Article · 8 min read

How Wikipedia's editors became the best AI detectors on the internet

Last updated: May 2026

No venture funding. No machine learning model. Just a group of volunteers who read a lot of articles and started noticing the same things over and over. The best guide to spotting AI writing didn't come from a tech company. It came from Wikipedia.

Sometime in early 2023, a Wikipedia editor named Ilyas Lebleu noticed something strange. Articles were coming in that didn't sound right. The grammar was fine. The structure was fine. But the voice was off in a way that was hard to name. The writing was too smooth, too even, too keen to tell you how significant everything was.

He wasn't the only one seeing it. Across the encyclopaedia, experienced editors were running into the same thing: a wave of submissions that read like nobody in particular had written them. Some had invented sources. Some described people who didn't exist, or put real places in the wrong country. One article about an Ottoman fortress turned out to be entirely fictional, and sat there for nearly a year before anyone caught it.

ChatGPT had launched in November 2022. The flood had begun.

A cleanup crew forms

Rather than panic, the editors did what Wikipedians do: they organised. In 2023 a group of them started WikiProject AI Cleanup, a volunteer effort to find and fix AI-generated content across the site.

What's worth understanding about this group is who they are. Wikipedia editors are, collectively, some of the most experienced readers of prose on the planet. They spend their free time reading, fact-checking, and rewriting text to a strict house style. They have opinions about commas. When a few thousand of these people start paying close attention to the same problem, they notice patterns the rest of us would miss.

And they had a particular advantage. Wikipedia is supposed to read a certain way: neutral, plain, sourced, no salesmanship. AI writing has the opposite instincts. It loves to editorialise, to admire its subject, to remind you why something matters. Against the flat backdrop of an encyclopaedia, those habits stand out like a highlighter mark.

"We started to notice a lot of articles which were written in a style that didn't match the style we usually saw on Wikipedia." — Ilyas Lebleu, in an interview with NPR

The guide that came out of it

As the editors compared notes, they started writing down what they were seeing. That document grew into a page called Signs of AI Writing, and it's now one of the most thorough catalogues of AI tells anywhere. It runs to something like 15,000 words. It is not light reading. But it is the real thing: a field guide built from thousands of actual examples, not a marketing page guessing at what AI might do.

What makes it good is that it goes beyond the obvious. Early on, the tells were easy. AI would leave in its own chatter ("Certainly! Here is a Wikipedia article about..."), or invent citations that led nowhere. Those are still around. But as the models improved, the editors had to get sharper, and the guide started cataloguing the subtler stuff:

None of these are wrong on their own. Plenty of humans write "rich cultural heritage." The skill the editors developed is reading the combination, and weighing it against what they know about how real writing behaves.

Why volunteers beat the commercial tools

Here's the part that surprised me when I first read into this. The Wikipedia editors are, in practice, better at this than most of the paid AI detectors on the market.

The commercial tools mostly work by statistics: they measure how "predictable" text is and convert that into a percentage. The trouble is that careful human writing is also predictable, so these tools throw a lot of false positives, especially at students and people writing in a second language. Wikipedia's own guidance is blunt about this. Their editors' guide explicitly warns that detectors like GPTZero "often suffer from false positives" and "should never be used as the sole evidence when accusing someone of AI usage."

The editors take a different approach, and it's one worth copying. They treat a tell as a reason to look closer, never as a verdict. Their guidance even does the maths on overconfidence: research they cite suggests that even expert LLM users only correctly identify AI writing about 90% of the time, which means if you confidently flag ten articles, you've probably just falsely accused one real person. So they build in caution. They check sources. They look at edit history. They ask before they accuse.

That combination, a sharp eye for patterns plus a deep wariness of their own certainty, is exactly what good detection looks like. It's also the opposite of a tool that flashes "97% AI" and leaves you to it.

What we took from it

Telltale is built on this guide. When you paste text into our tool, the patterns we check for are drawn directly from the work WikiProject AI Cleanup has done: the vocabulary, the structural habits, the promotional tone, the vague attributions. We codified their field guide into something you can run in a couple of seconds.

THE PART WE KEPT

We also kept the editors' humility about it. Telltale shows you the evidence and explains every pattern it found, but it won't tell you a verdict is certain, because it isn't. The tool is a starting point for a human judgement, not a replacement for one.

→ Try it on something

There's something fitting about where this guide came from. The most useful defence against machine-written text was assembled, by hand, by a few thousand people who care more about good writing than almost anyone else online. They didn't build a model. They just paid attention, wrote down what they saw, and shared it for free.

Which is, when you think about it, a very human thing to do.

Further reading

Published May 2026 · telltale-ai.com
All articles · Back to tool · Privacy · Terms