Wikipedia pulls the plug on ai articles after quality nosedives
Wikipedia’s grand AI experiment lasted six months. The fallout: a flood of half-true trivia, scrambled citations, and editors screaming into their talk pages. On 20 March the community voted 40-to-2 to yank large-language models out of article space. The English-language encyclopaedia now treats LLM text as radioactive—usable only for spell-checking your own prose, never for generating it.
How the ‘temple of the mind’ turned into a junkyard
Jimmy Wales once called the site a temple. By January the aisles were littered with auto-generated stubs about obscure Romanian mayors and 14th-century Mongolian grain tariffs. Sources didn’t match claims; hallucinated footnotes pointed to dead links. Veteran volunteers spent nights hand-curating paragraphs that bots had spun in seconds. The calculus collapsed: more words, less trust.
Traffic didn’t immediately dip, but page-rating tools told the real story. Reader-reported errors spiked 144 % between October and February. Internal quality scores for new articles fell below the threshold normally reserved for draft space. The signal was clear: scale had cannibalised accuracy.
Big tech’s cash and the failed sweeteners
Amazon, Microsoft, Meta and Perplexity had lined up partnerships, each hoping to gift users tidy AI summaries perched atop wiki pages. The Wikimedia Foundation rolled out the feature last autumn; click-through data showed readers opened the summaries, then bounced—never scrolling to the human-curated text below. Ad revenue from those modules never offset the reputational cost.
Meanwhile, suppliers such as Sony quietly halted SD-card shipments to Japan as silicon shortages bit. The timing was exquisite: just as hardware pipelines narrowed, Wikipedia discovered it had mass-produced the one resource the web already owns in surplus—low-grade text.
Editors divided, but the verdict is brutal
Volunteers who write for free split into camps. Inclusionists argued that large models speed translation, fill missing stubs, and democratise contribution. Deletionists countered with diffs: paragraphs that drift from source, circular citations, and the soft erosion of authorial voice. The vote settled it. From here on, any editor who pastes raw LLM output risks a ban; even paraphrased sections must carry a manual audit trail.
Exceptions exist, but they are cosmetic. Grammar bots may suggest a comma, not a claim. Translation tools can propose a wording tweak, yet the editor must re-check every clause against the original source. In practice, the workflow is slower than pre-AI days—by design.
The wider lesson: content inflation has a feedback loop
Wikipedia’s reversal is a canary for every platform that bet on generative text as a cheap growth hormone. The encyclopaedia’s corpus is still small versus the open web, yet its credibility currency is prized. If LLM sludge can corrode a site whose sole asset is trust, imagine the rot inside ad-driven blog farms or product-review mills.
Other publishers are already quietly drafting similar rules. Stack Overflow imposes permanent bans on unmarked AI answers; scientific journals are piloting detection pipelines that reject submissions with tell-tale statistical fingerprints. The message is uniform: the novelty phase is over, and quality is again the moat.
For Wikipedia, the cost is measured in volunteer hours. Senior editors predict a 15 % drop in new-article velocity this year. They welcome the slowdown. Better silence than static.