About 20 percent of new Wikipedia drafts were AI-generated; a volunteer team cut it to 15

Ellia Lebleu started WikiProject AI Cleanup after editors noticed articles with stock phrases like "stands as a testament".

In brief

With AI-detection company Pangram, Lebleu estimates about 20 percent of new articles and drafts were flagged in late 2025, falling to roughly 15 to 16 percent in early 2026. English Wikipedia now bans LLM-generated articles, and a source verifier checks whether citations support their claims. She cannot tell whether the drop is policy or models that evade the detector.

Commenters supplied the reader's side.

How the cleanup works

About 20 percent of new Wikipedia drafts were AI-generated;
About 20 percent of new Wikipedia drafts were AI-generated;

IEEE Spectrum profiles Ellia Lebleu, a founding member of the volunteer WikiProject AI Cleanup, launched in 2023. Around December 2023 editors noticed articles with phrases like "stands as a testament" and odd conclusion sections. By 2025 the tells were subtler: the rule of three, negative parallelisms, and accounts producing many long articles in minutes. Fabricated citations have become rarer, but real sources that do not support their attached claims are common.

Working with Pangram, Lebleu estimates about 20 percent of new articles and drafts were flagged in late 2025, falling to roughly 15 to 16 percent in early 2026. Pangram produced no false positives on pre-ChatGPT content, so she treats the numbers as a conservative lower bound. English Wikipedia now prohibits using LLMs to generate articles, a volunteer built an AI source verifier, and the Wikimedia Foundation is developing Paste Check.

What commenters saw on the ground

A commenter who tried to fix AI-generated text reported broken formatting, made-up text and links on wrong keywords, and said his edits were immediately reverted by Wikipedia's own AI bot, so he gave up. Another wrote:

"There will be a point in the future, possibly near future, when AI will be capable of reliably producing encyclopedic content, and we'll have tools to verify that the content produced passes reasonable standards."

He added that today is not that day, especially on verification. A third reader said a history-of-AIDS article he loved a decade ago is now pages of probably factual but unsatisfying text.

Lebleu's own line, from the article, is that AI "was something to clean up. Today, it's a tool to help us with the cleanup." Whether detectors can keep pace with newer models is the open number.

Get breakdowns like this every day

The top Hacker News stories, unpacked — by email, Telegram, or RSS.

Next story · 10 of 12 Read more
🚀Tech

Tailscale says most of its SSH bugs are privilege escalations, and a patch was due October 27

Tailscale calls the issues serious but is not recommending that everyone disable Tailscale SSH.

★ 198 · 💬 108 · tailscale.com · 2 min