Tu or Vous: Why AI French Writing Gets the Register Wrong

TL;DR: Tu, vous, AI French writing: the grammar in the draft is fine and the register is wrong. AI picks vous by default, opens with "Bonjour Madame, Monsieur" by default and closes with "Cordialement" by default, because each default is the centre of every French business email it has ever seen. Then it slides from tu to vous on the last line, because the closing formulas it memorised are all in vous. Asking for "professionnel mais chaleureux" does not help; that phrase describes thousands of writers and none of their habits. A voice lives in the concrete choices: which pronoun for which relationship, which address line, how the ask is shaped, how the mail ends. Count those from emails you already sent, hand them over as rules, and every future model executes them better than the last.

You ask ChatGPT for a quick follow-up to Pierre, a client you have tutoyé since the second meeting. It comes back in vous, opens with "Bonjour Monsieur," and closes with "Je vous prie d'agréer l'expression de mes salutations distinguées." You fix the pronouns by hand and notice halfway through that one "vous" survived in the middle of a sentence.

Nothing in that draft was incorrect. Everything in it was somebody else's.

Why Does AI French Writing Get Tu and Vous Wrong?

Because tu versus vous is a relationship decision, and a model that does not know the relationship decides by averaging. The average of written French business correspondence is vous, so vous is what you get, whoever you are writing to and however long you have known them. The model is not being formal. It is being central.

That centre has drifted away from a lot of real French workplaces. Internal mail at most tech companies and agencies runs on tu. Many client relationships live in the hybrid French corporate life invented for exactly this problem: vous plus first name ("Bonjour Pierre, pourriez-vous..."). Banking, law and public administration still hold the line at vous plus Madame or Monsieur. None of these is the average, and the average is what the model writes.

Here is what an AI draft typically gets right and wrong on a single email:

Every row in the right column is a register failure, not a grammar failure. Grammar checkers pass the draft. Pierre does not.

The second-to-last row is the failure people find hardest to explain. You told the model to use tu. It did, for six sentences. Then it closed with "Je vous remercie par avance" or "Bien à vous." The pronoun and the formulas are stored separately: the body follows your instruction, and the closing is retrieved from a very large pile of memorised endings, nearly all of them in vous, because formal letters are where French closing formulas come from.

What Changes Between Registers in French, Beyond the Pronoun?

Five things, and each writer sets them independently: the pronoun, the address line, the shape of the request, the closing formula, and the sentence rhythm. AI collapses all five into one setting called "formel" and applies it uniformly.

Take an account manager in Lyon writing three emails before lunch. To the client she has worked with for four years, the address line is "Bonjour Pierre," the pronoun is vous (they never made the switch and neither minds), the ask arrives in the second sentence as "Pourriez-vous me confirmer...", and the closing is "Bien à vous." To a prospect's finance director she has met once, the address line is "Bonjour Madame Martin," the ask is wrapped in "Je me permets de revenir vers vous" and "Seriez-vous disponible...", and the closing takes "Je reste à votre disposition" plus "Bien cordialement." To her own team, it is "Salut à tous," tu in the singular, an imperative with a "stp", and "Merci !" or nothing at the end.

Same person. Three registers. All three are professional and warm, and no adjective could tell them apart.

We wrote about the general version of this in why your writing voice changes when you switch languages, and we took apart the Japanese case in keigo, AI and the Japanese business email. French sits between the two: less grammaticalised than Japanese, far more than English, where "you" covers everyone. A French reader gets the relationship handed to them in the second word of the email and knows before the ask whether it is the right one.

