According to an IEEE Spectrum interview with Ellia Lebleu, a founding member of Wikipedia’s WikiProject AI Cleanup, the volunteer effort began in December 2023 after editors noticed articles with phrases like “stands as a testament” and sections titled “Why is this important?” At the time, those tells were obvious; by 2025, the markers have become more structural, including rule-of-three constructions and negative parallelisms.
Lebleu says misleading citations are now a major worry. AI systems often find a real source but misrepresent its content, and fully hallucinated citations are rarer than they once were. Volunteers are building an AI-based source verifier, and the Wikimedia Foundation is developing Paste Check to flag text pasted from chatbots. Some commercial chatbots now refuse to write Wikipedia articles, though open-weight models can still be adjusted to bypass restrictions.
Detection remains an open contest. Lebleu’s tests with AI-detection company Pangram found roughly 20 percent of new articles and drafts flagged as AI-generated in late 2025, dropping to 15 to 16 percent in early 2026. Lebleu cautions that the drop could mean policies are working, models are becoming less detectable, or both. The deeper question, Lebleu says, is whether Wikipedia’s identity as a human-written corpus can survive if AI-assisted writing would make it more comprehensive.