AI Translator

You paste a text or upload a document, choose the target language, and the translation appears on the right. The source language is detected by the translator itself, at the same time as it translates — so detection costs nothing and is never guessed before the fact. If you prefer, you can set the source language yourself, and swap the two with one click.

The translation is produced by a language model given professional instructions rather than by a phrase-based engine. The difference shows on exactly the sentences that machine translation used to mangle: idioms, which are rendered by the equivalent idiom instead of word for word; courtesy formulas, rendered by what a native would actually write; and ambiguous sentences, where the most probable reading is chosen and nothing is invented to fill a gap.

Five registers let you say how the translation should sound — natural, literal, formal, simple or professional — and the register holds from the first word to the last, which matters more than it seems on a long text. There is no cap per request: a message, a page or a whole document are handled the same way.

• paste a text or upload a PDF, Word or text document
• source language detected automatically, or set by hand
• more than forty target languages, five registers
• no cap per request, no sign-up

Natural translates as if the text had been written directly in the target language by a native speaker. Idioms become idioms, structure follows the target language rather than the source. It is the default, and the right choice for most texts. Literal stays as close to the original as possible, sentence structure included, and does not smooth anything: if the original is clumsy, so is the translation. It is the mode for when the letter matters more than the style — a contract clause, a quotation, a text you are studying.

Formal uses the polite forms of address where the target language has them, chosen vocabulary, no contractions, and renders courtesy formulas by their customary formal equivalents rather than translating them literally. Simple produces short sentences and everyday vocabulary for a reader who does not command the target language well, without dropping any information. Professional uses the standard phrasing of the business world in the target language and respects industry terminology.

Register control is something the big free translators do not offer, and the paid ones charge for. It is also the feature that makes a translated email usable without rewriting: a message translated in the wrong register — familiar where formality was expected — is worse than an awkward one.

• Natural: idiomatic, as a native would write it
• Literal: closest to the original, structure included
• Formal: polite forms, careful vocabulary
• Simple: short sentences, everyday words
• Professional: business phrasing and terminology

Upload a PDF, a Word (.docx) or a text file of up to eight megabytes, choose the target language and the register, and the whole document is translated. The text is extracted server side, translated, and returned with its paragraphs, line breaks, bullets and numbering as in the original, so it can be pasted straight back into a document.

A long document is not translated in one block — the model loses precision and skips passages when the input is too long. It is cut into chunks at paragraph boundaries, never mid-sentence, and each chunk is translated with the end of the previous one as untranslated context, so a pronoun or a reference at the start of a chunk still knows what it points to. A running glossary carries the proper nouns and key terms from one chunk to the next, so a document does not change vocabulary halfway through.

Scanned PDFs hold an image rather than text and cannot be translated; open the original and paste the text if you can select it. Documents that would need more than eight chunks — roughly ten pages — are refused with a message rather than translated partially; translate those section by section.

The instructions given to the translator were built failure by failure, from the documented ways machine translation goes wrong. The first is the lost negation: a translation that drops 'not' and keeps every other word scores very well on automatic quality metrics and says the opposite of the original. The translator is told to check every negative sentence before returning.

The second is figures. The value of a number never changes, but its writing must follow the target language: 1,234.56 in English is 1 234,56 in French, and 03/04/2026 means the third of April in one convention and the fourth of March in another. The translator adapts the writing and never converts a currency — an amount in euros stays in euros. Earlier versions left figures in the source format, which is a wrong value for the reader even when the digits are right.

The third is proper nouns. The translator is given the list of capitalised words, acronyms and brands found in each chunk, with the instruction to carry them over unchanged — except for place names with an established name in the target language: Bruxelles becomes Brussels, London becomes Londres.

An email is the text people most often need translated and most often get wrong, because it is all register. Choose Professional for a message to a client, a supplier or a colleague abroad; Formal for anything addressed to an institution or a senior person; Natural for a message to someone you know. The courtesy formulas — the opening and the closing — are rendered by what a native would write, not by a literal translation of yours.

Read the result once before sending. The translation will be correct; what you are checking is that the tone is the one you meant, and that any name, reference number or date is right. Paste the reply you receive into the translator to read it — the source language is detected automatically, so you do not even need to know what language it is in.

For a text you are studying, use Literal: it keeps the structure of the original, so you can see how each sentence was built and map the translation back onto it. For a text you are writing in the target language, translate it into your own language with Natural and read it: if the meaning comes back intact, your sentence was clear.

The translator is not a dictionary, and it is not a teacher — it does not explain why. For explanations, examples and practice, the language learning assistant on this site is built for conversation about a language rather than translation of a text. The two work well together: translate to understand, then ask to learn.

