The fanfiction community is at war with AI — and itself

Over the past week, a caller fanworks activity has kicked off, pinch the purpose to guidelines retired authors utilizing generative AI. But the discovery methods being implemented are questionable, and immoderate fanfic writer could beryllium caught successful the crossfire.

Broad distaste astir the usage of Claude, ChatGPT, and different AI devices has agelong been a thing successful creative communities, including the world of fanfiction. Readers and writers person passed astir tips for spotting supposedly AI-generated works, citing thing from em dashes to the broad concept of purple prose. But connected June 29th, an anonymous X relationship called @heatedrivalryai promised a seemingly much reliable solution. It posted a tegument — akin to an hold — for the celebrated fanfic repository Archive of Our Own (AO3) that would purportedly place coding artifacts near down by Anthropic’s Claude bot.

“When a Claude-generated consequence is pasted straight into AO3 from Claude, the matter is wrapped by a Claude-injected codification ‘font-claude-response-body,’” said the @heatedrivalryai account. “Its beingness indicates the usage of Claude definitively.” When a personification visits a page (like a activity of fanfic) pinch this code, the tegument turns the full inheritance red.

Several test posts person been published to AO3 that let users to cheque if it works. The surface instantly turned reddish erstwhile I tested the tegument against these examples myself, and I published a Claude-generated short communicative to tally my ain research conscionable successful case. The reddish surface appeared erstwhile I straight pasted from the chatbot into the editor and vanished if I pasted matter (including the nonstop aforesaid generated story) that didn’t travel consecutive from Claude.

The Claude detector station was accompanied by examples of fanfic wherever the artifacts were spotted, which the anonymous creator said was meant to show the strategy works, not “create an situation of mistrust aliases impeach peculiar users.” But fanfic communities person quickly mobilized to publically sanction and shame writers whose published useful were flagged by the tool, and its creator surely doesn’t see AI a affirmative thing. “Fandom is simply a uniquely connective, collaborative space. It thrives connected the quality constituent and the imaginative spark which drives it and feeds disconnected it,” they said. “If we unknowingly let AI to corrupt these spaces, what will beryllium near of them?”

Anthropic did not respond to my petition to verify if the fan-made Claude detector useful arsenic described. The methodology present does look sound, however, and our ain testing backs it up. There’s nary evident logic for the Claude codification to beryllium coming successful a communicative if the bot wasn’t utilized somehow. But there’s a clear consequence of some mendacious negatives and overgeneralizations.

The codification wrapping is only preserved if matter is copied straight from Claude into AO3’s editor, truthful it won’t drawback thing edited successful Google Docs aliases Microsoft Word and past moved to AO3 — and arsenic personification who writes for a living, I tin attest to really risky penning consecutive into a CMS is. Some writers who person been flagged person already updated their useful to region the artifacts, and early useful tin easy evade the tool.

Conversely, the tag doesn’t uncover really heavy Claude was utilized successful a fixed work. That flashbanged scarlet surface could mean the full communicative was afloat AI-generated, aliases that an writer pasted a fewer human-written sentences into Claude for spell-checking aliases translation, past moved them backmost into AO3.

That hasn’t mattered to immoderate fandom members, who position any usage of generative AI arsenic an inexcusable betrayal to the wider imaginative community. Many group mention concerns complete the biology effect of the exertion and really it’s trained by scraping the unfastened web, which apt includes fanworks uploaded to platforms for illustration AO3.

This peculiar tool’s applicability is constricted — AO3 isn’t the only level for publishing fanworks, and Claude is conscionable 1 of galore AI models. At slightest one personification claims they’ve written abstracted codification that tin observe “Claude, Deepseek, and immoderate ChatGPT” usage, but they haven’t released that solution to the nationalist aliases explained really it works. I asked Google and OpenAI if their models time off immoderate traceable artifacts successful matter procreation that could beryllium detected by akin means, but they haven’t responded.

In fact, it’d beryllium highly astonishing if a universally reliable strategy existed. I’ve been reporting connected the issues surrounding AI detection for a fewer years now, and to my knowledge, location isn’t presently a reliable technological solution for distinguishing generated matter from that typed retired by quality hands. Systems for illustration C2PA Content Credentials and Google’s SynthID are making immoderate advancement toward identifying generative AI successful images, videos, and moreover audio, but these trust connected invisible watermarks and metadata that don’t transportation complete for copy-pasted text.

That could alteration successful the future, and AI companies person each inducement to astatine slightest lick the problem internally. Early models were trained connected matter indiscriminately scraped from the internet, and arsenic quality penning is crowded retired by its synthetic counterpart, they could consequence a “model collapse” script that would degrade the accuracy of outputs.

For now, though, fandom communities are still mostly relying connected vibes. Most fanfics aren’t judged by a instrumentality for illustration the AO3 skin, but by “tells” that could see thing from circumstantial condemnation structures — for illustration the notorious “it’s not X, it’s Y” — to overuse of flowery metaphors. (At slightest cipher successful fandom, truthful far, has benches becoming men.) But we person to retrieve that AI often Writes Like That because it was trained connected worldly that existent group person written. It’s trying to replicate us. I’m not bold capable to stock my ain AO3 bookmarks, but I’ve decidedly publication immoderate overly grandiloquent fanfics successful the pre-ChatGPT net days that wouldn’t walk this dubious sniff test.

The champion solution for distinguishing AI useful connected AO3 is already available: the site’s robust tagging system. A “Created Using Generative AI” tag exists, and galore authors do see it to disclose the usage of devices for illustration Claude. That requires honorable transparency, though, and there’s small inducement for honesty fixed the backlash. It’s besides worthy remembering that fanfiction is expected to beryllium a hobby, not a regulated industry.

With these efforts to forestall AI from taking eyeballs distant from genuine human-driven creativity, authors who don’t conform to what’s deemed to beryllium an acceptable value of penning whitethorn go guiltless victims of the ongoing witch hunt. At slightest 1 writer has already been caught up successful this because different personification they trusted to edit their fic did truthful utilizing Claude. So if the adjacent fanfic you publication feels a small robotic, conscionable carnivore successful mind that it mightiness not actually beryllium the merchandise of a robot.