Society & safeguards

French Wikipedia wants to block 2,187 news sites flagged as AI-made

The question is who decided they were.

6 min read
French Wikipedia wants to block 2,187 news sites flagged as AI-made

Two footnotes in Allos

Allos is a village in the Alpes-de-Haute-Provence, in southeastern France. Its French Wikipedia page has a paragraph on the town's natural hazards, backed by two footnotes. The first points to a "municipal notice" that the reference attributes to the French Ministry of Ecology, accessed in 2012. The second points to the town's major-risk information document.

We clicked both on November 13. The first lands on the homepage of prim.net, which currently offers an article on buying a business and another on advances in eye surgery. The second opens bd-dicrim.fr, "news decoded", whose lead item is a travel guide to Alicante, "Spanish pearl".

In 2015, prim.net was the "Portal for the Prevention of Major Risks". The domain name stayed, the address too, and 382 articles on French Wikipedia, mostly pages for French towns, still link to it. We only checked what the pages show today. Both domains are among the 31 that volunteers plan to clean up in December, as part of French Wikipedia's push against sites that the French outlet Next classifies as AI-generated.

The story went public on November 12, in a text co-written by Next and Wikimédia France, the association that supports the encyclopedia in France. Since late 2024, Next has kept a list of French-language news sites that it estimates are generated "in whole or in part" or "massively" by AI. There were 1,000 in January 2025. Next now counts nearly 20,000.

A French Wikipedia contributor who signs as Jules* crossed that list with the encyclopedia's outbound links. By his count, 2,187 of those domains are cited in articles. According to the co-written text, more than 30,000 references and footnotes point to them.

The spread is very uneven. 932 domains appear only once. At the other end, 50 sites are cited more than a hundred times each, and the most-used one shows up 1,878 times. Picture a scattering of obscure sites across thousands of pages, and a few that are practically everywhere.

Two stories in one number

It's tempting to read "30,000 links" as 30,000 moments when a volunteer got fooled by a machine. Part of the story is exactly that. Jules* says in the Next piece (our translation from the French-language text) that beginners and veteran editors alike add these sites "in good faith", without noticing that the articles are produced by AI, often under invented bylines and plagiarized from real newspapers.

But there's a second story, the one in Allos. In his November 12 note on Wikipedia, Jules* points out that many links were added "at a time when they pointed to quality websites" that have since changed hands. A domain lapses, someone buys it for the audience it still drags along, and fills it with automated content. The reference hasn't changed by a single character. What sits behind it has.

Wikipedians have a word for these addresses: "vampirized" sites. In February 2025, Jules* and another contributor, Pa2chant.bis, went through 148 sites flagged by Next by hand, and 51 fell into that category. We don't know what share of today's 2,187 are in the same situation, and a 2025 sample of 148 is too old to extrapolate from.

A secret list for an open encyclopedia

The response has two parts. The first is to stop new links from going in. Next agreed to hand over its database on condition that it not be public, which required an agreement with Wikimédia France. So the list can't go on the encyclopedia's blacklist, which anyone can read. It sits instead in an anti-abuse filter, number 402, visible to about forty trusted contributors. Think of a bouncer who knows the faces but doesn't pin the list to the door.

The filter exists, but as of November 12 it wasn't doing anything. Jules* writes that it is "enabled but does nothing" until the community picks between warning the contributor and blocking the edit outright. The first opinions leaned clearly toward blocking. Jules* argues for it even for old, harmless links: keeping the original address sends readers, and traffic, to whoever occupies the site today. Better to cite an archived copy.

The second part is called "Christmas without AI": from December 1 to 31, one domain a day, starting with the most cited.

Volunteers have planned several treatments depending on the case. Remove every link when the site was never reliable. Remove only those added after the switch to AI. Or replace each link with the oldest available archive, when a respectable source has been hollowed out. For Allos, the third one applies.

Who decides a site is written by a machine?

Everything rests on Next's list, so it's worth a close look. Journalist Jean-Marc Manach says Next doesn't use automatic detectors, which it considers too unreliable. It works by eye, from a bundle of clues: an article count per journalist that no human could plausibly reach, journalists who don't exist, phrasing typical of AI, an expired domain name bought up by an SEO professional, no legal notice.

On reliability, the figure comes from Next itself: fewer than ten removal requests since the list launched, of which three were accepted as genuine false positives. That's the referee counting complaints against its own calls. A site that doesn't complain isn't proof it was rated correctly.

The co-written text says the volunteers' first checks confirmed the method, citing "the absence of false positives". The survey those volunteers published in February 2025, covering the first 148 sites examined, says something else: "2 are clearly false positives" (again our translation). Other sites, where AI use hadn't been established, were dropped from Next's list but not counted as errors, because they had other problems, such as plagiarism.

Two errors out of 148 is not much. But it's the only independent check that has been published, and it doesn't come up with zero.

And AI models?

The next question is obvious: Wikipedia is used to train AI models, so these sites would end up feeding the machines that wrote them. Researchers at the Wikimedia Foundation do note that the English edition often weighs heavily in training. But nothing published measures what the references on the French edition contribute. These are footnote links, and whatever they may have pushed into the text of articles hasn't been assessed.

The reverse direction is documented. Jules* observes that language models are poor judges of a source's quality and tend to cite sites that are themselves AI-generated. French Wikipedia, by the way, tightened its guideline on generative AI in September: the principle is now a ban on using it to write articles.

What's left is the lesson of Allos, which goes beyond AI. A footnote vouches for the source only on the day it was added. After that, it vouches for the address.

Was this article useful?

Topics covered:

EthicsAnalysis

Frequently asked questions

How many AI-generated news sites does French Wikipedia cite?
By the count of one contributor, Jules*, 2,187 domains that the French outlet Next classifies as AI-generated are cited in articles. According to the text co-written by Next and Wikimédia France on November 12, more than 30,000 references and footnotes point to them. The spread is very uneven: 932 domains appear only once, while 50 are cited more than a hundred times each.
Are these sites already blocked on Wikipedia?
Not yet. Next's list, which isn't supposed to be public, was placed in an anti-abuse filter, number 402, visible to about forty trusted contributors. But as of November 12, Jules* wrote that it is "enabled but does nothing" until the community chooses between warning the contributor and blocking the edit. The first opinions leaned clearly toward blocking.
Why did Wikipedia editors cite AI-generated sites?
There are two stories. Contributors, beginners and veterans alike, add them in good faith, according to Jules*. But many links were also added when they pointed to quality sites that have since changed hands: Wikipedians call them "vampirized" sites. In February 2025, 51 of the 148 sites checked by hand fell into that category; the share among today's 2,187 is not known.
How does Next decide a site is AI-generated?
Next doesn't use an automatic detector. It works by eye, from a bundle of clues (journalists who don't exist, an implausible publishing rate, an expired domain bought up, no legal notice). On reliability, the figure comes from Next itself: fewer than ten removal requests, three of them accepted. The co-written text cites "the absence of false positives"; the volunteers' February 2025 survey says "2 are clearly false positives" out of 148.
Do these sites end up in AI training data through Wikipedia?
Nothing published measures it. Wikimedia Foundation researchers note that the English edition often weighs heavily in training, but what the references on the French edition contribute has not been assessed. The reverse is documented: according to Jules*, language models judge sources poorly and tend to cite sites that are themselves AI-generated.
Katja Liersch

Katja Liersch

Co-founder & Journalist

Katja is a journalist and TV producer. With decades of experience in mainstream media, she brings to Declic Media the perspective of those discovering AI: curious, demanding and pragmatic. She ensures every piece of content truly speaks to everyone.

All articles by Katja →
The free AI newsletter