The Telltale Signs of AI Writing, According to Wikipedia’s Editors

Since 2023, a group of Wikipedia’s volunteer editors has been combing new submissions with one question in mind: did a person write this? Their working notes became a public field guide, “Signs of AI writing,” maintained by the editors of WikiProject AI Cleanup. TechCrunch’s Russell Brandom called it “the best resource I’ve found” for checking whether a suspicion about machine-written prose holds up.

This list is for the people who now do that checking for a living, whether they signed up for it or not: teachers grading essays, editors reading pitches, hiring managers sorting cover letters. We’ve picked the habits that the guide and the reporting around it treat as most telling, with an example of each, then ended on the part that matters most: what these signs can’t prove.

A disclosure, once. Our own style guide bans most of what follows. We hold our copy to this list too.

1. Everything is “pivotal”: inflated significance

The clearest tell is a reflex to explain why a subject matters, in terms so general they could apply to anything. According to TechCrunch’s reading of the guide, AI submissions lean on phrases like “a pivotal moment” or “a broader movement.”

Ilyas Lebleu, a volunteer editor in France who helped start the cleanup project, offered two more on NPR’s Morning Edition: “Some signs like rich cultural heritage or enduring legacy.” Picture a two-paragraph entry about a small-town bakery that somehow “stands as a testament to the region’s enduring legacy.” The bakery sells bread. The sentence is selling something else.

2. The trailing “-ing” clause that claims meaning

This one is subtler and, once you notice it, everywhere. A sentence states a fact, then adds a participle phrase announcing what the fact signifies. The guide’s examples, as TechCrunch quoted them: an event is “emphasizing the significance” of something, or “reflecting the continued relevance” of some general idea.

Try it on a sentence: “The city opened a new library in 2019, reflecting its continued commitment to education.” The first half is checkable. The second half is a claim nobody sourced. Brandom admitted it is “a bit hard to pin down, but once you can recognize it, you’ll see it everywhere.”

3. Prose that reads like a TV commercial

Language models absorbed a great deal of marketing copy, and it shows. In the guide’s world of AI text, Brandom wrote, “landscapes are always scenic, views are always breathtaking, and everything is clean and modern.” The editors’ own verdict on that register: “it sounds more like the transcript of a TV commercial.”

For an encyclopedia, that’s disqualifying: as NPR put it, Wikipedia is supposed to read like an encyclopedia “without opinions or promotional language.” For everyone else it’s a useful test. Strip the adjectives from a suspicious paragraph. If almost nothing is left, the adjectives were doing the work.

4. A résumé of minor press mentions

AI-written entries often pad a subject’s importance by listing small media appearances: a quote in a local paper, a podcast spot, a panel. TechCrunch noted this is “the kind of thing you’d expect from a personal bio, but not from an independent source.”

Hiring managers will recognize the pattern from cover letters. A list of places someone has been mentioned stands in for evidence of what they’ve done.

5. Sources and people that don’t exist

This is where style stops being a matter of taste. While combing new submissions, the cleanup volunteers found “errors, fake sources and people in places that were made up,” NPR reported.

Of everything on this list, it’s the one sign you can actually check. Click the citation. Search for the book. Look up the expert. A reference that leads nowhere is stronger evidence than any turn of phrase, and it’s a problem whoever wrote the text.

6. A voice that doesn’t fit the room

The project began with a mismatch, not a keyword. “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,” Lebleu told NPR. In 2024, Lebleu told 404 Media that a few editors “managed to replicate similar ‘styles’ using ChatGPT,” and that “discovering some common AI catchphrases allowed us to quickly spot some of the most egregious examples.”

The broader lesson holds outside Wikipedia. Every venue has a house register: a lab report, a student’s earlier essays, a newsroom’s wire copy. Text written to sound like everything at once tends to sound like nothing in particular.

7. Single “AI words” are weak evidence

A word of caution about the most famous tell. For a while, Brandom wrote, people were convinced that words like “delve” or “underscore” gave a model away, “but the evidence is thin,” and the telltale words have become harder to trace as models improve. The guide, by his account, focuses instead on habits and turns of phrase that are rare on Wikipedia but common on the wider internet, which is where the models learned to write.

Treat a vocabulary list the way you’d treat a single typo. It’s a reason to read more closely, not a finding.

8. Detector software is not a shortcut

On detection software, Brandom read the guide as confirming that automated tools are “basically useless.” OpenAI made a similar call about its own product. Its AI text classifier correctly flagged only 26% of AI-written text and wrongly labeled human writing as AI 9% of the time, and the company pulled it on July 20, 2023, citing “its low rate of accuracy.”

Nine percent sounds small until you apply it to a lecture hall of 200 honest students. That’s roughly 18 false accusations.

The catch: a sign is not a verdict

Every item above describes something human writers do too. People write ad copy. People pad their bios. Plenty of students were taught to end each paragraph by explaining its significance. On Wikipedia, a suspect article sits inside a system where other volunteers can check it against its sources and fix it. A teacher or employer acting on a hunch has no such backstop, and the cost of a wrong accusation lands on the person accused.

Our read: use the list to ask better questions, not to reach a conclusion. Ask a student to talk through a draft. Ask a candidate about a claim in their letter. Check every citation. Those steps catch fabrication whether a model or a person produced it.

And expect the signs to fade. The guide’s popularity has made it a style sheet for evasion: Humanizer, an open-source AI skill with tens of thousands of GitHub stars, says its pattern framework is sourced from the Wikipedia guide and uses it to rewrite AI text so it no longer trips those patterns. Brandom predicted the habits “can be disguised, but it will be hard to do away with them completely.”

NPR’s segment on the guide signed off with a joke only a human would bother to make. “In summary,” host Michel Martin said, “what you just heard was not written by AI.”

Sources

• Wikipedia: Signs of AI writing
• TechCrunch: The best guide to spotting AI writing comes from Wikipedia (Russell Brandom, Nov. 20, 2025)
• NPR Morning Edition: Wikipedia editors publish new guide to help readers detect entries written by AI (Sept. 4, 2025)
• 404 Media: The Editors Protecting Wikipedia from AI Hoaxes (Oct. 9, 2024)
• OpenAI: New AI classifier for indicating AI-written text (discontinued July 20, 2023)
• GitHub: blader/humanizer