for published translation
Never publish a machine translation into a language nobody on your side can read.
Not because it will definitely be wrong, but because you have no way to find out if it is. The failure mode is silent: the text looks fine, reads fluently, and says something slightly off — or occasionally something embarrassing — and the first people to notice are your customers.
Machine translation plus a native-speaker review is a genuinely good workflow. Machine translation alone, published, is a bet you cannot check.
Common questions
Is AI translation accurate? For understanding, generally yes in major languages. For publishing, it needs a native-speaker review.
Is it better than Google Translate? General models are better at tone and context; dedicated services are faster and more consistent at volume. Different tools for different jobs.
Can I use it for business documents? For internal understanding, yes. For contracts or anything legally binding, no.
Which languages work best? Major world languages are strong. Quality falls off sharply for less-resourced ones, and the model is equally confident either way.
How do I check a translation I cannot read? Back-translate it to your language and read the result, then have a native speaker review anything you will publish.
Should I translate my website with AI? As a first pass, yes. Publishing without a native-speaker review is a bet you cannot check.
Does it handle formality correctly? Only if you tell it to. Specify formal or informal address explicitly for languages that mark the distinction.