The fanfiction community faces internal conflict over AI involvement.
In the past week, a new movement has emerged within fan communities, aiming to identify authors who utilize generative AI in their writing. However, the methods being employed to detect AI-generated content raise concerns, and many fanfic writers could find themselves unfairly targeted.
The creative community, particularly in fanfiction, has long expressed skepticism about tools like Claude and ChatGPT. Fans have shared tips for identifying AI-generated works, pointing to stylistic elements like em dashes and the concept of “purple prose.” On June 29, an anonymous account on X, known as @heatedrivalryai, introduced a tool that claims to offer a more reliable detection method. This tool is essentially a skin for the popular fanfic archive Archive of Our Own (AO3), designed to identify coding artifacts associated with Anthropic’s Claude bot.
According to the @heatedrivalryai account, when text generated by Claude is pasted into AO3, it carries a specific code labeled ‘font-claude-response-body.’ The presence of this code is said to confirm the use of Claude. When users visit a fanfic page containing this code, the background turns red, signaling the detection.
Several test posts have been uploaded to AO3 to verify the tool’s effectiveness. I tried it out myself, and the background indeed turned red when I pasted text directly from Claude. However, if I pasted the same text that had been copied elsewhere, the red screen did not appear.
The post introducing the Claude detector included examples of fanfics where these artifacts were found, with the creator emphasizing that the goal was to demonstrate the tool’s functionality, not to foster distrust among users. Despite this, some members of the fan community have quickly begun to publicly call out writers whose works were flagged, and the creator of the tool has expressed a strong aversion to AI. They stated, “Fandom is a uniquely connective, collaborative space. It thrives on the human element and the creative spark which drives it and feeds off it. If we unknowingly allow AI to corrupt these spaces, what will be left of them?”
Anthropic has not responded to inquiries about the accuracy of the fan-made detector. However, the methodology appears sound, and my own tests support its claims. The presence of the Claude code seems to indicate that the bot was used in some capacity. Still, there are significant risks of false negatives and overgeneralizations.
The detection code is only preserved when text is copied directly from Claude into AO3’s editor, meaning it won’t catch anything that has been edited in other programs like Google Docs or Microsoft Word before being pasted into AO3. As a writer, I can attest to the challenges of composing directly in a content management system. Some authors flagged by the tool have already revised their works to eliminate the artifacts, and future submissions can easily bypass detection.
Moreover, the detection method does not indicate the extent to which Claude was utilized in a work. A red screen could suggest that an entire story was generated by AI, or it might simply mean that an author used Claude for minor edits or translations before transferring the text back to AO3.
This distinction is lost on some fans, who view any use of generative AI as a betrayal of the creative community. Concerns about the environmental impact of AI and its reliance on data scraped from the web, including fanworks, are frequently cited.
The tool’s effectiveness is limited, as AO3 is not the sole platform for fanfiction, and Claude is just one of many AI models. Some claim to have developed separate code that can detect usage of Claude, Deepseek, and certain ChatGPT outputs, but this solution has not been made public, nor has its functionality been explained. Inquiries to Google and OpenAI about whether their models leave detectable artifacts have gone unanswered.
It would be surprising if a universally reliable detection system existed. After years of covering AI detection issues, I can confidently say that no dependable technological solution currently distinguishes AI-generated text from that written by humans. While systems like C2PA Content Credentials and Google’s SynthID are making strides in identifying generative AI in images and videos, they rely on invisible watermarks and metadata that do not transfer to copied text.
This situation may evolve, as AI companies have strong incentives to address these challenges internally. Early AI models were trained on text indiscriminately gathered from the internet, and as human writing is overshadowed by synthetic content, there’s a risk of a “model collapse” that could compromise output quality.
For now, fan communities largely rely on subjective judgment. Most fanfics are evaluated based on stylistic “tells,” which can range from specific sentence structures to an overabundance of flowery language. It’s important to remember that AI often mimics these styles because it has been trained on human-written content. It’s attempting to replicate our voices. While I won’t share my own AO3 bookmarks, I can recall reading some overly elaborate fanfics from the pre-ChatGPT era that would likely fail this dubious detection test.
The most effective way to identify AI-generated works on AO3 is through the site’s tagging system. A “Created Using Generative AI” tag exists, and many authors use it to disclose their use of tools like Claude. However, this requires honesty, and the backlash against AI use creates little incentive for transparency. It’s also essential to remember that fanfiction is meant to be a hobby, not a regulated industry.
As efforts to keep AI from overshadowing genuine human creativity continue, authors who don’t meet the community’s standards may become unintended victims of this ongoing scrutiny. At least one writer has already faced backlash because an editor they trusted used Claude to revise their work. So, if the next fanfic you dive into feels a bit robotic, keep in mind that it might not be the product of a machine.
