A translation is a register, not a word list
Choose the target language, the register and the kind of text. You get a translation that reads as if it had been written in that language, a more literal version for comparison, the four word choices worth knowing about and the places where the meaning does not survive the move. Names, numbers and units stay exactly as you wrote them.
The AI Translator produces a translation in one of the supported languages, with a register and a text type chosen so it reads as a message, a document or a caption rather than a word-for-word rendering. The result is editable text.
One request, three sentences, all correct
The English source is four words long - \"please send the documents by Friday\" - and it becomes three different German sentences depending on who is reading. All three are grammatically right. Two of them are wrong for the situation.
Wir bitten Sie, die Unterlagen bis Freitag einzureichen.
a letter, a contract, HRBitte reichen Sie die Unterlagen bis Freitag ein.
a good news articleSchick die Unterlagen bitte bis Freitag.
a message to a colleagueSame request, same deadline, three sentences a translator would defend. The words are almost identical - the register is the translation.
This is what a translator is actually paid for, and it is why a translation tool that only takes a language is not finished. \"Sie\" and \"du\", \"vous\" and \"tu\", keigo and plain form: the choice is made by the relationship, not by the sentence, and the wrong choice is the kind of mistake a native reader notices before the meaning.
The vocabulary is the easy part
A dictionary already knows that \"invoice\" is Rechnung. What a dictionary cannot know is that your invoice is a pro-forma document for a customer in another country, that the German word carries a legal expectation about payment terms, and that using it in a marketing email reads as a demand. The word is right; the sentence is wrong.
The same thing happens in the other direction. Turkish has no single word for \"privacy policy\" that carries the same legal weight, Japanese has no natural plural for most nouns, and French punctuation puts a space before a colon. A translation that keeps the source's sentence structure will be understandable everywhere and idiomatic nowhere, which is exactly what a reader means when they say \"it sounds translated\".
So the tool keeps two outputs. The translation is the version to publish. The literal version is the diff: read them side by side and the decisions become visible - where a sentence was split, where a passive became active, where a name was kept because translating it would be wrong.
What is protected, and what is decided
Kept exactly
Names, product names, numbers, dates, units, code, file names and anything in quotation marks. A translated price is a wrong price.
Decided by the register
The form of address, contractions, sentence length and how much is spelled out. This is the control that changes the most.
Decided by the text type
A legal text keeps its repetition because the repetition is the clause. Marketing copy does not. The domain selector sets that expectation.
Flagged, not fixed
Puns, idioms, units without a local equivalent. You get a short note instead of a confident wrong answer you would publish.
Which register belongs to which document
| Document | Register | What changes | What must not |
|---|---|---|---|
| Contract, terms, privacy notice | Formal | address form, complete clauses, no contractions | nothing: the meaning has to be identical |
| Business email to a client | Neutral | shorter sentences, one ask per paragraph | the greeting convention of that country |
| Support reply | Neutral to informal | plain verbs, second person | technical terms the user needs to search for |
| App interface | Informal | labels shorten to the longest language | string length limits |
| Ad or landing page | Neutral | idiom replaced, not translated | claims you did not make in the source |
| Academic abstract | Formal | passive voice, nominal style | terminology that the field has fixed |
Two rows are worth reading twice: interface strings and advertising. Interface text has to fit the space the layout gives it, and German and Finnish are longer than English by a third. Advertising is where a literal translation goes from awkward to illegal, because a claim that is allowed in one market is not allowed in another - and a translator is expected to know which is which.
Getting a translation you can publish
Set the pair and the register
The register is not a style preference. If the reader is a customer and the document is a notice, it is formal.
Paste the whole paragraph
A sentence translated alone loses the referent. Two or three paragraphs give the model the subject of the verbs.
Compare with the literal version
Where the two differ a lot, look again. Usually it is a good decision; occasionally it is a missing fact.
Read the term notes
Four lines about the words where the obvious choice was wrong. That is the part a dictionary cannot give you.
Where machine translation still trips
\"Eventually\" is not eventuell, \"sensible\" is not sensibel, \"gift\" in German is poison. A model with the full sentence usually picks the right sense, a model with a fragment does not - which is the practical argument for pasting paragraphs rather than isolated sentences.
Miles, Fahrenheit, dates and decimal commas. The tool keeps your units on purpose: converting 12 miles into 19.3 kilometres is a decision about the reader, not a translation, and it should be made once, deliberately. If the target audience uses metric, convert after translating, not instead of translating.
