Wikipedia Volunteers Say 15% of New Articles Were Flagged as AI-Written in Early 2026

About 15 to 16 percent of new English Wikipedia articles and drafts were flagged as AI-written in early 2026, down from roughly 20 percent in late 2025, according to Ellia Lebleu, a founding member of WikiProject AI Cleanup. Lebleu gave the figures to IEEE Spectrum in a Q&A published October 21, 2026, and called them a conservative lower bound.

The drop coincides with a written rule. English Wikipedia's guideline on writing with large language models says using an LLM to generate or rewrite article content is "strictly prohibited," with two exceptions.

How the 20 and 15 percent figures were produced

Lebleu used Pangram, a commercial AI-detection tool, to score new articles and drafts. The method comes with one check: Lebleu told Spectrum that Pangram produced no false positives on text from before ChatGPT existed. That is why the percentages are treated as a floor and not a midpoint.

Lebleu is also pursuing a Ph.D. in AI at the Inria Center of the University of Rennes, per the piece, which was written by Eliza Strickland. The project is a volunteer effort that launched in 2023.

Period	Share of new articles and drafts flagged
Late 2025	about 20%
Early 2026	15 to 16%

What the cleanup volunteers look for

The project began around December 2023, when editors noticed phrases such as "stands as a testament" and "rich cultural heritage." Early giveaways also included invented conclusion sections and headings like "Why is this important?"

By 2025 the signs were subtler, Lebleu said: the "rule of three" and negative parallelisms. Editors also look at publishing volume. Fabricated citations have become rarer because of retrieval-augmented generation. The common problem now is real sources that do not support the claim attached to them.

What the rule allows and what enforces it

The guideline permits two uses. Editors may ask an LLM to "suggest corrections to their own writing," but only for spelling, punctuation, capitalisation and grammar, and only after human review. They may also use an LLM to translate an article from another-language Wikipedia into English, following separate translation guidance, which the page warns often causes the same policy violations as content generated from scratch.

Persistent use outside those exceptions "is considered disruptive" and can lead to a block. Because writing style alone can resemble LLM output, the page says sanctions need more than stylistic indicators. Reviewers are told to consider an editor's overall pattern and whether edits comply with core content policies.

The page's sidebar links a speedy-deletion criterion, G15, labelled "LLM-generated pages without human review." I did not read that policy, so what it requires is not covered here.

Most AI-generated submissions come from beginner editors, Lebleu said, so the rule is promoted in welcome messages and editing help.

The tools, and what nobody knows yet

Spectrum names three. An AI source verifier, built by another editor, checks whether cited sources match the claims, and Lebleu said it is not guaranteed to be accurate. The Wikimedia Foundation's Paste Check flags text pasted from some chatbots. The Foundation is also developing an attribution framework for people who cite Wikipedia in AI output.

The open question is causation. Lebleu said it is unclear whether the fall from 20 percent to 15 or 16 percent reflects the policy or simply better models that leave fewer tells for Pangram and editors to catch. She also raised whether Wikipedia should offer an AI-assisted option while keeping its human-written identity.

Models keep changing: our coverage of OpenAI's GPT-6.1 Terra shows how fast the output side moves, and local tools such as Hob put the same generation on a desktop.