Wikipedia's AI cleanup volunteers explain how they spot machine writing now

In an IEEE Spectrum interview, Ellia Lebleu, a founding member of the volunteer WikiProject AI Cleanup, says early AI text was easy to catch from phrases like "stands as a testament", but by 2025 the signs became structural, such as the "rule of three" and negative parallelisms, plus behaviour like several long articles written in minutes. The bigger worry now is citations: hallucinated sources are much rarer, but real sources that do not support the claim are, in Lebleu's words, extremely commonplace. Using the detector Pangram, Lebleu estimates roughly 20% of new articles and drafts were AI-generated in late 2025 and 15 to 16% in early 2026, which they call a conservative lower bound; English Wikipedia has since prohibited using LLMs to generate articles. Editors also use an AI source checker, and the Wikimedia Foundation is building Paste Check to flag text pasted from some chatbots. On Hacker News (26 top-level comments) there was no consensus. Several commenters backed the cleanup and said AI cannot yet write or verify encyclopedia content reliably; others asked why it matters who wrote a page if the content is good, and replies said volunteer reviewers cannot keep up with the volume. One commenter said they had saved a copy of Wikipedia from April 2024 as a clean reference, and another doubted the cleanup can keep up at all. Sorted take: the citation problem is the one to copy at work; a real link that does not say what the text claims is harder to catch than a fake one.