WikiProject AI Cleanup, the volunteer effort to find and remove AI-generated material from Wikipedia, states that the guideline WP:LLM has prohibited the use of large language models to generate or rewrite article content since March 2026, with limited exceptions for translation and basic copyediting (Wikipedia). In an interview published on 2026-10-21, founding member Ellia Lebleu told IEEE Spectrum that the detection company Pangram flagged perhaps 20 percent of new articles and drafts as AI-generated in late 2025, and 15 or 16 percent in early 2026 (IEEE Spectrum).
Key Facts
| Item | Detail | Source |
|---|---|---|
| Policy | WP:LLM prohibits LLMs to generate or rewrite article content since March 2026 | Wikipedia |
| Exceptions | Translation and basic copyediting | Wikipedia |
| Late 2025 estimate | Pangram flagged perhaps 20 percent of new articles and drafts | IEEE Spectrum |
| Early 2026 estimate | Pangram flagged 15 or 16 percent | IEEE Spectrum |
| Tool in development | Wikimedia Foundation is working on Paste Check | IEEE Spectrum |
Timeline
- 2022-11-30: ChatGPT was released, the date before which article text is very unlikely to be AI-generated (Wikipedia).
- 2023-12-04: The first founding members added their names to the project participant list (Wikipedia).
- 2026-10-21: IEEE Spectrum published its interview with Lebleu (IEEE Spectrum).
Details (as Reported)
The project’s stated goals include identifying AI-written text, identifying AI-generated images, keeping track of AI-using editors, educating editors, and protecting against disruption from AI agents (Wikipedia). A lot of LLM-generated text inserted from 2022 to 2026 remains on the site, according to the project page, and pages clearly entirely LLM-generated can be nominated for speedy deletion under WP:G15 (Wikipedia).
Early articles repeated phrases such as “stands as a testament” or “rich cultural heritage”, while by 2025 the signs had become subtler and more structural (IEEE Spectrum). Lebleu called fake or misleading citations a major worry, adding that AI more often finds a real source but gets its content completely wrong (IEEE Spectrum).
Lebleu explained that Pangram was tested on Wikipedia content from before ChatGPT and found no false positives, so its results provide a very conservative lower bound (IEEE Spectrum). The vast majority of AI-generated content is submitted by beginner editors, Lebleu added, and the rule has been made visible in welcome messages, editing help and advice (IEEE Spectrum).
On tooling, another Wikipedia editor has built an AI source verifier that checks whether the sources on a page match the claims, and the Wikimedia Foundation is developing Paste Check to flag text pasted directly from some AI chatbots (IEEE Spectrum). The Foundation is also working on an attribution framework that could show AI users when an article was last edited, how comprehensive it is and whether editors have flagged problems (IEEE Spectrum).
What Is Not Yet Confirmed
- Lebleu stated that it is still a little early to tell whether the rule prohibiting AI writing is working (IEEE Spectrum).
- It is not yet known whether the drop in flagged content reflects the recent policies or models getting more powerful before Pangram has had time to adapt, according to Lebleu (IEEE Spectrum).
- Because the AI source verifier is itself an AI model, there is no guarantee it will be correct every time, Lebleu stated (IEEE Spectrum).