Wikipedia bans ai-generated text after flood of errors and low-grade junk

Jimmy Wales thought machine prose would patch the gaps in the world’s largest volunteer encyclopedia. Instead he watched it fill up with poisoned summaries, phantom references and garbled history. On 20 March he asked his unpaid army to vote: keep the bots or exile them. The tally was 40–2. Overnight, Wikipedia’s English edition became the first global knowledge base to slam the door on large-language-model authorship.

How the ‘temple of the mind’ turned into a junkyard

Wales once called Wikipedia a sanctuary for human reason. That was before ChatGPT derivatives began burping out 10,000-word entries on medieval coinage in seconds. The lure was obvious: volunteers could finish backlogs while sipping coffee. The outcome was grim. Climate-science pages sprouted citations to non-existent journals. Biographies of living scientists ballooned with AI hallucinations. Vandal patrollers stopped counting malicious edits; they were too busy chasing algorithmic ghosts.

Editors coined a term for the slurry: LLM-laundered text. It reads fluently, slips past plagiarism detectors and collapses under scholarly scrutiny. One veteran told me he spent a Saturday night surgically removing 37 paragraphs on Ottoman tax law that cited a 1920 source that never existed. “The machine dreamed it,” he said, voice raw from caffeine. “We’re the janitors of its nightmares.”

The knife-edge vote that rewrote the rules

The knife-edge vote that rewrote the rules

Behind the scenes, the Wikimedia Foundation convened Zoom calls that felt like emergency wards. Data analysts showed curves no one wanted to see: error rates doubling every three months since GPT-4 integration. Wales put the ban proposal on-wiki and stepped back. The comment section turned into a battlefield. Purists waved the flag of epistemological purity. Tech-savvy pragmatists argued for tighter guardrails, not prohibition. One admin posted a screenshot of an AI draft that described the Battle of Trafalgar as a 19th-century naval skirmish between Spain and Finland. That jpeg killed the middle ground.

When voting closed, even opponents admitted the result restored a strange calm. Articles stabilised. The queue for new submissions shrank by half; quality, not quantity, became the bragging right.

What survives under the new regime

What survives under the new regime

The prohibition is surgical, not absolute. Volunteers may still run grammar bots trained on public-domain prose. Translators can auto-fill tedious infoboxes if every fact is cross-checked against sources. The key line in the policy manual now reads like a haiku of distrust: “No LLM may originate narrative.” Interpret that too loosely and you lose your mop; repeated violations trigger site-wide bans.

Amazon, Microsoft, Meta and Perplexity—each paying for live Wikipedia firehose access—received updated licensing terms last week. The small print obliges them to flag any synthetic text scraped since January. Compliance is voluntary; enforcement is another matter. Inside the foundation, engineers are testing cryptographic watermarks that twitch when an AI fingerprint is detected. The project has no deadline; the arms race rarely does.

The aftershock felt across silicon valley

The aftershock felt across silicon valley

OpenAI cancelled its long-promised ‘adult mode’ hours after the Wales vote. Staffers whisper about internal panic that bad training data—much of it harvested from Wikipedia—could poison future models. Meanwhile, Google Discover’s algorithms suddenly favour pre-2021 sources. Coincidence? A product manager laughed, then stopped laughing. “We’re all downstream of Wiki,” she admitted, glancing at her badge to check if it was still visible.

Sony’s halt on SD-card orders in Japan is unrelated on paper, yet the timing feeds a broader narrative: the information supply chain is fracturing. When the world’s default reference work chooses human friction over silicon speed, chip makers and cloud giants feel the chill.

Back in the volunteer trenches, an editor named Maria posted a screenshot of her dashboard: zero pending AI flags for the first time in 14 months. She captioned it with a single word: “Quiet.” The encyclopedia that anyone can edit just became the one that bots can’t touch, and the silence is deafening.