Fanfiction’s AI purge sparks detector doubts and collateral flags

AO3 Claude – A new “Claude detector” skin for AO3 has gone viral in fan communities, turning pages red when a specific artifact is detected. But the method depends on how text is pasted, can’t measure how much AI was used, and has already led to public naming and shaming—r

For the third time in a week, the fanfiction page suddenly felt less like a story and more like a crime scene.

A new fanworks movement has been gaining momentum, aimed at rooting out authors using generative AI. The flare-up centers on a so-called detector built as a skin for Archive of Our Own (AO3), designed to identify “coding artifacts” left behind by Anthropic’s Claude bot.

It all began on June 29th, when an anonymous X account, @heatedrivalryai, posted the method it claimed would work. The account said that when a Claude-generated response is pasted directly into AO3 from Claude. the text is wrapped by a Claude-injected code called “font-claude-response-body.” In the account’s description. the presence of that code is what the detector treats as proof. When a user visits a page with the skin enabled, the entire background turns red.

To test whether that claim held up, several test posts were published to AO3. One experimenter says the screen turned red immediately when the skin was tested against those examples. and a Claude-generated short story was also published for the same purpose. The red screen appeared when the Claude response was pasted straight into the AO3 editor. It vanished when the same kind of generated text—pasted in a way that didn’t come straight from Claude—was used instead.

The detector post came with examples meant to demonstrate how the system behaves. and the anonymous creator insisted the intention wasn’t to “create an environment of mistrust or accuse particular users.” Still. fandom moved fast once the tool existed. Fanfic communities have mobilized to publicly name and shame writers whose works were flagged by the red-screen skin. The creator behind the detector framed AI as a threat to what fandom is supposed to be.

“Fandom is a uniquely connective, collaborative space. It thrives on the human element and the creative spark which drives it and feeds off it,” the @heatedrivalryai account said. “If we unknowingly allow AI to corrupt these spaces, what will be left of them?”

Anthropic did not respond to a request to verify whether the fan-made Claude detector works as described. But the method, as described, does look technically plausible—there’s “no apparent reason” for the Claude-injected code to appear in a story if the bot wasn’t used somehow.

The problem is what the detector can’t reliably prove.

The code wrapping is only preserved if text is copied directly from Claude into AO3’s editor. That means it won’t catch text that is edited in tools such as Google Docs or Microsoft Word and then moved into AO3. For someone writing for a living. the risk of pasting content into a content management system without it behaving exactly like “direct copy from Claude” is described as all too familiar. And that limitation cuts both ways: some writers who were flagged have already updated their works to remove the artifacts. and future writes can potentially evade detection.

Even when the red screen appears. it still doesn’t answer a bigger question: how heavily Claude was used in the work. The tag doesn’t show whether an entire story was generated. or whether a few human-written sentences were pasted into Claude for spell-checking or translation and then returned to AO3. For some fandom members, that distinction doesn’t matter. They treat any use of generative AI as an inexcusable betrayal.

That harsh view is fueled by multiple concerns beyond detection itself. Many people cite worries about the environmental impact of the technology and how models are trained by scraping the open web—a process that likely includes fanworks uploaded to platforms like AO3.

But this particular tool has a built-in narrowness. AO3 is not the only platform for fanworks, and Claude is only one of many AI models. Another person has claimed they wrote separate code that can detect “Claude. Deepseek. and some ChatGPT” usage. but they haven’t released the solution publicly or explained how it works.

The broader search for “traceable artifacts” is also stalling elsewhere. A request was made to Google and OpenAI to ask whether their models leave any traceable artifacts in text generation that could be detected by similar means, and they did not respond.

The underlying belief driving the community conflict—that there must be a universally reliable way to separate machine output from human typing—runs into hard limits. The write-up points to years of reporting on AI detection problems and says there is. to the best of knowledge. no reliable technological solution today for distinguishing generated text from human-written text in copy-pasted form. Systems such as C2PA Content Credentials and Google’s SynthID are described as making progress for images. videos. and even audio. but they depend on invisible watermarks and metadata that don’t survive copy-paste into plain text.

And that’s where the conflict turns from technical debate into a credibility crisis.

AI companies have an incentive to solve internal detection problems—early models were trained on text scraped indiscriminately from the internet. and the write-up notes that a shift away from human writing could lead to a “model collapse” scenario where output accuracy degrades. Still. for now. fandom is largely relying on “vibes” and stylistic cues. from specific sentence structures such as “it’s not X. it’s Y” to patterns of flowery metaphors. Even the mention of a joke-like example (“At least nobody in fandom. so far. has benches becoming men.”) lands as a reminder that these guesses rarely behave like evidence.

The write-up argues that AI writes the way it does because it learned from real people’s writing. It’s trying to replicate what it saw. That leads to an uneasy takeaway: writers could be judged as “robotic” based on patterns that were already part of human fandom long before ChatGPT.

If there is a solution available on AO3. it’s not a skin—it’s the site’s “robust tagging system.” A “Created Using Generative AI” tag exists. and many authors include it to disclose the use of tools like Claude. That requires honesty, however, and honesty can be hard to find in a climate where backlash is immediate and blunt. Fanfiction is also framed as a hobby rather than a regulated industry.

In the end, the detector’s scarlet verdict risks becoming a shortcut people use to protect themselves from uncertainty. The write-up warns that efforts to stop AI from taking attention from human creativity can still sweep up innocent writers—especially when the person who edits or formats a fic uses Claude without the author realizing.

At least one writer is described as having been caught up in that exact situation, where a trusted editor used Claude to edit their story.

So if the next fanfic you read feels a little robotic, the only safe conclusion is the simplest one: it might be a robot, but it might also be someone caught in the crossfire of a tool that can only see one kind of trace—and nothing else.

fanfiction AI Claude AO3 generative AI detection Anthropic @heatedrivalryai cybersecurity-like detection digital communities