Wikipedia's volunteer editors noticed something strange in late 2023: a rash of new articles all containing the same verbal tics. Phrases like "stands as a testament" and "rich cultural heritage" appeared with unnatural frequency, alongside structural giveaways like conclusion sections and headings reading "Why is this important?" These were ChatGPT's fingerprints, and they were everywhere. The WikiProject AI Cleanup, founded by Ellia Lebleu and fellow editors in 2023, began as a pattern-matching operation and has evolved into something closer to an institutional immune response. The detection challenge has escalated rapidly. Early AI-generated articles were trivially identifiable by their clichés. By 2025, the tells had become structural rather than lexical — overuse of the "rule of three," negative parallelisms, and other rhetorical patterns that individually mean nothing but collectively signal machine authorship. Behavioral signals now supplement textual analysis: five thousand-word articles submitted within five minutes from a single editor is a red flag no prose detector needs to catch. The citation problem may be more corrosive than the prose problem. Outright hallucinated sources have become rarer thanks to retrieval-augmented generation, but a subtler failure mode has taken their place: real sources cited in support of claims they don't actually make. This is harder to detect at scale and more damaging to Wikipedia's core function as a verifiable knowledge base. An AI source verifier built by another editor checks whether cited sources match their claims, but it is itself an AI model with no guarantee of correctness. The numbers tell a story of partial progress and deep uncertainty. Working with AI-detection company Pangram, Lebleu tested the tool against pre-ChatGPT Wikipedia content and found zero false positives, establishing a conservative lower bound. In late 2025, Pangram flagged roughly 20 percent of new articles and drafts as AI-generated. By early 2026, that figure had dropped to 15-16 percent. But the causal mechanism is unknown — it could be policy deterrence, improved editor awareness, or simply that newer models evade Pangram's classifiers. English Wikipedia has formally prohibited editors from using LLMs to generate articles, and the rule is prominently displayed in welcome messages and editing guidance. The strategic picture extends beyond content quality. Wikipedia now positions itself as "the last good place on the internet" — a purely human-written corpus. This identity has commercial value: AI companies need clean, human-authored training data, and Wikipedia's commitment to remaining human-written makes it a uniquely valuable resource. Lebleu notes that even AI companies find this positioning interesting, creating a peculiar dynamic where the encyclopedia's resistance to AI-generated content increases its value to the AI industry. The existential threat isn't AI-generated articles — it's AI-mediated access. If Google's AI Overviews and similar products reduce direct Wikipedia traffic to near zero, the pipeline of new editors collapses. Users who only see Wikipedia content filtered through AI answers may never realize the encyclopedia is a living project they can edit. The Wikimedia Foundation is developing an attribution framework that would surface metadata — last edit date, comprehensiveness ratings, editor-flagged problems — through AI interfaces. Lebleu envisions readers eventually contributing through AI models themselves, submitting edit requests when they spot errors in AI-generated answers that cite Wikipedia. The trajectory Lebleu describes is striking in its pragmatism: "Yesterday, AI was something to clean up. Today, it's a tool to help us with the cleanup. Tomorrow, it will be the air we breathe." Wikipedia is attempting something rare in the platform era — maintaining institutional integrity while absorbing a technology that fundamentally threatens it, without pretending the technology will go away.
