内容来源：https://www.theverge.com/tech/960854/ai-fanfiction-ao3-claude-detector

内容总结：

同人圈掀起“AI清剿”运动：检测工具存争议，无辜作者或遭误伤

过去一周，一场旨在揪出使用生成式AI创作的同人圈新“清剿”运动悄然兴起，但其采用的检测手段备受质疑，可能导致大量同人作者无辜“中弹”。这场运动正将同人小说社区推向一场与AI、乃至与自身的内战。

长期以来，创意社群对Claude、ChatGPT等AI工具普遍持反感态度。6月29日，一个名为@heatedrivalryai的匿名X账号发布了一款针对知名同人作品库AO3的“皮肤”（类似浏览器扩展），声称能识别Anthropic公司Claude机器人留下的编码痕迹。该工具的原理是：当用户直接从Claude复制粘贴文本到AO3编辑器时，文字会被一段名为“font-claude-response-body”的代码包裹。启用该皮肤后，访问包含此代码的页面时，整个背景将变红。经测试，直接粘贴Claude生成内容确实会触发红色警报，但若文本经过中转编辑（如谷歌文档）再上传，则能轻松规避检测。

该工具的局限性显而易见：它仅能识别直接复制粘贴的Claude生成内容，无法判定AI在创作中的使用程度——一篇文章全由AI生成，与作者仅用AI进行拼写检查或翻译后粘贴回AO3，都会亮起同样的“红灯”。然而，部分激进的同人圈成员已开始根据此工具公开点名、羞辱被标记的作者，将任何AI使用行为视为对创意社群的背叛。一位被波及的作者解释，自己仅仅是让另一位信任的朋友用Claude帮忙校对，便遭到了误判。

此外，当前不存在万能的AI检测技术。谷歌、OpenAI等公司虽在图像、音视频的水印识别上取得进展，但对纯文本的复制粘贴仍缺乏可靠方案。同人小说社区目前更多依赖“感觉”判断，如特定句式（“不是X，而是Y”）、过度华丽的辞藻等。然而，这些“特征”往往源于AI对真人写作风格的模仿，许多AI时代的“可疑”文风，在ChatGPT诞生前的互联网上就已存在。

专家指出，AO3现有的“使用生成式AI创作”标签系统本是最佳解决方案，但面对激烈的抵制声浪，作者如实标注的意愿极低。随着这场“猎巫”行动的升级，那些无法达到所谓“人类创作标准”的普通作者，正面临着成为无辜牺牲品的风险。正如已有案例所示，下一次当你觉得某篇同人文读起来略显“机械”时，它或许并非出自机器之手。

