Home / Blog

Translation

Why DeepL misses idiomatic French (and how to catch it)

Charline BrainezCharline Brainez19 October 20265 min read

When a client sends me a machine translation to post-edit, the French usually reads smoothly. That is exactly the problem. A sentence can be grammatical, fluent and confident, and still say something your English never said. So is DeepL good enough for English to French? Often, for a first draft. Here is what it does well, where it slips, and how I catch the slips before your French readers do.

What DeepL does well

Let me be fair first. DeepL is a serious tool, and I work with whatever my clients use. Here is what DeepL documents on its own site.

  • File translation: according to DeepL’s document translation page, it handles PDFs (scanned or digital), Word, PowerPoint, Excel, HTML, SRT and TXT files, and preserves the original formatting and layout.
  • Glossaries: DeepL glossaries let you set your preferred translation for key terms, product names and company phrases, applied with the context of the sentence.
  • Formality and context: the DeepL API documentation describes a formality setting (more formal or more informal) for target languages that support it, and an option to pass extra context with the text.

DeepL also states that its models outperform major AI translation platforms in blind tests. That is DeepL’s own claim, on its own page. I have no reason to argue with it, and it is not the point. The point is what happens after the machine has done its part.

Where the settings stop

Where it is available for French, a formality setting chooses between “tu” and “vous”. A glossary fixes your product names. Neither one knows your reader, your brand or the joke in your headline. That part is still a human decision.

Fluent French can still be wrong French

Machine errors are rarely ugly. They are plausible. On my AI content editing page, I sum it up this way: false friends, outdated phrasing, subtly wrong terminology. Nobody notices until a French reader does. Here are the traps I check in any machine output, DeepL included. Every example goes from English into French, my only language pair.

False friends

  • “Eventually, the app will support ten languages.” A literal rendering gives “Éventuellement, l’application prendra en charge dix langues.” In French, “éventuellement” means possibly. The right version: “À terme, l’application prendra en charge dix langues.”
  • “The actual cost was lower.” Watch for “le coût actuel”, which means the current cost. The right version: “Le coût réel était inférieur.”
  • “She made a sensible decision.” A French “décision sensible” is a delicate or sensitive one. The right version: “Elle a pris une décision raisonnable.”
  • “I borrowed it from the library.” A “librairie” is a bookstore. The right version: “Je l’ai emprunté à la bibliothèque.”

Each wrong version is correct French. That is why a spellchecker will never flag it.

One English word, two French meanings

English is full of words that play two roles. “Book” is a noun and a verb. “Quote” is a quotation and a price estimate. “Range” is a product line and a kitchen stove. A machine has to guess, and in my French article on the limits of automatic translators and correctors, I list typical errors I correct: “Please find attached our quote” turning into “notre citation” instead of “notre devis”.

Calques and idioms

A calque copies the English structure word for word. The French is understandable, and it sounds translated.

  • “We need to address this issue.” The calque “adresser ce problème” is a classic. The right version: “Nous devons traiter ce problème.”
  • “Let’s touch base next week.” No base to touch in French: “Faisons le point la semaine prochaine.”
  • “It’s not rocket science.” The French idiom is “Ce n’est pas sorcier.”
  • “You hit the nail on the head.” The French idiom is “Vous avez mis dans le mille.”

Idioms are hard for machines in general. The organizers of the WMT24 general machine translation task report that an idiom test suite submitted by the Árni Magnússon Institute (English to Icelandic, not French) produced relatively low scores, with considerable room for improvement.

Tone and register

Picture a fitness app with this line: “Ready to crush your goals?” A literal version gives “Prêt à écraser tes objectifs ?” Three decisions hide in it. Tu or vous? Masculine “prêt” or feminine “prête”? And does “écraser” sound motivating or violent? A natural option avoids all three traps: “Envie de vous dépasser ?”

Can a dropdown menu make those choices for you? I don’t think so. It depends on your brand, your audience and the country where they live.

MTPE: how I turn machine output into publishable French

Machine translation post-editing (MTPE) is one of my specialties. Your content is machine-translated first, then I verify and rewrite it to human quality. It is faster than translating from scratch, and far more reliable than raw output. Here is my order of work.

  1. Source first: I read your English and any glossary or style guide you have.
  2. Meaning: I compare each French sentence with the English. Omissions, additions, false friends and wrong senses get fixed here.
  3. Facts: names, numbers, dates, units and product terms are checked one by one.
  4. Naturalness: calques, idioms, rhythm and register get rewritten for readers in France.
  5. Proofreading: grammar, agreement, French punctuation and spacing. This is where my Certificat Voltaire score of 994/1000, expert level, does its job.

Behind that step list, there is training too: I hold the certified Translator title, level 7 (Edvenn, 2023). And living in France helps a lot. I catch the reference that fell flat and the phrasing that aged badly. A translator abroad, or an AI, works from memory of the language. I work from inside it.

When MTPE is not the right choice

Sometimes post-editing costs more effort than a fresh translation. A slogan, a pun or a brand manifesto usually needs to be rewritten, not repaired. In that case I tell you, and I translate from the English directly. You will find both options on my English to French translation page.

How to check a DeepL translation before you publish

You do not read French? You can still spot the risky zones.

  • Flag every idiom, pun and cultural reference in your English before translating.
  • Build a glossary for product names and key terms, and apply it every time.
  • Decide on tu or vous for your brand, once and for all.
  • Check that every number, date and name survived the trip.
  • Have a native French reader compare the French with the English, not just read the French.

That last point matters most. A reviewer who only reads the French cannot see what went missing. A ZEISS project lead wrote that I “consistently demonstrated exceptional attention to detail” on multiple English-to-French translation and proofreading projects. That attention is what post-editing is made of.

Do you have a DeepL draft waiting in a folder? Get your translation post-edited, or send me your file: you get a firm quote within 24 hours.

Sources

Charline Brainez
Need a native eye?
I proofread and translate your French so it reads as if it had been written in France.

Certified proofreader (Certificat Voltaire 994/1000), living in France. Written quote within 24 hours.Charline

Questions. Short answers.

Is DeepL accurate for English to French?

Often, for a first draft. The French is usually fluent, but false friends, calques, idioms and tone can still be wrong. I would not publish it without a bilingual human review.

What is machine translation post-editing (MTPE)?

Your content is machine-translated first, then I compare it with the English and rewrite it to human quality: accurate meaning, natural French, consistent terms.

Is post-editing cheaper than a full translation?

It can be, when the machine output is a good base. When a text is creative or full of idioms, a fresh translation is often the better choice. I tell you which one fits in my quote.

Get a quote