French Wikipedia will block AI-generated news sites flagged by Next: how it works and what to check

In one sentence: If you cite online sources, for Wikipedia, a paper, an article or an internal note, the partnership between Wikimédia France and the French outlet Next is a reminder that a news site can look like a real publication without anyone writing or checking it, and that experienced editors fall for it too.

What was announced

On 12 November 2026, Wikimédia France, the association that supports Wikimedia projects in France, and Next, a French online technology news publication, published a joint text (in French) announcing a partnership. Next hands over its list of news-site domains it considers wholly or partly generated by AI; the volunteers of French-language Wikipedia add them to their tools so that these sites can no longer be cited as sources.

The cross-check has already been run. Out of a list of 19,926 sites supplied by Next, 2,187 domains appear in the references or footnotes of French Wikipedia, in more than 30,000 places, Next reports. That is about 11% of Next’s database already present in the encyclopaedia, by our calculation.

The issue is not French alone. Generated sites that copy the look of the press are aimed at advertising revenue and, sometimes, influence campaigns, the joint text says. What is specific here is the response: a news outlet’s database plugged into a Wikipedia edition’s filters under a formal agreement. For other European debates on AI and information, see our Society & Law section.

How the block works

Wikipedia normally moderates after the fact: an edit goes live, then volunteers review it. Abuse filters are the exception. They use logical conditions and regular expressions to block or flag an edit before it is saved, and they are only visible to, and editable by, about forty trusted French-speaking contributors, according to an editor quoted by Next.

The process described by Wikimédia France has three steps:

• Next sends its list of domains, matched against the sources cited on Wikipedia;
• Wikimédia France passes the list securely to the volunteers who maintain the filters, under an agreement it drafted for this ongoing exchange of data;
• those volunteers add the domains to the filters, which block or flag any attempt to add such a link.

The association does not decide for the community: it supplies data and a legal framework. Its role is to act “as a provider of tools, means and resources for the community”, says its executive director, Rémy Gerbet, in the joint text (our translation). No date has been given for the filters to go live.

Existing links: Noël sans IA in December

Filters only stop new links. For references already in articles, the Noël sans IA project (“an AI-free Christmas”) sets out, from 1 to 31 December 2026, 31 sites in 31 days among the most used as sources, each handled collectively: removing doubtful links, replacing them with archived copies or finding better sources. The domains stay hidden until their day.

The breakdown published by Next shows where the effort pays: of the 2,187 domains, 932 are cited only once, or 43% by our calculation, while 50 sites are each cited between 101 and 1,878 times. Cleaning those fifty removes most of the links.

Why French Wikipedia got here

The number of French-language AI-generated news sites detected by Next grew from 1,000 in January 2025 to nearly 20,000, and the share of the French population visiting the 250 most-visited ones each month rose from 25% to nearly 40% between October 2025 and April 2026, according to Médiamétrie measurements cited by Wikimédia France. Médiamétrie is France’s audience measurement company; its study itself is not reproduced in the partners’ text.

French Wikipedia has also tightened its own rule. Its guideline on generative AI now states that generative AI offers no guarantee of reliability, free reuse or verifiability, and that using it to write articles is therefore forbidden. Two exceptions remain: spelling correction is allowed, and careful use as a drafting aid is possible, under conditions, for editors who know the rules. The guideline names the usual failures: plagiarism, false information, and sources that are misused, low-quality or do not exist.

The two problems feed each other, the editor quoted by Next explains: large language models judge source quality poorly and tend to cite sites that are themselves generated by AI. French Wikipedia’s Observatoire des IA project tracks these uses and is linked to the English-language WikiProject AI Cleanup, where English Wikipedia editors are building a guide to identifying AI-generated writing.

What EU law says about AI-generated news text

Outside Wikipedia’s own rules, the EU AI Act has a transparency duty that applies here. Under Article 50(4) of the regulation, deployers of an AI system that generates text “published with the purpose of informing the public on matters of public interest” must disclose that the text has been artificially generated. The duty does not apply where the content “has undergone a process of human review or editorial control” and a person holds editorial responsibility for its publication.

These transparency rules have applied since 2 August 2026, as set out in our AI Act timeline. They do not make a generated site unusable as a source, and they do not oblige Wikipedia to do anything: the encyclopaedia’s ban is its own editorial choice. For brands, the same article of the AI Act is applied to advertising in our guide to labelling AI-generated ads.

How Next identifies these sites

Next says it does not use AI-text detectors, because of false positives and false negatives; it found that even the best of them, such as Pangram, too often took AI-generated content for human writing (Wikimédia France). Its method rests on a cluster of signs: the number of articles per bylined journalist, whether those journalists exist, wording typical of generative AI, expired domains bought up by marketing or search-engine-optimisation firms, membership of a network of sites built for advertising, and no legal notice. Investigative journalist Jean-Marc Manach says Next has had fewer than 10 removal requests since launch and accepted three, which were false positives (Next). Our French explainer on how reliable AI detectors are (in French) covers those limits.

A first list of just over 800 sites, shared with Wikipedia the previous year, produced no false positives when volunteers checked it, according to Wikimédia France. A list kept by a single outlet is still an editorial judgement: the community decides to add it to the filters, and a domain can be challenged.

How to check a source before citing it

French Wikipedia’s help page on identifying an AI-generated source gives criteria that work well beyond the encyclopaedia, and in any language:

• Does the author exist? Search for the name elsewhere. Social media profiles can be created for authors who do not exist, and a real name proves nothing: some sites switched to AI or borrow identities.
• How much do they publish? An author putting out around ten articles a day, other than news briefs, is described by the page as an extremely reliable sign of generative AI use.
• What does the legal notice say? Look for a real publisher and a postal address that exists. Without them, the site is not a quality source.
• Where are the sources? A serious outlet cites its source in the text; a list of links at the bottom of the page, outside any sentence, is typical of sites that rewrite other media.
• Do the comments ring false? Posted at intervals too regular to be human, always enthusiastic, never replying to one another.

Formatting signs on their own prove nothing, the help page adds: several are needed. When in doubt, do not use the source, and check with the Internet Archive’s Wayback Machine what a page said before a possible domain takeover.

For an automatic warning, Next offers a free extension, Alerte sur les sites GenAI, for Firefox and Chromium-based browsers. It flags sites in Next’s database, which mostly covers French-language outlets, and lets users report new ones; more than 5,000 reports have been received, according to Wikimédia France. The same work by Next led to nine French AI-generated news sites being dropped from Google Discover (in French) in early November.

What we checked

We read the joint text published by Wikimédia France and by Next on 12 November 2026 and, as they stood on 13 November 2026, French Wikipedia’s guideline on generative AI, its help page on AI-generated sources, the Noël sans IA and Observatoire des IA project pages, the English WikiProject AI Cleanup page, and Article 50 of the AI Act in the Official Journal. The 11% and 43% figures are our calculations from the numbers published by Next.

What we do not know: the list of 2,187 domains is not public, nor are the abuse filters, and no date has been given for them to take effect. The audience figures come from Médiamétrie, cited by the partners.