The five moving parts, spelled out:

1. Pronoun. Tu or vous per relationship. Some writers tutoient everyone below director level and vouvoient everyone above it. Some switch the day a client suggests it and never before. This is a rule per relationship, not a personality trait.
2. Address line. "Bonjour Pierre" / "Bonjour Monsieur Durand" / "Monsieur," / "Cher Pierre" / "Salut Pierre". Each carries information, and most writers use two of them for their whole career.
3. Request shape. Imperative ("Envoie-moi le devis"), plain question ("Peux-tu m'envoyer..."), conditional ("Pourriez-vous..."), or fully cushioned ("Serait-il possible de..."). AI defaults to the cushioned form regardless of the relationship.
4. Closing. "Cordialement" / "Bien à vous" / "Belle journée" / "À très vite" / "Je vous prie d'agréer...". The set is small and public, so the choice within it is a signature. A writer who closes client mail with "Bien à vous" and internal mail with nothing has a fingerprint, and it is not in any adjective.
5. Rhythm. Sentence length, how many subordinate clauses, whether the ask is its own line. This is where French AI output is furthest from most working writers.

Our measurements bear out that last point. Across the 320-sample benchmark behind our per-language baselines, French AI output scores 75 on sentence complexity, the highest of the four languages we measured, and 32 on conciseness, the lowest. Its formality score is 42, also the lowest, which sounds like good news for the tu-writer until you see what it means: a draft that is structurally elaborate and tonally loose at once, dense clauses wrapped around a register that belongs to no particular relationship. Those figures are averages, and the average is the problem: the model does not write at your sentence length, it writes at 32.

Why Doesn't "Professionnel Mais Chaleureux" Fix It?

Because the phrase is a pointer to a register the model already knows, and what it points to is the centre of everything written under that label. Type "professionnel mais chaleureux", or "formal but friendly, appropriate for French business", and the model retrieves its average of French business correspondence. Your colleague who types the same words gets the same email. You two write nothing alike.

AI does not know your tone. It knows words that describe tone.

This is the same failure we laid out for English in why tone adjectives are not your voice, and French sharpens it in one specific way: the adjective cannot carry the pronoun. "Chaleureux" does not tell the model tu or vous. "Professionnel" does not tell it Bonjour Pierre or Bonjour Monsieur. The two words that decide how the email lands are exactly the two an adjective cannot express, so the model falls back to the default for both.

The common workarounds do not close the gap either:

• Uploading past emails as reference files. The model paraphrases their content when it judges them relevant. It does not absorb the habits. Five emails cannot tell it whether your "Bien à vous" is a rule or a mood.
• Asking for "the same tone as my last email." Works for one draft, evaporates in the next conversation, and copies whatever was unusual about that one email.
• Writing in English and asking for a French translation. English has one "you". The translator has to invent the pronoun, and it invents vous. It also maps "Best regards" onto a formal French closing, so a two-line note to a colleague arrives with "Je vous prie d'agréer" stapled to the end. We covered the mirror image in why AI English sounds translated.
• Switching to a newer model. A more capable model follows instructions more faithfully. Give it an adjective and it produces a more polished rendering of the same average.

That last point deserves its own section, because it is the one people bet on.

Will a Better Model Learn My French Register?

No, and the reason is structural. Model capability and user-data conditioning are separate layers. The lab improves the first with every release. It cannot supply the second, because the second is made of emails you sent and the lab has never seen them.

What a stronger model delivers is fidelity to instructions. Feed it "professionnel mais chaleureux" and it executes that instruction more thoroughly than the last model did: smoother French, more consistent vous, cleaner formulas. It has moved closer to the average, not closer to you. People who report that their French drafts got "more generic" after a model upgrade are describing exactly this.

The same property works in your favour the moment the instruction carries something precise. A profile that says "tu with Pierre, vous plus first name with every other client, never Madame, Monsieur; ask in the second sentence as Pourriez-vous, never Serait-il possible; close with Bien à vous, never Cordialement" is a set of rules a stronger model hits more accurately than a weaker one. The better the model gets at following nuanced instructions, the better it executes your profile. We made the general argument in why better AI models still don't know your voice. French is a case where it bites early, because the pronoun is binary: a model either honours your tu or it does not, and there is no partial credit.

So the model dropdown is the last thing to change. Fix what the model is being asked to execute, and every future release executes it better.

How Do You Give AI Your Own French Register?

Measure it from emails you have already sent, then hand the measurements over as rules. Your register is countable. Every one of the five moving parts above can be read off ten to fifteen emails in an hour, and each one becomes an instruction with no adjective in it.

