# Wikipedia is Cleaning Up 8,000 Suspected AI-Generated Articles

*Technology · Technology Desk · Published 2026-11-09T09:40:36Z*

Canonical: https://amk.news/article/wikipedia-is-cleaning-up-8-000-suspected-ai-generated-articles

> Wikipedia is currently scrubbing thousands of AI-generated entries from its platform to combat a rise in sophisticated digital hoaxes. The cleanup project began after volunteers discovered entire fabricated histories, including a nonexistent Ottoman fortress, maintained by large language models.

## Key points

- On December 4, 2023, Wikipedia volunteers established WikiProject AI Cleanup after deleting a fraudulent ChatGPT-generated article about a nonexistent Ottoman fortress named Amberlihisar.
- Ellia Lebleu noted that early AI-generated articles in 2023 featured flowery patterns, whereas modern models by 2025 revealed themselves through repetitive sentence structures and three-part lists.
- Wikipedia introduced the G15 criteria in August 2025 to allow administrators to swiftly remove pages featuring obvious hallmarks of artificial intelligence generation without formal discussions.
- New Wikipedia guidelines implemented in March 2026 explicitly prohibited writing articles or rewriting existing text using large language models, restricting exceptions to translations and minor corrections.

On January 2, 2023, a new article appeared on English Wikipedia detailing a nonexistent Ottoman fortress in Trabzon. According to the text, the fictional "Amberlihisar" was commissioned by Sultan Mehmed the Conqueror in 1466 as a defensive bulwark against the Ak Koyunlu state. Spanning over 2,000 words, the article described the fortress’s name, construction details, various sieges, and its purported restoration after Russian shelling during World War I. It was complete with impeccably formatted footnotes and had even passed the initial review process used by volunteer editors for new submissions.[1](#ref-1)[2](#ref-2)

_The true fortress of Trabzon consists of the walls in the Ortahisar district. There is no historical record of the Amberlihisar fortress described in the fraudulent article. Photo: İlayda.tr, Wikimedia Commons, CC BY-SA 4.0._
The castle does not exist. On December 4, 2023, the article was deleted as an "obvious hoax"; the entire entry had been generated by ChatGPT, and none of the sources were authentic.[1](#ref-1b) On the same day, a group of editors established WikiProject AI Cleanup, a volunteer initiative dedicated to identifying and removing AI-generated content from Wikipedia.[3](#ref-3)

Key Insights

- How early AI-generated articles were identified and why they are harder to detect today.
- Why footnotes linked to real articles are considered a more insidious threat than entirely fabricated sources.
- The evolution of Wikipedia’s editorial policies over two years regarding machine-generated content.
- How errors introduced into Wikipedia can propagate through chatbot responses.

## How can readers and editors distinguish machine-generated text?

Ellia Lebleu, a co-founder of the project and a doctoral researcher in AI at Inria in Rennes, France, noted that the earliest suspicious articles from 2023 were identified by recurring, flowery patterns. Models often described bridges or towns as "evidence of a rich cultural heritage" and generated "conclusion" sections that are rarely found in traditional encyclopedia entries.[4](#ref-4) By 2025, however, the indicators shifted from word choice to structural rhythm. Modern models often reveal themselves through repetitive sentence structures, such as using three-part lists in every paragraph or constructing rigid contrast sentences using negation. While no single feature serves as definitive proof, their accumulation within a single text raises significant suspicion. Volunteers also analyze behavioral patterns, such as accounts that create five articles of thousands of words each within a five-minute window, which are automatically flagged for review.[4](#ref-4b)

Footnotes present a greater challenge. While entirely fabricated citations like those in the Amberlihisar hoax became rarer as AI models gained internet-browsing capabilities, a more common issue has emerged: linking a real source to a claim the author never made.[4](#ref-4c) One example documented by the project involves an article about the insect species *Estola albosignata*. The AI-generated paragraphs were supported by genuine German and French publications, but the French source actually described an entirely different organism.[3](#ref-3b) Unless a reader cross-references the specific citation, the deception remains invisible. Consequently, rapid deletion policies now prioritize source verification: dead links, invalid ISBNs, irrelevant DOI numbers, or news reports dated after an event occurred are treated as evidence that the page was generated without human oversight.[5](#ref-5)

## How has Wikipedia’s policy on AI evolved?

Following a community discussion in July 2025, Wikipedia introduced the G15 criteria for speedy deletion in August. This policy allows administrators to remove pages—without opening a formal deletion discussion—if they feature hallmarks of AI generation, such as "Here is your Wikipedia article" prompts, references to a model’s training cutoff date, or placeholders like "[Date of Birth]" left unfilled.[5](#ref-5b)[6](#ref-6)

