In recent days a fresh campaign has surfaced among fan creators seeking to identify those who rely on generative AI for their stories. Yet the approaches used to spot such works invite skepticism and risk sweeping up innocent writers in the process.

Longstanding aversion to tools like Claude and ChatGPT already runs deep across creative circles that include fanfiction. Participants have circulated advice on recognizing supposed AI output by pointing to elements such as em dashes or overly ornate language. On June 29 an anonymous account on X named @heatedrivalryai unveiled what appeared to be a more dependable option. The post shared a custom skin for Archive of Our Own that claims to flag hidden markers left by Anthropic's Claude model.

The account explained that text copied straight from Claude into AO3 carries a specific wrapper labeled font-claude-response-body. Its appearance signals Claude involvement without question. When the skin detects this marker on a story page the background shifts to red.

Sample entries on AO3 let users verify the skin's function. Direct tests confirmed the red background on flagged examples. A separate trial with a freshly generated Claude story produced the same result when pasted directly yet showed no change when the identical text was inserted without the original wrapper.

The announcement included instances of detected markers and stressed the goal was to illustrate the method rather than foster suspicion toward individuals. Communities nevertheless began publicly calling out authors whose works triggered the flag. The account's creator voiced clear opposition to AI noting that fandom depends on human connection and that unchecked AI risks eroding that foundation.

Anthropic offered no comment on the detector's accuracy. The underlying logic appears consistent and independent checks support its function. Nothing suggests the marker would appear without some use of the bot. Still both missed detections and broad assumptions remain possible.

The marker survives only with direct copy and paste so it misses text first revised in other programs before reaching AO3. Many writers prefer external editors to avoid platform risks. Several flagged authors have already stripped the markers from existing works while new stories can avoid the skin entirely.

The indicator also fails to show the extent of Claude involvement. A full red screen might reflect complete AI generation or merely a few sentences run through the tool for editing or language support.

Some community members treat any AI assistance as unacceptable regardless of degree. They cite environmental costs and concerns over training data scraped from public sites including fan archives.

The tool's reach stays narrow since AO3 is one platform among many and Claude represents only a single model. Separate claims of code that spots additional systems have surfaced yet remain private and unexplained. Inquiries to other AI firms about similar markers went unanswered.

A universally dependable detector seems unlikely. Years of coverage on AI detection show no current technical method reliably separates generated text from human writing. Progress on image and audio watermarking exists but does not extend to plain text.

AI firms hold strong reasons to address this issue from within. Early training drew from broad web scrapes and rising synthetic content could lead to degraded model performance over time.

Communities continue to lean on subjective signals such as repetitive phrasing or excessive metaphors. AI produces such patterns because it learned from human examples. Pre-AI fanfiction already contained ornate styles that would trigger today's informal checks.

AO3 already offers a practical option through its tagging system. An official tag for generative AI use exists and some authors apply it voluntarily. Honest disclosure remains essential yet backlash reduces the motivation to use it. Fanfiction functions as a pastime rather than a formal profession.

Efforts to shield human creativity may inadvertently penalize writers whose style deviates from prevailing standards. One case involved an author whose trusted beta reader applied Claude without disclosure. A story that reads as mechanical may simply reflect ordinary human variation rather than machine involvement.

Follow topics and authors from this story to see more like this in your personalized homepage feed and to receive email updates.

Jess Weatherbed