Start by separating your samples. French in one folder, English in another, and within French, label each email by relationship: long-standing client, new contact, internal, supplier, hierarchy above you. Do not pool them. A profile averaged across your client mail and your team mail is a profile of nobody, which is the mistake you are trying to escape.

Then count, per register:

Pronoun. Tu or vous, and for whom. Write it as a list of relationships, not as a policy: "Tu: team, agency contacts, Pierre. Vous plus prénom: all other clients. Vous plus Madame/Monsieur: first contact only, until they reply."

Address line. Which greeting per relationship, and which ones you never use. Absence is invisible in a sample and has to be stated: "never Cher, never Madame, Monsieur."

Request shape. For each ask, which form did you use? Tally imperative, Peux-tu, Pourriez-vous, Serait-il possible. You will find a firm habit per relationship, and a habit is a rule. Note your maximum number of softeners before an ask.

Closing. List them by register. Most writers have two, sometimes three, and a long tail of formulas they would never type.

Rhythm. Words per sentence, typical and range. Where the ask sits. How often a sentence opens with a subordinate clause, which is the habit French AI leans on hardest.

A finished block for one register looks like this:

Registre : client existant, contact hebdomadaire. Ouvrir par « Bonjour Pierre, », jamais « Monsieur ». Tutoiement acquis ; ne jamais revenir au vous, y compris dans la formule de fin. Demande en deuxième phrase, forme « Peux-tu… », un seul adoucisseur maximum (« si possible »), jamais « Serait-il possible ». Phrases de 10 à 20 mots, une idée par phrase, pas de subordonnée en tête de phrase. Fermer par « À très vite, » ou « Merci ! ». Jamais « Cordialement », jamais « Je vous prie ».

Paste that where your instructions live: ChatGPT custom instructions, a Claude Project, a Gemini Gem, a Custom GPT's instructions field. OpenAI's own custom instructions documentation describes the field as a place to say how you want the model to respond, and a register block is the most literal use of it there is. Keep one block per register and tell the model which one applies. The "never revert to vous in the closing" line is doing real work: it is the one instruction that stops the seam described earlier, and the model needs it stated because its defaults pull the other way.

This is also what a Style Profile from MyWritingTwin.com does, per language. A multilingual user's French samples get their own Writing DNA analysis, measured against a French AI baseline (formality 42, conciseness 32) rather than an English one (formality 58, conciseness 42), across more dimensions than the five above. The output is a block of constraints, and your Writing Twin for ChatGPT, Claude and Gemini is that block deployed inside the tools you already use. Your English samples get a separate profile. Neither one is a translation of the other, because your two voices are not translations of each other.

A measured profile can be checked: "was that chaleureux enough?" has no answer, and "was it tu throughout, including the closing?" does.

How Do You Test Whether the Draft Is Yours?

Give the model the brief for a French email you already sent, put its draft beside yours, and count. No tools, five minutes, and the answer is a set of numbers rather than a feeling.

Pick one email per register from the last month. Write a two-line brief for each: who it was to, what you needed, what they already knew. Ask the model for the email in French and compare:

• Pronoun: yours, or vous by default?
• Pronoun consistency: the same from first line to last, or a switch at the closing?
• Address line: yours, or Bonjour Madame, Monsieur?
• Ask position and shape: which sentence, and imperative, Peux-tu, Pourriez-vous or Serait-il possible?
• Closing: yours, or Cordialement?
• Words per sentence: typical, shortest, longest

If the counts differ, the model does not know your register, whatever it says when you ask it directly (it will say yes; it always says yes). Adjective-based setups usually pass the first-contact email, because the average is closest to that register, and fail the internal and long-standing-client ones, because those are where your habits diverge from the centre and where tu shows up.

Run the same test after replacing the adjective with the measured block. The counts converge on the first draft and keep converging as you tighten single rules. That convergence is the difference between an email with your signature under it and an email with you in it.

The grammar was never the problem. The register was, and the register was never handed over.

See how far your AI drafts sit from your real writing with the Humanity Score test, or upload your French samples and let a Style Profile do the counting.