An anonymous X account called @heatedrivalryai released a site skin for Archive of Our Own on June 29th that claims to definitively identify fanfiction written with Claude. The skin turns the entire page background red when it detects a specific code artifact — 'font-claude-response-body' — that Anthropic's chatbot injects into text pasted directly from its interface into AO3's editor. Within days, fandom communities began using the tool to publicly name authors whose works triggered the flag.
The detection method is technically sound, at least for the narrow case it covers. Direct testing confirms the skin fires reliably when Claude output is pasted straight into AO3 and stays quiet when the same text is routed through a different editor first. There is no obvious reason for the Claude code wrapper to appear in a fanwork if the model was not involved somewhere in the chain.
The problem is the chain. The tag only survives a direct copy-paste from Claude to AO3's editor. Any author who moves the text through Google Docs, Microsoft Word, or effectively any other editor first — which is standard practice for anyone who writes seriously — strips the artifact entirely. Several flagged authors have already updated their works to remove the code, and any future user aware of the detector can evade it in one extra step.
“When a Claude-generated response is pasted directly into AO3 from Claude, the text is wrapped by a Claude-injected code 'font-claude-response-body,'”— @heatedrivalryai, creator of the AO3 Claude-detector skin
Key facts
- 01An anonymous X account, @heatedrivalryai, released a browser-style skin for Archive of Our Own on June 29th that flags Claude-generated text.
- 02The skin works by detecting a leftover 'font-claude-response-body' tag that [Claude](/claude) injects when its output is pasted directly into AO3's editor.
- 03Any author who routes Claude text through Google Docs or Microsoft Word first strips the tag, evading the detector entirely.
- 04AO3 already offers a 'Created Using Generative AI' tag for voluntary disclosure, but adoption relies on authors self-reporting.
- 05No current AI detection system, including C2PA Content Credentials or Google's SynthID, reliably identifies AI-generated text after copy-paste.
The false-negative rate is one issue. The false-positive interpretation is a bigger one. The presence of the tag reveals nothing about how heavily Claude was used. A fully AI-generated story throws the same red screen as one where an author pasted a paragraph into Claude for a spell-check or translation pass and moved it back. Fandom members treating the flag as proof of wholesale AI authorship are making an inference the tool cannot support.
Anthropic did not confirm whether the code wrapper functions as described, and the detector's creator says the goal is not to accuse individual users. That framing has not held. Flagged authors have been named publicly, and the tag is being read as a binary verdict rather than a signal that Claude touched the document at some point.
The creator's stated motivation is protective rather than punitive.
Beyond AO3, the detector has no reach. It does not cover other fanfiction platforms, and it does not cover other models. One person claims to have written separate code that flags Claude, DeepSeek, and some ChatGPT output, but has not released it or explained the method. Google and OpenAI have not said whether their models leave comparable artifacts in generated text.
A universally reliable text detector does not currently exist. Systems like C2PA Content Credentials and Google's SynthID are making progress on identifying AI-generated images, video, and audio, but those approaches lean on invisible watermarks and metadata that do not survive copy-paste in plain text. Text detection remains an unsolved problem, and the AO3 skin is not a solution — it is an accidental fingerprint from one specific product's front-end.
AI companies have their own reasons to solve this internally. As synthetic text crowds out human writing on the open web, models trained indiscriminately on scraped data risk model collapse, where output quality degrades because the training corpus is increasingly recursive. Reliable provenance for generated text is a self-preservation question for the labs, not just an enforcement question for fandoms.
AO3 already offers a 'Created Using Generative AI' tag that authors can apply voluntarily. Adoption depends on transparency, and transparency is scarce when the community response to disclosure is a public shaming campaign. The incentive structure the detector has created works against the disclosure mechanism that would actually address the concern.
The broader dynamic here matters more than the specific skin. A hobbyist community has decided that AI use is a betrayal serious enough to warrant surveillance tooling, and it has adopted the first available signal — regardless of what that signal can and cannot prove. Writers who never used AI but happen to write in flowery prose are already being accused on vibes alone. The detector adds a technical veneer to a judgment that was mostly aesthetic to begin with.
For Anthropic, this is a minor UX artifact turning into a reputational surface it did not choose. The 'font-claude-response-body' tag is a small front-end detail that will likely disappear in a future update, and once it does, the detector stops working entirely. What remains is the community's appetite for enforcement, which will outlast any specific technical tell. Every model provider now has a preview of what happens when their product's exhaust ends up in a space that treats its presence as evidence of a crime.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
