AI translation: when it is good enough and when it is not

Machine translation is now good enough to read with, good enough to draft with, and not good enough to sign with. For a news article, a menu or an incoming email it beats most people's half remembered school language. For a contract, a dosage instruction, an immigration form or a slogan it fails in ways that look completely fluent on the page. The deciding question is not the tool or even the language pair. It is what a single wrong word costs you, and whether you would notice it.

What changed, and why the jump was so large

Older systems worked in fragments. They matched short phrases against a large table of phrase pairs harvested from documents that already existed in two languages, then stitched the pieces together. Each fragment was defensible, the sentence was not, and the output read like a ransom note.

Neural systems changed the unit of work. The model reads the whole sentence before it produces anything, so it can tell which noun a pronoun points at and which sense of an ambiguous word the context demands. General purpose language models went further, holding a long document in view and keeping terminology consistent from the first page to the last. How a language model actually works covers the prediction step underneath, and that step is what to keep in mind: the system produces the most probable rendering of your text, it does not verify that the rendering carries your meaning. Fluency improved faster than accuracy, and that gap is what makes the remaining mistakes hard to spot.

Where quality still drops

• Low resource languages. Quality tracks how much text in that language the model was trained on. Widely written languages such as Spanish, French, German, Portuguese, Mandarin, Japanese and Arabic do well. Regional and indigenous languages, and many languages of Africa, Central Asia and the Pacific, do noticeably worse, and a pair with no major language in it is worse again because the system effectively pivots through English and loses detail twice. Where training data comes from explains why the coverage map looks the way it does.
• Idiom and wordplay. Puns do not survive. Proverbs come out either painfully literal or swapped for one that carries a different implication.
• Names and job titles. Systems translate what should have been left alone, and transliterate the same name two ways in one document. Company suffixes, court names and government departments have official renderings a general model will not always pick.
• Formality and pronouns. French, German, Spanish, Russian and Japanese all force a choice between formal and familiar address that English never makes. The system guesses, and a guess that is too casual in a business letter reads as rude in a way you will never see in your original.
• Gender. Going from a language with no gendered pronouns into one that has them means inventing information, and the invention tends to follow the stereotype in the training data. It is one of the clearest everyday examples of how AI bias shows up.
• Terms of art. Legal and medical language gives ordinary words fixed technical meanings. Consideration, discharge, negative, relief and party all mean something specific in context, and a general rendering can invert the sense entirely.
• Numbers, dates and units. Decimal commas, thousands separators, day first against month first date order, and metric against imperial units all pass through untouched when they should have been converted or flagged.

All of these fail quietly. A model will drop a subordinate clause, add a qualifier that was not there, or render a guess in confident prose, and the output still reads smoothly. It is the same mechanism behind why AI makes things up, applied to a task where you cannot judge the result by reading it.

How to check a translation you cannot read

Back translation. Translate the result back into your own language using a different system from the one that produced it. Meaning reversals, dropped clauses and invented sentences show up immediately. The different system matters, because a model asked to undo its own work will often repeat its own mistake and hand you a clean result. This check says nothing about tone or naturalness, so treat a clean round trip as a floor rather than a ceiling.

Ask for alternatives and the reasons. Instead of asking for the translation, ask for three renderings of the key sentence with a note on how they differ in formality and emphasis, then ask which words in your original were ambiguous and what was assumed. The answer exposes the decisions you would have wanted to make yourself, such as whether a request reads as polite or peremptory.

Spot check the fixed items. Names, numbers, dates, amounts, product codes and addresses should look identical in both versions. Comparing them needs no language skill, and they are where silent corruption hurts most. The wider habit of checking specific claims instead of the whole output is in how to fact check an AI answer.

If you can find any speaker of the language, even an amateur, sixty seconds of their time beats an hour of your checking. Ask one question: does this sound like something a person would write?

Risk by use case

The same tool is excellent or reckless depending on what you point it at.

The middle rows are the interesting ones. Nobody sends raw machine output to a court. Plenty of people send it to a client.

When a human translator is not optional

Two things push a job over the line. One is accuracy under consequence. The other is accountability: a professional translation comes with a named person who stands behind it, and for official purposes that signature is the actual product.

Many institutions require a certified translation, and that word means different things by country. In the United States it generally means a signed statement of accuracy attached to the translation. In several European countries the work must come from a sworn or court appointed translator on an official register, and nothing else is accepted. In the United Kingdom, certification usually comes from a translator or company confirming accuracy and giving its credentials. Ask the receiving body what it accepts before paying anyone. This is general information rather than advice for your situation.

A middle path is now standard in the industry: machine translate first, then pay a bilingual professional to post edit. You keep most of the speed and most of the safety for less than translating from scratch. Where health, money or law are involved, asking an AI about health, money or law the safe way applies before you act on any translated instruction.

The privacy question when you paste a document

Pasting a document into a translation box sends it to somebody else's computer. Not automatically a problem, but a problem more often than people assume, because the documents worth translating tend to be contracts, medical letters, identity papers and client correspondence.

Three things vary by service and plan: whether the text is stored, for how long, and whether it can be used to improve the model. Consumer and business tiers often differ sharply on the last point, and the default is not always the one you would choose. Where your prompts actually go explains the path between your keyboard and the answer.

Before you paste, ask three questions. Is this mine to share? An employment contract belongs to two parties and a patient letter belongs to the patient. Does my employer have a rule? A workplace AI policy usually names the categories that may never leave the building. Can I remove what identifies people? Swapping names and account numbers for placeholders, translating, then swapping back takes a minute and removes most of the exposure.

Where the text is genuinely sensitive, a model that runs on your own device is a reasonable answer, because nothing is transmitted. Offline quality sits below the best online systems, but for understanding a document rather than publishing one it is often enough. Page translation extensions deserve the same scrutiny: translating a page means sending it, including anything you were signed in to see.

A workflow worth using today

For reading, just read. Accept that idioms will be rough and do not over interpret one strange sentence.

For anything you will send, use four steps. Give the context first: who is writing, who is reading, how formal it should be, what the text is for. Ask for the translation plus a note on the formality choices and any ambiguity it resolved. Back translate with a second system and compare sentence by sentence. Check names, numbers and dates by eye.

For anything binding, hire someone. The test: if you would not be comfortable having the translated version read aloud in a dispute about what you agreed to, it needs a human name attached. Everything short of that is a question of how much checking the stakes deserve, and the checks above take ten minutes.