The second person is the hardest thing to carry. Turkish has a full honourific system, Japanese has three, and English has one word for all of them. A translation that picks the wrong one is grammatically perfect and socially wrong, which is why the register selector sits next to the language selector rather than in an advanced panel.
Translating text that has to be published
Translation is rewriting in another language
The best translation of a good English paragraph is a good paragraph in German that makes the same point in the same order. That is not what word-for-word produces, and it is why the literal version exists as a separate output rather than as the default: it is a diagnostic, not a deliverable.
A useful habit is to read the translation without looking at the source. If it reads like a text written for the reader, it is done. If it reads like a text that is explaining itself, something in the structure is still English.
Decide the terminology once, then keep it
Every field has words that must be translated the same way every time: \"dashboard\", \"workspace\", \"subscription\", \"onboarding\". Pick the term, write it down, and translate the rest of the product against that list. Readers learn a term from the first screen they see, and a second translation of the same concept reads as a second feature.
The term notes in the result are a good place to start that glossary. Four decisions per text, collected over a project, become the style guide nobody had time to write.
Keep the structure where the structure is the message
A contract's numbered clauses, a policy's defined terms, a form's field labels: these are not prose and should not be rewritten for flow. Translate the words, keep the order, keep the numbering. In marketing and editorial text the opposite is true - the paragraph order is a choice, and a translator who reorders it is doing the job.
Numbers, names and the things that must not move
Product names are usually not translated, but the names of features sometimes are, and the decision belongs to whoever owns the brand. Prices, dates, version numbers and legal references keep their form. A translation that converts a price is a price the customer will quote back at you.
Quotations are the other trap. A quotation attributed to a person has to be the established translation of that quotation if one exists, and if it does not, the original with a note. Retranslating a famous line produces something that reads as a mistake to anyone who knows it.
Review by a native reader, always
Not to check the grammar - the grammar is usually fine. To check the register, the collocations and the parts that are technically correct and say the wrong thing. A native reader catches in ten minutes what no amount of rereading the source will reveal, because the source is not the problem.
For anything legal, medical or financial, the review is not optional and the tool's output is a draft for a professional, not a replacement. A model produces fluent text with confidence, which is exactly the quality that makes an error expensive.
Small texts, in volume
Most translation work is not a document; it is forty interface strings, twelve support macros and a release note. Paste them as a numbered list, translate once, and keep the numbering - it is the cheapest way to turn a series of one-off translations into a consistent vocabulary. When a string comes back longer than the space allows, shorten it in the target language rather than translating the English again.
Five runs a day is enough for a small release. Run the whole batch in one request, so the terminology stays consistent inside the batch, and save the output with the source beside it.
Before you publish the translation
How good is the translation?expand_more
Why is one output called literal?expand_more
Does it keep formatting?expand_more
Can I translate into a language that is not in the list?expand_more
Are my texts stored?expand_more
Translate the page, keep the register
Pick the language, the register and the kind of text, and compare with the literal version.
fact_checkQuick facts
- Translation keeps the meaning, not the word order: a literal rendering of an idiom reads as an error to a native speaker.
- Names, units and dates are the parts a machine translation most often gets wrong, so they are worth checking by hand.
- A language model reads text in tokens, and one token is roughly four characters of English, so a 300-word answer is about 400 tokens.menu_bookOpenAI Cookbook
- The output depends on the prompt: the same request with a different tone or length instruction produces a different text, not a different answer quality.menu_bookOpenAI Cookbook
- A single emoji can be several code points joined together, so “count the characters” gives a different number from “count what you see”.menu_bookUnicode Consortium
helpQuestions people ask
How accurate is AI translation?expand_more
Why does a translation sound unnatural?expand_more
Can I translate a website with this tool?expand_more
hubWhat this page is about
- Large language modelopen_in_new
A model that predicts the next token of a text; the site sends the prompt to a hosted model and returns its answer.
- Unicodeopen_in_new
The standard that gives every character of every script a code point, and defines what counts as one user-perceived character.
library_booksOfficial sources cited on this page
- How to count tokens with tiktokenOpenAI Cookbook
- The Unicode StandardUnicode Consortium
Keep going with the next tool
The tools in this group that do the next piece of the same work, all switched on right now.
Switched-off tools are not listed here: only what works right now.