Wikipedia draws a red line: human authors only, ai need not apply
Wikipedia’s volunteer army just declared war on the machine. After quietly purging AI-generated articles for months, the encyclopedia’s editors now publish the rationale: large-language models rewrite reality faster than they fact-check it, and the collateral damage is the truth.
The silent edit wars no one noticed
Between January and March, unnoticed by most readers, thousands of machine-written drafts vanished from public view. The reason wasn’t vandalism or spam; the prose looked fluent, even elegant. But veteran editors spotted the tell: a slight shift in date, a nuance sanded off, a source bent to fit the sentence. One example: a bot recast the Siege of Sarajevo as a ‘military stalemate’, erasing the civilian death toll. It took a human historian in Sarajevo three days to notice the rewrite.
That micro-incident became the smoking gun. Editors ran a controlled test: feed the same set of footnotes to ChatGPT-4 and to a first-year history student. The student misquoted one statistic; the AI invented two battles that never happened and cited them to out-of-print books no one could verify. The score: human 1, robot 0 – but only because a human was watching.
Why plagiarism detectors fail against synthetic prose
Traditional copy-and-paste fraud leaves fingerprints: identical strings, identical URLs. AI leaves none. Every sentence is freshly minted, every citation shuffled like a deck of marked cards. The best detector tools, tested internally, hover around 68 % accuracy on Wikipedia’s sample set. That’s worse than a coin flip when the stakes are global knowledge.
And hallucinations travel fast. A false claim seeded at 2 a.m. Berlin time can echo through thirty language editions before breakfast in Mexico City. Volunteers call it ‘whack-a-mole with smoke’.
The cost of keeping the lights human
Wikipedia’s traffic keeps climbing – 8.5 billion page views last month – yet the number of active editors has plateaued at 120 000. AI promises volume; the foundation wants veracity. The compromise: bots may spell-check, format tables, or flag broken links, but they may not author narrative text. Enforcement is social, not algorithmic. Any user can challenge an edit; if the author refuses to confirm human origin, the passage dies.
The policy is already reshaping recruitment drives. New onboarding banners now read: ‘Humans wanted, no salary, eternal glory.’ The joke masks a sobering reality: someone has to keep the frontier of fact a manned outpost.
Meanwhile, the rest of the web sinks deeper into synthetic sludge. YouTube’s search results brim with AI voice-overs reading Wikipedia knock-offs that were themselves AI-written, a recursive ouroboros of junk. The encyclopedia’s refusal to join the race is, for now, the only brake pedal in sight.
Expect courtship from start-ups promising ‘fact-aligned’ models. Expect open letters claiming censorship. Expect, above all, more quiet deletions at 3 a.m. by volunteers who still believe that knowledge is a handshake between people, not a probabilistic parlor trick. Wikipedia just bet its future on that conviction. The servers stay human; the truth, if we’re lucky, follows.