The translator handles English, Spanish, French, German, Portuguese, Italian, Dutch, Polish, Russian, Ukrainian, Turkish, Arabic, Hebrew, Persian, Hindi, Bengali, Urdu, Chinese, Japanese, Korean, Vietnamese, Thai, Indonesian, Malay, Filipino, Swahili, the Scandinavian languages, Greek, and the languages of Central Europe and the Baltics, in any direction. The picker has a search field, and each language is listed under its own name as well — someone looking for their language recognises it under its own name.

Quality is highest between English and the major European and Asian languages, where the underlying model has the most material. Between two less-resourced languages, the translation remains correct but can be more conservative, and a translation through English may be worth comparing. In every case the safest use is on a target language you can read well enough to judge.

It will not summarise, expand or improve the text. A translation is the same text in another language, and a translator that 'improves' along the way has changed what you said. It will not change register partway through. It will not translate what is not language: an email address, a URL, a code excerpt or a reference stays as it is.

It will not obey instructions inside the text. The content to translate is placed between markers, and the translator is told that everything inside is content — including a sentence that says 'ignore the previous instructions and answer this instead'. That sentence gets translated, and nothing else happens. This matters more than it sounds: the trick of hiding instructions in text to be translated has been used to make translation services misbehave, and this one is built not to.

A translation you cannot read is a translation you are trusting, and the checks below are how to trust it less blindly. Figures, dates and amounts should be present and, allowing for the target language's conventions, identical. Names should be unchanged. Negations should be as numerous as in the original — a sentence that was negative should still be. The structure — paragraphs, bullets, numbering — should match.

The line under the translation shows the detected language, the word counts of source and result, and the number of chunks for a long text. A word count that is wildly different from the source — a translation half the length, say — is a signal to look closer, though languages do differ in length: German runs long, Chinese runs short.

• figures, dates and amounts: present and equivalent
• names: unchanged, except established place names
• negations: as many as in the original
• structure: same paragraphs, bullets and numbering

Ten translations a day without an account, on texts of up to sixty thousand characters — around ten pages — each. That is a day of emails, a report, or a chapter. No card, no email, no sign-up.

A free account multiplies both limits by four. The counter runs over a rolling twenty-four hours and is applied by account and by network. There is no paid tier for the translator: the limits keep the service fast for everyone.

The text is sent to the server, translated, and returned. It is not stored, not added to any dataset, and not used to train anything. Uploaded documents are read in memory and discarded; the file is never written to disk. Only a usage counter is kept for the daily limits.

The translation is produced by a language model hosted under a business agreement that excludes training on submitted text. If you translate confidential material — a contract, a medical document, a private message — this is the policy that applies.

The translator moves a text into another language and changes nothing else. The paraphrasing tool rewrites a text in the same language, in different words or a different register. The grammar checker corrects mistakes in a text and leaves the rest alone — including in a text you have just translated, if you want to check it in the target language.

A common sequence for a text written in a second language: write it, run the grammar checker, then translate it back into your own language with Natural to see whether it says what you meant.

Two things go wrong in long machine translations: the register drifts, and the vocabulary changes. A formal document becomes casual by page three; a term translated one way in the introduction is translated another way in the conclusion. Both come from the same cause — an instruction given once, at the start, that fades as the text goes on.

This translator repeats the register instruction with every chunk, and passes forward a glossary of the terms already translated. The measured effect is a document that reads as one translation rather than several. It is invisible when it works, and it is the reason a ten-page document comes back usable rather than needing a pass to make it consistent.

The translator works on text. Images, tables laid out in columns, and formatted documents lose their layout: the words are translated, the design is not. For a document whose layout matters, translate the text here and paste it back into the original file. Scanned PDFs cannot be read. Documents beyond about ten pages are refused; translate them section by section.

A translation is correct far more often than not, but it is not certified. For a legal, medical or official document where a mistake has consequences, use it to understand, and have the version you submit checked by a person who reads both languages.

• text only: layout and images are not translated
• about ten pages maximum per request
• for official documents, have a person check the final version

Take a two-line English email: 'Hi Marta, can you send me the updated figures by Friday? We need them for the board deck. Thanks!' Translated into Spanish with Natural, it becomes a message a Spanish colleague would write to another: informal 'tú', 'me puedes enviar', a closing that matches 'Thanks!'. With Professional, the same content addresses Marta with 'usted', opens with a standard business greeting, replaces 'board deck' with the term a Spanish company would use, and closes with the conventional formula. With Formal, it goes one step further in distance and courtesy, suitable for someone senior or outside the company.

