An anonymous account called @heatedrivalryai posted a browser extension on June 29, 2026 that identifies text generated by Anthropic's Claude when it appears on Archive of Our Own, the largest fanfiction platform. The fanfiction community has since mobilized around it in ways the creator did not anticipate.

The detection method is narrow and specific. When you paste text directly from Claude into AO3's editor, Claude embeds a CSS class — a piece of code that marks its container — called font-claude-response-body. The @heatedrivalryai extension scans any AO3 work page for that marker. If found, the entire page background turns red. The Verge tested the tool and confirmed the behavior: red background on direct Claude-paste, normal background on the same story with the artifact removed. Several test works were published on AO3, including at works/87682756, for community verification.

The limitation is significant but easy to bypass. The detection only works on text copied directly from Claude into AO3's editor. Route the same text through Google Docs or Microsoft Word first and the font-claude-response-body marker does not survive the transfer. The tool catches direct, unchanged use more than it catches intention — a distinction that matters.

The creator stated the original goal plainly: demonstrate that detection is technically feasible, not create a culture of suspicion or accusation. That intent has not held up in practice. Community members have already begun publicly identifying writers whose works triggered the red screen, and some of those writers have quietly edited their posts to remove the Claude artifact. The tool became a social enforcement mechanism faster than its author expected.

AO3 already has systems for voluntary disclosure. The platform maintains browsable tag pages for 'AI-Generated Text', 'Created Using Generative AI', and 'AI-Generated Images' — so the infrastructure for self-identification has existed. The real question the community has wrestled with is enforcement. What happens when labeling is not voluntary? The Organization for Transformative Works addressed this tension in May 2023, acknowledging community concern and mentioning existing detection tools like ZeroGPT and GPTZero. The Claude skin targets a specific artifact rather than using probabilistic guessing, but it sits in the same contested zone.

Anthhropic did not respond to The Verge's request to confirm whether the artifact the tool detects is genuine or whether it persists in newer Claude versions.

The silence matters practically. CSS class names embedded by a web-based AI interface are implementation details, not deliberate signatures. They can be renamed, removed, or randomized in any product update. The current detector works against specific Claude behavior at a specific moment; it is not a reliable long-term forensic tool. Anyone using it for enforcement rather than proof-of-concept is relying on something Anthropic could change without notice.

What the detector actually exposes is this: the fanfiction community faces a trust problem that AI detection tools — whether probabilistic or artifact-based — cannot cleanly solve on their own. AO3 is built on an understanding that works represent human creative effort, often deeply personal, always connected to source material and to other fans. The resistance to AI-generated fiction is not fundamentally about quality. It concerns the relational contract at the center of fan culture: that authorship is human. A red screen proves a specific paste operation occurred. It does not answer what the community actually wants to decide, which is where the boundary should sit between human authorship and machine generation in a space that has always treated the former as foundational.