Image Credits:Riccardo Milani / Hans Lucas / Hans Lucas via AFP / Getty Images8:32 AM PST · November 20, 2025
We’ve each felt nan creeping suspicion that thing we’re reference was written by a ample connection exemplary — but it’s remarkably difficult to pin down. For a fewer months past year, everyone became convinced that circumstantial words for illustration “delve” aliases “underscore” could springiness models away, but nan grounds is thin, and arsenic models person grown much sophisticated, nan telltale words person go harder to trace.
But arsenic it turns out, nan folks astatine Wikipedia person gotten beautiful bully astatine flagging AI-written prose — and nan group’s nationalist guideline to “Signs of AI writing” is nan champion assets I’ve recovered for nailing down whether your suspicions are warranted. (Credit to nan writer Jameson Fitzpatrick, who pointed retired nan archive connected X.)
Since 2023, Wikipedia editors person been moving to get a grip connected AI submissions, a task they telephone Project AI Cleanup. With millions of edits coming successful each day, there’s plentifulness of worldly to tie on, and successful classical Wikipedia-editor style, nan group has produced a section guideline that’s some elaborate and dense connected evidence.
To commencement with, nan guideline confirms what we already know: automated devices are fundamentally useless. Instead, nan guideline focuses connected habits and turns of building that are uncommon connected Wikipedia but communal connected nan net astatine ample (and thus, communal successful nan model’s training data). According to nan guide, AI submissions will walk a batch of clip emphasizing why a taxable is important, usually successful generic position for illustration “a pivotal moment” aliases “a broader movement.” AI models will besides walk a batch of clip detailing insignificant media spots to make nan taxable look notable — nan benignant of point you’d expect from a individual bio, but not from an independent source.
The guideline flags a peculiarly absorbing quirk astir tailing clauses pinch hazy claims of importance. Models will opportunity immoderate arena aliases item is “emphasizing nan significance” of thing aliases other, aliases “reflecting nan continued relevance” of immoderate wide idea. (Grammar nerds will cognize this arsenic nan “present participle.”) It’s a spot difficult to pin down, but erstwhile you tin admit it, you’ll spot it everywhere.
There’s besides a inclination towards vague trading language, which is highly communal connected nan internet. Landscapes are ever scenic, views are ever breathtaking, and everything is cleanable and modern. As nan editors put it, “it sounds much for illustration nan transcript of a TV commercial.”
The guideline is worthy reference successful full, but I came distant very impressed. Before this, I would person said that LLM prose was processing excessively accelerated to pin down. But nan habits flagged present are profoundly embedded successful nan measurement AI models are trained and deployed. They tin beryllium disguised, but it will beryllium difficult to do distant pinch them completely. And if nan wide nationalist gets much savvy astir identifying AI prose, it could person each sorts of absorbing consequences.
Russell Brandom has been covering nan tech manufacture since 2012, pinch a attraction connected level argumentation and emerging technologies. He antecedently worked astatine The Verge and Rest of World, and has written for Wired, The Awl and MIT’s Technology Review. He tin beryllium reached astatine russell.brandom@techcrunch.com aliases connected Signal astatine 412-401-5489.
17 hours ago
English (US) ·
Indonesian (ID) ·