中文翻译：

过去一周，一场新的同人创作运动悄然兴起，旨在清除使用生成式AI的作者。然而，其实施的检测方法存在争议，任何同人小说作者都可能因此受到波及。
同人小说社群正与AI——也与其自身——展开战争
读者们正匆忙研发检测方法，以判断同人作品是否借助生成式AI完成。这些方法的效果存疑。
同人小说社群正与AI——也与其自身——展开战争
读者们正匆忙研发检测方法，以判断同人作品是否借助生成式AI完成。这些方法的效果存疑。
在包括同人小说在内的创意社群中，对Claude、ChatGPT及其他AI工具的普遍反感由来已久。读者和作者们曾互相传授识别疑似AI生成作品的技巧，从破折号的使用到所谓的“紫色散文”泛化概念，无所不包。但6月29日，一个名为@heatedrivalryai的匿名X账号发布了一种看似更可靠的解决方案：它为流行同人作品库AO3（Archive of Our Own）提供了一款类似扩展功能的皮肤，据称能识别Anthropic公司Claude机器人留下的编码痕迹。
“当Claude生成的回复被直接粘贴到AO3时，文本会被Claude注入的代码‘font-claude-response-body’包裹，”@heatedrivalryai账号表示，“若存在此代码，即可明确判定使用了Claude。”当用户访问含有此代码的页面（如某篇同人作品）时，皮肤会将整个背景变为红色。
已有数篇测试文章发布至AO3，供用户验证该皮肤是否有效。当我亲自用这些例子测试皮肤时，屏幕瞬间变红；为确保万无一失，我还发布了一篇由Claude生成的短篇小说进行实验。当我从聊天机器人直接复制文本到编辑器时，屏幕呈现红色；而若粘贴非直接来自Claude的文本（包括完全相同的生成故事），红色背景则消失。
该Claude检测工具的发布帖还附带了被检测出编码痕迹的同人作品示例，匿名创作者声称此举意在展示系统可靠性，而非“制造不信任氛围或针对特定用户”。但同人社群迅速行动起来，公开点名批评被该工具标记的作者，其创作者显然并不将AI视为积极事物。“同人圈是一个独特联结、协作的空间。它依赖人性元素与驱动、滋养自身的创意火花，”他们表示，“若我们无意中允许AI侵蚀这些空间，最终还剩下什么？”
Anthropic未回应我关于验证该粉丝自制Claude检测工具是否如描述般有效的请求。不过，其方法学看似可靠，我们自己的测试也证实了这一点。若未使用机器人，故事中显然没有理由出现Claude的代码。但“假阴性”和“过度泛化”的风险显而易见。
只有直接从Claude复制文本到AO3编辑器时，包裹代码才会保留；因此，在谷歌文档或Microsoft Word中编辑后再上传至AO3的内容不会被检测到——作为以写作为生的人，我可以证明直接在CMS中写作有多冒险。一些被标记的作者已更新作品以移除编码痕迹，而未来的作品也极易规避该工具。
反之，该标签无法揭示Claude在作品中的使用程度。那刺眼的猩红屏幕可能意味着整篇故事完全由AI生成，也可能只是作者将少量人工撰写的句子粘贴至Claude进行拼写检查或翻译，再移回AO3。
这对部分同人社群成员而言并不重要，他们将任何使用生成式AI的行为视为对整个创意社群不可饶恕的背叛。许多人担忧该技术的环境影响及其通过抓取开放网络（很可能包括上传至AO3等平台的同人作品）进行训练的方式。
该特定工具的适用范围有限——AO3并非发布同人作品的唯一平台，而Claude也只是众多AI模型之一。至少有人声称已编写出能检测“Claude、Deepseek及部分ChatGPT”使用痕迹的独立代码，但尚未公开该方案或解释其原理。我询问谷歌和OpenAI其模型在文本生成中是否会遗留可通过类似手段检测的痕迹，但未获回应。
事实上，若存在一套普遍可靠的系统，反倒令人意外。我报道AI检测相关问题已有数年，据我所知，目前尚无可靠技术方案能区分生成文本与人类手打文本。C2PA内容凭证、谷歌SynthID等系统在识别图像、视频甚至音频中的生成式AI方面取得进展，但这些依赖不可见水印和元数据，对于复制粘贴的文本而言并不适用。
AI公司有充分动力至少从内部解决该问题
未来情况可能改变，而AI公司有充分动力至少从内部解决该问题。早期模型基于从互联网不加甄别抓取的文本进行训练，随着人类写作被其合成副本挤占空间，它们可能面临“模型崩溃”风险，导致输出准确性下降。
不过，目前同人社群仍主要依赖“直觉判断”。多数同人作品并非通过AO3皮肤等工具评判，而是依靠“破绽”——从特定句式结构（如臭名昭著的“不是X，而是Y”）到过度使用华丽隐喻。（至少至今同人圈内尚未出现“长椅变成男人”的荒谬案例。）但我们需谨记，AI之所以常“那样写作”，恰恰因其训练素材源自真人创作——它在竭力模仿我们。我虽不敢公开自己的AO3书签，但确实在ChatGPT诞生前的互联网时代读过一些过于浮夸的同人作品，它们恐怕难以通过这种可疑的“直觉测试”。
区分AO3上AI作品的最佳方案早已存在：该网站强大的标签系统。已有“使用生成式AI创作”标签，许多作者也确实会标注以表明使用Claude等工具。但这需要诚实的透明度，而面对群起攻之的声讨，诚实缺乏激励。同样值得铭记的是，同人创作应是爱好，而非受监管的产业。
在这些阻止AI抢夺人类创意关注度的努力中，那些写作水准不符合所谓“可接受标准”的作者，可能成为这场持续猎巫行动的无辜受害者。已有至少一位作者因此遭殃——因一位受托编辑其作品的友人使用了Claude。因此，若你下一篇阅读的同人作品略显机械，请谨记：它未必真是机器产物。

