French Wikipedia bans AI from writing its articles

The introduction rewritten at one in the morning

Sunday night bled into Monday, 12:58 a.m., when a volunteer editor who signs as Jules* opened a page on French Wikipedia and swapped out its introduction. His edit summary: "New introduction, per the survey results." The page is called "Wikipédia:Intelligence artificielle générative."

By the next day, the news had spread across French-language media. Wikipedia bans AI, the headlines said. The word really is there: French Wikipedia's page now states, translated from the French, that "the principle is a ban on the use of generative artificial intelligence [...] for writing articles." Until that night, the same page said it was merely "strongly discouraged."

Below that introduction, the rest of the text barely moved. That's not an editing oversight: the document that organized the vote says so itself, before the first ballot was even cast.

104 votes for, 2 against

The survey opened on August 31. The last comment is dated September 13 at 10:55 a.m. On question 1, rewording the introduction: 104 for, 2 against, 7 neutral. On question 2, adding an AI section to talk pages: 90 for, 5 against, 7 neutral. To vote, editors needed an account at least 90 days old, with 500 edits on the clock.

113 votes cast, against the 6,551 accounts that made at least five edits on French Wikipedia in August, might look thin. It's actually the normal turnout this community gets on this exact topic: 107 and 115 votes on the two January 2024 polls, 138 in May 2025, when the text was adopted as a guideline. What changed over two years is the opposition: on AI-generated text specifically, 23 votes against in 2024, 14 in 2025, 2 this time.

One of those two dissenting votes came with a four-point argument. Its author notes he wrote it with help from Claude Opus 5, though the vote itself is his own. His first point takes aim at the word the whole new introduction rests on: "'Quasi-systematically' isn't established. We only see the detected uses, meaning the bad ones; properly sourced AI-assisted contributions are invisible by construction." Another editor replied that generative AI isn't a neutral tool, and that French Wikipedia's core value remains "collaboration between humans."

Four tolerated uses before, four after

The guideline carries a list of AI uses it tolerates. Before the vote: summarizing a source you've already read, help with rewording, fixing spelling and typography, checking what a foreign word means.

After the vote: the same four, in the same order. Three are identical character for character. The fourth, summarizing sources, got tightened rather than expanded: editors must now verify that the summary matches the source. The four principles framing these uses, compliance, restraint, responsibility, and transparency, are also unchanged. The list even moved up in the page's outline, from a subsection to a full section of its own.

One last detail, the clearest in the whole file. The only sentence in the text that names a penalty says misuse "can lead" to an account being blocked from editing. It's worded identically, character for character, before and after. It appeared twice in the old version and once in the new one: the copy that disappeared was, precisely, the one in the introduction.

The warning sign got a fresh coat of bright red paint. The barrier itself hasn't moved an inch.

What the voting document says about itself

None of this is hidden. The proposal put to voters states its intent up front, translated from the French: "Reword and expand the introduction, while keeping the same pragmatic spirit (no absolute ban on the use of generative AI)." A note follows, just in case: "These tolerated uses are already described in the guideline, and this survey does not propose changing them."

The reasoning sits right above it, in a section called "Problems identified": some readers only ever read the introduction of a guideline page, and AI use that's merely "strongly discouraged" reads, the survey's authors write, "probably as 'allowed' for many readers, especially newcomers, misleading them."

So it's an editorial trade-off, made in the open: between the exact wording and the message a newcomer will take away, the community chose the message. It's a bit like painting DANGER on a workshop door instead of posting the full safety rulebook. A neutral voter summed it up mid-vote, without any snark: "It's banned, but there are exceptions, so it's not really banned." The survey's author replied that same evening that this had already been the spirit of the previous text, "but in my opinion that spirit came through imperfectly, too ambiguously."

What does a guideline actually require?

The page still carries the "Guideline" banner, unchanged as well. French Wikipedia distinguishes policies, adopted through a formal decision process, from guidelines, adopted by simple consensus. The reference text is explicit: "Unlike policies, guidelines may be followed less strictly and cannot always be enforced against an editor who doesn't respect them."

A guideline still carries weight, though. An internal community essay notes that some guidelines get cited routinely in blocking decisions, while properly adopted policies get broken with no consequences at all. On French Wikipedia, a text's label announces how formal it is, not how hard it actually gets enforced.

One thing did genuinely change in kind, though, and it comes from the second question. Messaging another volunteer with AI-generated text is now "considered a breach of etiquette rules." That paragraph didn't exist before September 14.

The caveat gets lost in the first recap

The interesting part is what happens to a text once it travels. Frandroid published an accurate piece Tuesday morning, spelling out in black and white that "the ban isn't total, however," and listing the tolerated cases. KultureGeek did the same the evening before, with a subheading devoted to the exceptions.

That same morning, an AI news site published its own version of the story, noting under the headline that it was written by its own agents "based on Frandroid" and "published without prior human review." All four tolerated uses had vanished. The guideline had turned into a policy, one the community would "now have to strengthen its vigilance and detection tools to enforce."

A strong word placed at the top of a text survives every recap. The exceptions that soften it, three paragraphs down, almost never do. The editors of French Wikipedia already knew this: it's the exact reasoning they wrote into their own survey before casting a single vote.