AO3 Readers Found Claude Code Hidden in Fan Fiction — Then the Community Turned on Itself

Fan-fiction readers hunting for undisclosed artificial intelligence use on Archive of Our Own uncovered something far more concrete than suspiciously slick prose: hidden Claude chatbot code embedded in dozens of posted stories.

On 29 June, an anonymous X account released a 25-page document titled 'Fandom Has a Hidden Generative AI Problem'. It described how text copied directly from Anthropic's Claude chatbot into AO3's rich-text editor could retain distinctive HTML formatting that ordinary readers would not see.

The document identified the formatting in 38 works. Many came from the rapidly expanding Heated Rivalry fandom, although works connected to 'The Pitt', 'Pluribus', 'Off-Campus' and 'Stranger Things' were also included.

Its authors said they wanted writers to disclose their AI use so readers could make informed choices. They also condemned harassment and ostracism of the people identified in the dossier. Those qualifications did little to stop the document from becoming evidence in a growing argument over authorship, honesty and who gets to oversee fan spaces.

What The Claude Trace Proved

The discovery was more persuasive than conventional AI detection because it did not depend on judging a writer's style. The document supplied an AO3 'site skin' – custom code that changes how the archive appears to an individual user – which could reveal the Claude-specific HTML wrapper.

Its presence indicated that text had apparently travelled directly from Claude into AO3's rich-text editor. Some stories reportedly contained only a few instances, while others had hundreds.

But the method was far from a universal AI detector. It could not find Claude output that had first been pasted into another document, text from which the wrapper had been removed, or material created through ChatGPT, Gemini or another model. It also could not establish why Claude had been used or determine how much of an entire story was generated by AI.

One writer later said Claude had been used for proofreading and formatting rather than generative writing. Another reportedly acknowledged using AI instead of a human beta reader. Those explanations revealed the difficulty of treating every detectable AI interaction as equivalent to outsourcing an entire story.

The dossier nevertheless named specific stories, turning a narrow technical finding into a public judgement about their writers. Some readers felt vindicated after defending authors who had previously denied using AI.

Others warned that publishing lists would invite dogpiling regardless of the document's request that fans refrain from harassment. More works were subsequently deleted or restricted, and at least one writer temporarily disabled comments after receiving death threats, Fansplaining reported.

That was the point at which an investigation intended to restore transparency began deepening the community's distrust. Readers had evidence of a particular Claude workflow, but no shared process for deciding what that evidence meant or what consequences should follow.

Why Disclosure Matters

The intensity of the reaction reflects how fan fiction functions. Stories are commonly shared without payment, with readers returning comments, recommendations and encouragement. AI therefore raises more than questions about the perceived quality of a work: it unsettles the expectation that the exchange is taking place between people.

A study of 157 active fan-fiction readers and writers found that 86% believed authors should disclose AI involvement. Among 155 respondents asked how they would feel after retrospectively learning that a story was AI-written, 72.2% reported negative feelings and 58% said they would feel deceived.

However, 66.2% were unsure that they could accurately identify AI-generated writing. The researchers consequently proposed nuanced, voluntary disclosure systems rather than automatic detection, reflecting the gap between the community's demand for transparency and its ability to prove undisclosed use.

The English-language sample also skewed towards North American participants, limiting how widely its findings can be generalised.

AO3's rules leave that gap largely to fans themselves. The platform permits fully or partly AI-generated works when they otherwise qualify as fanworks, citing its commitment to 'maximum inclusivity'. It encourages optional tags including 'Created Using Generative AI' and 'AI-Generated Text', but has not made AI disclosure mandatory.

The Organization for Transformative Works said in June that generative-AI works remained allowed because a ban would be difficult to enforce and could expose suspected writers to further abuse.

Its Policy and Abuse team cannot reliably determine whether a work was AI-generated, while general detection tools risk false positives that disproportionately affect inexperienced, neurodivergent and non-native English writers.

The Claude trace briefly gave readers something those broad detectors could not: evidence of one direct route from chatbot to archive. Yet publicising the method also showed writers how easily it could be defeated by using another document or removing the formatting.

The hidden code may now become less common. The collapse of trust that drove readers to search for it will be considerably harder to address.