Wikipedia blocks ai-written entries to stop silent erosion of facts
Wikipedia’s volunteer army drew a line in the sand last week: no machine-generated prose, no algorithmic paraphrasing, no synthetic citations. The reason is starkly simple—AI text doesn’t just repeat errors, it rebuilds reality with the confidence of a seasoned historian and the memory of a goldfish.
The mirror that rewrites what it reflects
Editors discovered that large-language models don’t copy; they hallucinate. A bot asked to summarize the causes of the 1973 Chilean coup delivered three paragraphs that read like textbook material, only the quoted CIA cable never existed and the death toll was off by 1,200. Human reviewers caught it, but only after the paragraph had already circulated on social media as “proof” of U.S. involvement. Multiply that slip by 300 languages and 60 million articles and the encyclopedia risks becoming a fun-house mirror.
YouTube is already drowning in that distortion. Search for “Roman concrete” and the top results are faceless channels narrated by synthetic voices reciting ChatGPT scripts. The visuals are stock, the facts half-right, the ad revenue real. Wikipedia sees the same sludge seeping into its footnotes and wants the valves shut.
Detection is a cat-and-mouse game the mouse keeps winning
Turnitin-style detectors boast 98 % accuracy until someone pings the latest model, feeds it a thesaurus and runs the output through a paraphrase layer. The text is new, the syntax shuffled, the fingerprints dissolved. Volunteers then waste evenings playing forensic linguist instead of polishing articles on medieval irrigation.
What’s left is asymmetry: a single bad actor with a GPU can carpet-bomb a niche page in minutes, while a scholar needs an afternoon to source one Persian manuscript. The community voted 312-to-27 to treat AI text as non-reliable unless its origin is transparent and every clause is cross-checked—effectively a ban, because who will admit the bot wrote half the entry?
A human exception carved in wikicode
The guideline isn’t absolute. Editors may use AI to fix grammar, translate sources or brainstorm outlines—essentially a spell-checker with delusions of grandeur. Authorship, however, must stay biological. The rule even carries a whiff of poetry: “Wikipedia is written by people, for people, about people.” Strip the humanity and you’re left with a tautology machine that spits out ersatz knowledge faster than we can vet it.
The decision lands at the worst possible moment. Traffic from Google’s AI Overviews is already cannibalizing clicks; volunteer recruitment is aging into extinction; donation banners blink redder each December. Banning the cheapest content factory on earth feels like self-amputation. Yet the alternative is a slow bleed where trust—Wikipedia’s only currency—evaporates one hallucinated footnote at a time.
Other platforms will flinch. Reddit threads, Stack Overflow answers, even peer-reviewed journals face the same flood. Wikipedia just chose to sink a stake in the ground before the frontier reaches the citadel. History will record whether that was prescience or hubris, but at least the record will still be written by someone who can be named, shamed or thanked.