Wikipedia’s AI Cleanup volunteers
Pangram flags ~15–16% of new English articles as AI-written in early 2026, down from ~20% in late 2025
Partly confirmed
Status
- Claim
Partly confirmed
- Our reporting
- Medium confidence
- Importance
- 2 of 5
- Last verified
- 28 October 2026
Your AI and this story
- GPT-6.2 Astra143 days after its cutoff
- Claude Opus 5.5113 days after its cutoff
- Gemini 4 Hydrogen204 days after its cutoff
- Grok 4.7143 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 113 days before it.
Key facts
- Source: IEEE Spectrum (Eliza Strickland), ‘How Wikipedia’s Editors Keep the Online Encyclopedia Human-Written’, Oct 21, 2026; reached 483 points on Hacker News on Oct 27
- WikiProject AI Cleanup started in 2023; Lebleu noticed the pattern around December 2023 (articles with ‘conclusion’ and ‘Why is this important?’ sections)
- Pangram estimates (as reported): ~20% of new articles and drafts AI-flagged in late 2025, ~15–16% in early 2026; Lebleu calls them a conservative lower bound (no false positives on pre-ChatGPT text) and says the drop may reflect the policy or newer models Pangram has not adapted to
- English Wikipedia prohibits using LLMs to generate articles (adoption date not given in the article); most AI text comes from beginner editors
- Typical problem has shifted from invented citations to real sources that do not support the claim, which Lebleu links to retrieval-augmented generation
- Tools: a volunteer-built AI source verifier; the Wikimedia Foundation’s ‘Paste Check’ to flag text pasted from chatbots; a planned attribution framework showing AI users an article’s last edit date, completeness and flagged problems
- Lebleu: Claude has sometimes refused, saying ‘I can’t write a Wikipedia article; it would go against Wikipedia’s policies’
What happened
IEEE Spectrum interviewed Ellia Lebleu, a founding member of English Wikipedia’s WikiProject AI Cleanup, about how volunteers spot and remove machine-written text. Detection has moved from stock phrases (“stands as a testament”, “rich cultural heritage”) to structural tells (the “rule of three”, negative parallelisms) and output volume, such as five multi-thousand-word articles in five minutes. Pangram’s figures, as relayed by Lebleu, put the AI-flagged share of new articles at about a fifth in late 2025 and a sixth in early 2026.
Why it matters
Wikipedia is one of the largest sources of training and retrieval data for language models, so how much model-written text it absorbs feeds back into future models. The numbers are one vendor’s detector, relayed in an interview, not an audited count.
Sources
2 sources from 2 sites. Numbers match the chips in the text.
2 sources: 1 press, 1 reaction
Press
- IEEE Spectrum: How Wikipedia’s editors keep the online encyclopedia human-written (Oct 21, 2026)spectrum.ieee.org, press
Reactions
- Hacker News discussion (483 points)news.ycombinator.com, discussion
Changes
- Filed (Hacker News)