Wikipedia spent 2 years building a system to catch AI writing. Then a developer turned it into a training manual.
Since late 2023, a group of Wikipedia editors called WikiProject AI Cleanup has been reading thousands of articles, flagging the AI-generated ones, and writing down exactly what gave them away.
They published a field guide running nearly 15,000 words. Dozens of patterns across five categories. Every one pulled from real articles they actually caught.
NPR covered it. TechCrunch called it "the best guide to spotting AI writing."
Then a developer named Siqi Chen read the page.
He saw perfectly specific detection rules. And he built Humanizer, an open-source tool that feeds those exact rules to Claude and rewrites AI text so it doesn't trip a single one.
The guide built to catch AI is now teaching AI to sound human.
Worth understanding why.
The detection system (and why it's different)
Most "AI detection" is vibes. Someone reads a paragraph and says "that sounds like ChatGPT."
Wikipedia's editors did something different.
WikiProject AI Cleanup was co-founded by France-based editor Ilyas Lebleu. The group reviewed over 500 articles suspected of being AI-generated.
No GPTZero. No watermark scanning. No algorithms.
They just read. Thousands of hours. And they documented what they found.
The patterns are specific, searchable, and countable. Here are the ones that show up everywhere.
Em dashes are the easiest tell. Humans use commas and parentheses. AI uses em dashes for everything. Two per 500 words is normal. Six is a red flag.
Inflated language is next. "Stands as a testament." "A watershed moment." "Underscores its importance." If your page reads like a museum plaque, something went wrong.
Then there's the rule of threes. Three adjectives. Three list items. AI defaults to three because its training data is full of persuasive writing that uses tricolon.
Certain words show up at 10x human frequency. "Delve." "Tapestry." "Pivotal." "Foster." "Multifaceted." The words aren't wrong. They're just overused.
Compulsive summaries are another giveaway. "Overall, the importance of..." AI restates what it just said because "helpful" in training data often means "redundant."
And then negative parallelisms. "It's not just about X, it's about Y." Sounds deep. Says nothing.
The full guide covers dozens more across five categories: language, style, communication artifacts, markup, and citations. All documented with examples from real flagged articles.
Google's E-E-A-T framework rewards exactly this kind of hands-on expertise.
Here's where it gets interesting
Siqi Chen goes by @blader on GitHub. He read Wikipedia's detection page and realized something simple:
If you tell AI exactly what makes it sound like AI, it stops doing those things.
So he built Humanizer. It's a Claude Code skill. Free. MIT-licensed. Takes a couple minutes to set up.
How it works:
- Point it at your text
- It scans for Wikipedia's detection patterns
- It rewrites every flagged section while keeping the meaning
- It runs a final audit: "What still sounds AI-generated?" Then fixes that too
Here's the uncomfortable part.
The specificity that makes Wikipedia's guide good for detection makes it equally good for evasion.
Wikipedia published a "how to not get caught" guide. They just didn't frame it that way.
See what's actually working on your site.
Get your full site analysis →The before and after (this is the part that matters)
Talk is cheap. Here's what the transformation looks like.
Take inflated symbolism. Before: "The Statistical Institute of Catalonia was officially established in 1989, marking a pivotal moment in the evolution of regional statistics in Spain." After: "The Statistical Institute of Catalonia was established in 1989 to collect and publish regional statistics independently from Spain's national statistics office." Same information. Half the words.
Or vague attribution. Before: "Experts believe it plays a crucial role in the regional ecosystem." After: "The river supports several endemic fish species, according to a 2019 survey by the Chinese Academy of Sciences." "Experts believe" means "I made this up." A named source with a date means "I did the work."
Or promotional fluff. Before: "Nestled within the breathtaking region of Gonder in Ethiopia, Alamata Raya Kobo stands as a vibrant town with a rich cultural heritage and stunning natural beauty." After: "Alamata Raya Kobo is a town in the Gonder region of Ethiopia, known for its weekly market and 18th-century church."
Every rewrite does the same thing. Cut the performance. Add the specifics.
Most people miss this part: the "fixed" version is also better writing. More useful to a reader. More credible to Google. More likely to get cited by AI search engines.
The detection patterns ARE the bad writing patterns. Fix one, you fix the other.
"But doesn't this make detection useless?"
No. And here's why.
Wikipedia published their framework on a public wiki page. It was going to get reverse-engineered. That was always going to happen.
But look at what the patterns actually are. Em dashes. Inflated language. Compulsive summaries. Vague attributions.
These aren't secret codes only AI produces.
They're bad writing habits. AI just produces them at scale.
Removing them doesn't make AI deceptive. It makes AI output less annoying to read. An article without "stands as a testament" and "vibrant tapestry of cultures" is a better article. Period. Doesn't matter who fixed it.
The question was never "can we detect AI?"
The question was always "does this content add value?"
Google's helpful content system doesn't run an AI detector. It checks whether content has experience, expertise, and originality. Content loaded with AI patterns fails that test. Not because it's AI. Because it's generic.
Why this costs you money
This isn't a writing style problem. It's a revenue problem.
Ahrefs found that only 12% of AI-cited URLs rank in Google's top 10 for the same query. The overlap ranges from 6% for ChatGPT to 29% for Perplexity.
AI search engines pick their own sources. They prefer content that sounds like a person who knows the subject wrote it.
Content that reads like unedited ChatGPT output gets skipped. By Google. By AI search. By the reader who bounces in three seconds.
We see this in our own SEO audit reports. Sites with clean, specific writing score higher on content quality. Sites that read like raw AI output score lower. Every time.
Every day your content sounds like AI, you're losing clicks to competitors whose content doesn't.
Two ways to fix it
Option 1: Install Humanizer
If you use Claude Code, clone the repo:
git clone https://github.com/blader/humanizer.git ~/.claude/skills/humanizer
Claude Code auto-discovers it. Point it at your content. It runs Wikipedia's patterns as a rewriting checklist. Done.
Option 2: Do it yourself
No Claude Code? Here's the manual version. Open your last blog post.
Ctrl+F these words: "delve," "tapestry," "landscape," "foster," "moreover," "furthermore," "pivotal." Three or more hits? Flag the piece.
Count your em dashes. More than two per 500 words means most need to go.
Check if every list has exactly three items. Real writing varies the rhythm.
Find anything that "stands as a testament" or "underscores its importance." Delete it. Say what actually matters instead.
Delete the last paragraph. If you don't lose any information, leave it deleted.
Read the first two paragraphs out loud. Do you sound like a person? Or a press release?
Last one. Is there a single specific story, number, or example from your real work anywhere in the piece? If not, the piece fails E-E-A-T. Doesn't matter what else you did.
Option 3: Have someone do it for you
If you'd rather skip the checklist entirely, we write AI-optimized blog content that passes every one of these checks before it ships. Same patterns, same audit, already done.
The patterns will change. The principle won't.
Models will get better at sounding human. Wikipedia's editors will find new tells. New tools will get built.
But generic content that adds nothing original will keep getting filtered out. By readers, by Google, and by AI search engines picking what to cite.
The fix is always the same. Write about what you know. Have a point of view. Put in the specific, messy details from your actual work that a language model can't make up.
If you use AI to draft, edit against the checklist until it sounds like you. Not like a press release.
Our content strategy guide covers what makes content rank in 2026. If you want content that's already been through the checklist, check out our blog writing packages.
Need blog content that doesn't read like AI?
We write AI-optimized articles that pass every check on this list before they ship. Strategy, writing, images, publishing.
See blog writing packages →
