The fanfiction world just turned on itself over AI, and the weapon is a homemade Claude detector. According to The Verge AI, an anonymous X account called @heatedrivalryai posted a browser “skin” for Archive of Our Own (AO3) on June 29th that flags stories written with Anthropic’s Claude. When it detects the bot, it floods the entire page background in bright red. The Verge AI tested it and confirmed it works.
Here’s the mechanism. When you paste text straight from Claude into AO3’s editor, the bot leaves behind a hidden code artifact, a wrapper called “font-claude-response-body.” The skin scans for that tag. Present means Claude was used. The Verge AI ran its own experiment: pasting directly from the chatbot turned the screen scarlet, while the same story routed through another app came back clean. As red flags go, it’s impossible to miss.
Why this landed so hard
Creative communities have distrusted ChatGPT, Claude, and the rest for a long time. But detection has always been guesswork, readers hunting for “tells” like em dashes, purple prose, or the now-infamous “it’s not X, it’s Y” sentence structure. This tool promised something the guesswork never could: a definitive signal. That’s the appeal, and that’s the danger.
Within days, fandom members used the flags to publicly name and shame flagged writers. The creator says the tool was meant to prove the method works, not to “create an environment of mistrust.” That line didn’t hold. For a slice of the community, any AI use at all is a betrayal, whether over environmental cost or the fact that models were trained by scraping the open web, fanworks included.
The detector’s real limits
What stands out here is how narrow the tool actually is, and how confidently people are wielding it anyway. A few problems, per The Verge AI:
- It only catches direct pastes. Copy from Claude into Google Docs or Word first, then move it to AO3, and the artifact disappears. Nothing gets flagged.
- It can’t measure how much AI was used. That red screen could mean a fully machine-written story, or an author who ran two human sentences through Claude for spell-check and pasted them back.
- It’s trivial to evade. Flagged writers are already stripping the artifacts, and future works dodge it easily.
- It covers one model on one platform. AO3 isn’t the only home for fanfic, and Claude is one bot among many.
So you get false negatives for the careful and public shaming for the careless, with no way to tell a heavy user from someone who fixed their grammar.
The bigger signal for AI
This is significant because it exposes where AI detection actually stands. The Verge AI, which has covered this beat for years, is blunt: there is no reliable way to tell generated text from human writing. Watermarking systems like C2PA Content Credentials and Google’s SynthID are making progress on images, video, and audio, but those signals don’t survive a copy-paste of plain text. The AO3 skin isn’t detecting AI writing at all. It’s detecting a sloppy paste that left metadata behind. Clean up the paste and the “detector” is blind.
Anthropic didn’t respond to The Verge AI’s request to confirm how the artifact works. Google and OpenAI didn’t say whether their models leave similar traces. One person claims to have code that spots Claude, DeepSeek, and some ChatGPT output, but hasn’t released it or explained how.
There’s a reason AI companies may eventually solve this themselves. Early models trained on scraped internet text, and as synthetic writing floods the web, they risk “model collapse,” where training on AI output degrades future output. Detecting their own generations internally protects the product.
What to watch
For now, fandom is policing itself on vibes and one brittle tool. Expect three things: writers changing their paste habits to slip past detection, the shaming to outrun the accuracy, and copycat detectors that lean on the same fragile metadata tricks. The uncomfortable truth underneath it all is that AI “writes like that” because it was trained on how real people write. Plenty of florid, overwrought human fanfic from the pre-ChatGPT era would fail this same sniff test. Full details are at the original report from The Verge AI.