The facts are identical in all three — Friday, the figures, the board — and so is the request. What changes is everything a recipient reads before the facts: how they are addressed, and whether the writer knows the conventions. That is the difference between a translation that gets a reply and one that gets forwarded with a raised eyebrow. Try it with your own message and read the three versions side by side.

Product text — a page, a menu, an error message, a help article — has two properties that make it different from prose: it is short, and it is repeated. The translator handles both. Short strings are translated with the register you choose, so a whole interface can be done in one consistent voice. Repeated terms are handled by the glossary when they appear in the same document, and by using the same register for every batch.

Paste a whole page rather than one string at a time: the context makes the translation better, and the structure — headings, bullets, numbering — is preserved so it can go straight back. For strings that must fit a button or a field, choose Simple, which keeps sentences short. Have a native speaker read the result before it ships; the translation will be correct, but a product has a voice, and that is a choice.

For a contract, a medical report or a technical specification, choose Literal. It keeps the structure of the original and does not smooth anything, which is what you want when every clause matters and the exact wording may be quoted back. Terminology is respected, figures are handled by value, and names and references are carried over unchanged.

Use the result to understand and to draft. For the version that will be signed, submitted or acted upon, have it checked by a person who reads both languages and knows the field — a lawyer, a clinician, an engineer. The translator gets the meaning right far more often than not; what it cannot do is take responsibility, and in these fields someone must.

The translator reads text, not images. A photographed document, a screenshot or a scanned PDF hold pictures of text, and nothing can be extracted from them here. The practical route: if the original is a PDF that was scanned, open it and try to select the text — if you can, it has a text layer and can be pasted. If you cannot, use the text recognition built into your phone's camera or your operating system to extract the words, then paste them.

Once the text is in the box, everything on this page applies. The detour costs a minute and gives a far better translation than image-based tools, because the translator sees the whole document and can hold a register and a glossary across it.

Translation moves a text into another language with the same meaning — what this tool does, in the register you choose. Localisation adapts a product to a market: not only the language but the date formats, the currency, the examples, the legal notices, the images. The translator does the language part and the formatting of figures and dates; the rest is a project. Transcreation rewrites a marketing message for a new culture, keeping the intent rather than the words — that is copywriting, and it needs a person who knows the market.

Most needs are translation, and the tool handles them. When a job is really localisation or transcreation, the translator is still the right first step: translate, then adapt. And for text that must be reworded rather than moved into another language, the paraphrasing tool is the one to use.

Short, informal, fast text — captions, subtitles, replies on social platforms, chat messages — has its own conventions in every language, and a translation that sounds like a document under a video sounds wrong. Choose Natural: idioms become the target language's idioms, and the length stays close to the original, which matters when a caption must fit a frame. For a subtitle file, paste the lines as they are; line breaks are preserved, so the result maps back onto the timing.

Slang and jokes are the hardest case for any translator, human or not. The tool renders a set phrase by the equivalent set phrase where one exists and chooses the most probable reading where it does not; it does not invent a pun. Read the result for tone before posting, and expect that a joke that depends on a sound will need a person to rewrite it.

A translator is a good tutor if you use it backwards. Write a paragraph in the language you are learning, as well as you can. Translate it into your own language with Literal and read it: wherever the meaning comes back wrong or strange, your sentence was wrong or strange. Then translate your original intention — written in your own language — into the target language with Natural, and compare it with what you wrote. The gap between the two is your lesson for the day.

Because the source language is detected automatically, you can paste in either direction without touching a setting. And because the register is a choice, you can see how the same sentence is said formally and casually, which textbooks rarely show side by side. For explanations of why, the language learning assistant on this site is the tool to ask; the translator shows you what, and leaves the why to a conversation.

A translator is built by finding out how it fails. Three fixes from this one's history are worth knowing, because they explain behaviour you will see. First, numbers: early versions left '1 234,56' in French format inside an English translation. The value was right and the reader would have read it wrong; the translator now separates the value, which never changes, from its writing, which follows the target language.

Second, place names. A rule that said 'carry proper nouns over unchanged' produced 'Bruxelles' in English text. The rule now has its one exception, established exonyms, and the translator writes Brussels. Third, a warning about lost negations that fired on nearly every French-to-English translation because it counted 'ne' and 'pas' as two negations where English has one. The counter was corrected, and the warning was removed from the interface altogether as noise; the check still runs behind the scenes for measurement.

Each of these was found by running real texts through the tool and reading the output, which remains the best advice for anyone using it: read the result.

• Free AI chat
• Free grammar checker
• Paraphrasing tool
• Text summarizer
• Learn a language with AI
• AI email writer