A more significant step followed in March 2026. New guidelines explicitly prohibit writing articles or rewriting existing text using large language models. The only permitted exceptions are for translations and minor mechanical corrections like spelling and punctuation, with "improving" prose being strictly excluded.[7](#ref-7) During the same month, the persistent use of language models was added to the list of grounds for user blocking. By May 2026, citing the backlog of work faced by the cleanup team, the community agreed that AI-generated contributions could be reverted without having to prove the error in every single instance.[6](#ref-6b) The Wikimedia Foundation also entered the debate; in June 2025, the organization considered placing AI-generated summaries at the top of articles but suspended the project following opposition from the editor community.[6](#ref-6c)

## Notes and Sources

- Wikipedia (2026). "List of hoaxes on Wikipedia: Less than one year", Amberlihisar entry; archived text: "Wikipedia:List of hoaxes on Wikipedia/Amberlihisar". [en.wikipedia.org](https://en.wikipedia.org/wiki/Wikipedia:List_of_hoaxes_on_Wikipedia/Less_than_one_year); [archive](https://en.wikipedia.org/wiki/Wikipedia:List_of_hoaxes_on_Wikipedia/Amberlihisar) [↩](#ref-1)
- Maiberg, E. (2024, October 9). "The Editors Protecting Wikipedia from AI Hoaxes". 404 Media. [404media.co](https://www.404media.co/the-editors-protecting-wikipedia-from-ai-hoaxes/) [↩](#ref-2)
- Wikipedia (2026). "WikiProject AI Cleanup". Project page, founding member records December 4, 2023. [en.wikipedia.org](https://en.wikipedia.org/wiki/Wikipedia:WikiProject_AI_Cleanup) [↩](#ref-3)
- Strickland, E. (2026, October 21). "How Wikipedia’s Editors Keep the Online Encyclopedia Human-Written". IEEE Spectrum, interview with Ellia Lebleu. [spectrum.ieee.org](https://spectrum.ieee.org/wikipedia-wikiproject-ai-writing-cleanup) [↩](#ref-4)
- Wikipedia (2026). "Wikipedia:Speedy deletion", G15: LLM-generated pages without human review; "Administrators’ newsletter", August 2025. [en.wikipedia.org](https://en.wikipedia.org/wiki/Wikipedia:Speedy_deletion#G15); [newsletter](https://en.wikipedia.org/wiki/Wikipedia:Administrators%27_newsletter/2025/8) [↩](#ref-5)
- Wikipedia (2026). "Wikipedia:Artificial intelligence", timeline of AI discussions (June 2025 machine summaries, July 2025 G15, March 2026 directive and blocking policy, May 2026 default removal). [en.wikipedia.org](https://en.wikipedia.org/wiki/Wikipedia:Artificial_intelligence) [↩](#ref-6)
- Wikipedia (2026). "Wikipedia:Writing articles with large language models" (WP:LLM). [en.wikipedia.org](https://en.wikipedia.org/wiki/Wikipedia:Writing_articles_with_large_language_models) [↩](#ref-7)
- Wikipedia (2026). "Category:Articles containing suspected AI-generated texts" subcategories and "Category:AI-generated content proposed for deletion"; counted via MediaWiki API on October 26, 2026. [en.wikipedia.org](https://en.wikipedia.org/wiki/Category:Articles_containing_suspected_AI-generated_texts) [↩](#ref-8)
- Brooks, C., Eggert, S. and Peskoff, D. (2024). "The Rise of AI-Generated Content in Wikipedia". Proceedings of the First Workshop on Advancing Natural Language Processing for Wikipedia (WikiNLP), pp. 67–79, Miami. [doi:10.18653/v1/2024.wikinlp-1.12](https://doi.org/10.18653/v1/2024.wikinlp-1.12) [↩](#ref-9)
- Turkish Wikipedia (2026). "Vikipedi:Büyük dil modelleriyle madde yazma" (proposal, Jan 3, 2026) and "Vikipedi:Büyük dil modelleri" (essay). [tr.wikipedia.org](https://tr.wikipedia.org/wiki/Vikipedi:B%C3%BCy%C3%BCk_dil_modelleriyle_madde_yazma); [essay](https://tr.wikipedia.org/wiki/Vikipedi:B%C3%BCy%C3%BCk_dil_modelleri) [↩](#ref-10)
- Miller, M. (2025, October 17). "New user trends on Wikipedia". Wikimedia Foundation. [wikimediafoundation.org](https://wikimediafoundation.org/news/2025/10/17/new-user-trends-on-wikipedia/) [↩](#ref-11)

---
Source: AMK News (https://amk.news). Reuse requires attribution and a link to the canonical URL.