Which tool to choose
The choice isn't "the best one," it's "the best one for that text." The two philosophies are different, and that's where all the difference lies.
| DeepL | ChatGPT | |
|---|---|---|
| Underlying logic | faithfulness to the text, predictable output | rephrasing and adaptation |
| Same sentence repeated | returns the same translation | may vary from one point to another |
| Idioms | often rendered literally | adapted to the usage of the target language |
| Tone control | formal/informal switch | free requests in conversation |
| When it wins | contracts, manuals, technical texts | newsletters, dialogue, emails meant to persuade |
Go with DeepL when you're translating already-formatted documents or technical and legal texts, where the right term isn't negotiable. It's built for one task only, and on a contract that shows: if a sentence appears twice, it normally returns the same translation. That's the stability you want when every word carries legal weight.
Go with ChatGPT when the text lives on tone: a social post, a dialogue, an email that has to "sound" a certain way. With idioms and cultural references DeepL often translates word for word, while ChatGPT recognizes figurative language and adapts it to a sensible equivalent in the target language, provided you give it context and instructions.
A practical signal: if your text would be read by a lawyer, choose DeepL; if it would be read by a client you need to persuade, choose ChatGPT.
How to do it
Path A — Translating a whole document with DeepL
From browser or desktop app the path is the same:
- Open DeepL and go to the file translation section (the document icon next to the text box).
- Drag in the file. It accepts PDF, .docx and .pptx and keeps the original layout.
- Set the target language. For languages that distinguish register (German, French, Spanish) look for the formal/informal option and choose the tone.
- Start and download the translated file.
Check: open the downloaded document. Tables, bold text and images should be in place. If the layout is broken, the source file was a scanned PDF (an image, not text): convert it to Word first, see below.
The limit to know: the free plan translates up to 1,500 characters at a time in the text box, with a reduced number of documents per month. And free-plan content is used for a limited period to train DeepL's neural networks. For confidential material this is a problem; I'll come back to it in the dedicated section, with the way out.
Path B — Translating with tone control on ChatGPT
The mistake people make with ChatGPT is pasting the text and writing "translate." That gives you a generic translation. The tool's strength is the instruction given beforehand.
The operational syntax:
Translate the text below from Italian to French.
Context: it's an email to an important client who complained about a delay.
Register: formal but warm, no defensive tone.
Keep the paragraph structure and don't add sentences that aren't in the original.
For technical terms, leave the standard French term used in the logistics sector.
If an Italian sentence sounds idiomatic and has no direct equivalent, adapt it
to French usage instead of translating it word for word, and flag in square
brackets your choices where you had to adapt.
Text:
[paste here]
Check: read the square brackets. They show you where the AI interpreted. If a choice doesn't convince you, reply in the same chat ("make sentence X more concise") and regenerate just that one. This in-conversation refinement is the advantage DeepL doesn't have: asking for "a bit more casual" or "more formal" and tweaking until it fits.
A concrete example
Marta runs a small cosmetics e-commerce and has to translate a product description into German. The Italian text says: "Una coccola quotidiana che non pesa sulla pelle" (A daily caress that doesn't weigh on the skin).
First attempt with a "literal" translator: the German comes out with "eine Liebkosung" and a verb that, literally, says the cream "doesn't weigh" — correct in the dictionary, clumsy on a label.
Marta switches to ChatGPT with the prompt above and adds: "German register for cosmetic packaging, light and sensory tone, avoid the literal translation of 'coccola'." The AI returns a phrasing that conveys lightness and daily care with a brand-like expression, and notes in square brackets that it replaced "coccola" with a term more natural in the beauty sector. Marta has a native-speaker acquaintance check it: two words tweaked and the description is online. Ten minutes instead of half a day. The native-speaker confirmation is the step you don't skip, and it's the reason the method worked.
When it does NOT work (and how to fix it)
If ChatGPT invents or adds sentences
ChatGPT is built to generate text, not to translate deterministically: variability is by design, and sometimes that leads it to rephrase in ways that drift from the meaning or structure. The first defense is already in the prompt ("don't add sentences that aren't in the original"). The second: translate in blocks of 4-5 paragraphs and visually compare the number of sentences. If the translation has more, something got in that wasn't there.
If the document is a scanned PDF
DeepL translates text, not images of text. A PDF born from a scan or photo is, to the software, an image. First run text recognition (OCR, the technology that turns the image of a word into editable text): upload the PDF to Google Drive and open it as a Google Doc, or use Word, which imports PDFs and makes them editable. Then give the cleaned-up file to the translator.
If the text is confidential (contracts, client data)
On the free plan your texts may end up in training. The verified route is the paid plan: data is not used to train the AI and texts are deleted right after translation, a condition that makes it suitable for contracts and company documents. DeepL Individual costs $8.74/month billed annually.
If you don't want to pay, the concrete alternative is to anonymize before pasting: replace names, amounts and sensitive data with placeholders ([CLIENT], [AMOUNT], [ADDRESS]), translate, and put the real values back into the result. The text that leaves your computer no longer contains anything identifiable.
If the language is "minor" or niche
On regional slang and languages with little digital data, both tools drop off: the less a language or culture is represented, the more they get wrong. On the major European languages the results are excellent. Fix: for niche languages a native speaker's check isn't optional, it's the mandatory final step. No prompt replaces it.
If technical terms change from one page to the next
ChatGPT may render the same term in two different ways at different points in the text. DeepL solves this with the glossary (a list of terms with a fixed translation that you impose). On ChatGPT you recreate it by hand: at the top of the prompt add a list like "always render 'fattura' as 'invoice', never 'bill'." It's less automatic, but it holds.
A tip from someone who actually uses it
Use the two tools in sequence, not as alternatives. Translate the block with DeepL first to get a precise, aligned base, then paste that result into ChatGPT asking it only to "soften the tone and make it more natural for reader X, without changing the meaning." You get DeepL's faithfulness and ChatGPT's naturalness in the same text. It's the logic agencies apply too: for critical translations they never rely on a single tool and cross-check the results to catch the obvious errors.
Frequently asked questions
How many characters can I translate for free?
On free DeepL there are two distinct limits: the text pasted into the box has a cap per single translation (1,500 characters), while uploaded files are counted separately and in reduced numbers per month. For basic web translation you don't even need an account. For long texts, break them into blocks or switch to ChatGPT, which handles larger inputs in one go.
Do I really have to pay to get a good translation?
No. The quality of the free model is already high. You pay for three concrete reasons: larger volumes, editable layout and — most importantly — the guarantee that texts are neither kept nor used for training. If you translate little and nothing confidential, the free version is enough.
How do I keep the same tone throughout a long document?
On ChatGPT give it a sample: paste a paragraph already written in the right tone first, then ask it to translate the rest aligning with that ("keep the tone of this example"). On DeepL you adjust the tone with the formal/informal switch and with the glossary for recurring terms.
Can AI really replace a human translator?
For internal emails, product descriptions, FAQs and drafts, yes: it's more than enough. But there's a real limit, not a detail. The AI doesn't know whether its translation holds up before a native speaker or whether it introduces an ambiguity at a delicate point in a contract. For legal, medical texts or a public message that can't get it wrong, the AI gives you the draft in a few minutes — the final review is done by a person who knows the language and the context. It's not a rejection of the tool: it's the right way to use it, and it still saves you 80% of the time.