英文来源：

Over the past week, a new fanworks movement has kicked off, with the aim to root out authors using generative AI. But the detection methods being implemented are questionable, and any fanfic writer could be caught in the crossfire.
The fanfiction community is at war with AI — and itself
Readers are scrambling to develop ways to detect whether generative AI was used to write fanworks. The results are questionable.
The fanfiction community is at war with AI — and itself
Readers are scrambling to develop ways to detect whether generative AI was used to write fanworks. The results are questionable.
Broad distaste around the use of Claude, ChatGPT, and other AI tools has long been a thing in creative communities, including the world of fanfiction. Readers and writers have passed around tips for spotting supposedly AI-generated works, citing anything from em dashes to the broad concept of purple prose. But on June 29th, an anonymous X account called @heatedrivalryai promised a seemingly more reliable solution. It posted a skin — similar to an extension — for the popular fanfic repository Archive of Our Own (AO3) that would purportedly identify coding artifacts left behind by Anthropic’s Claude bot.
“When a Claude-generated response is pasted directly into AO3 from Claude, the text is wrapped by a Claude-injected code ‘font-claude-response-body,’” said the @heatedrivalryai account. “Its presence indicates the use of Claude definitively.” When a user visits a page (like a work of fanfic) with this code, the skin turns the entire background red.
Several test posts have been published to AO3 that allow users to check if it works. The screen immediately turned red when I tested the skin against these examples myself, and I published a Claude-generated short story to run my own experiment just in case. The red screen appeared when I directly pasted from the chatbot into the editor and vanished if I pasted text (including the exact same generated story) that didn’t come straight from Claude.
The Claude detector post was accompanied by examples of fanfic where the artifacts were spotted, which the anonymous creator said was meant to demonstrate the system works, not “create an environment of mistrust or accuse particular users.” But fanfic communities have quickly mobilized to publicly name and shame writers whose published works were flagged by the tool, and its creator certainly doesn’t consider AI a positive thing. “Fandom is a uniquely connective, collaborative space. It thrives on the human element and the creative spark which drives it and feeds off it,” they said. “If we unknowingly allow AI to corrupt these spaces, what will be left of them?”
Anthropic did not respond to my request to verify if the fan-made Claude detector works as described. The methodology here does look sound, however, and our own testing backs it up. There’s no apparent reason for the Claude code to be present in a story if the bot wasn’t used somehow. But there’s a clear risk of both false negatives and overgeneralizations.
The code wrapping is only preserved if text is copied directly from Claude into AO3’s editor, so it won’t catch anything edited in Google Docs or Microsoft Word and then moved to AO3 — and as someone who writes for a living, I can testify to how risky writing straight into a CMS is. Some writers who have been flagged have already updated their works to remove the artifacts, and future works can easily evade the tool.
Conversely, the tag doesn’t reveal how heavily Claude was used in a given work. That flashbanged scarlet screen could mean the entire story was fully AI-generated, or that an author pasted a few human-written sentences into Claude for spell-checking or translation, then moved them back into AO3.
That hasn’t mattered to some fandom members, who view any use of generative AI as an inexcusable betrayal to the wider creative community. Many people cite concerns over the environmental impact of the technology and how it’s trained by scraping the open web, which likely includes fanworks uploaded to platforms like AO3.
This particular tool’s applicability is limited — AO3 isn’t the only platform for publishing fanworks, and Claude is just one of many AI models. At least one person claims they’ve written separate code that can detect “Claude, Deepseek, and some ChatGPT” usage, but they haven’t released that solution to the public or explained how it works. I asked Google and OpenAI if their models leave any traceable artifacts in text generation that could be detected by similar means, but they haven’t responded.
In fact, it’d be highly surprising if a universally reliable system existed. I’ve been reporting on the issues surrounding AI detection for a few years now, and to my knowledge, there isn’t currently a reliable technological solution for distinguishing generated text from that typed out by human hands. Systems like C2PA Content Credentials and Google’s SynthID are making some progress toward identifying generative AI in images, videos, and even audio, but these rely on invisible watermarks and metadata that don’t carry over for copy-pasted text.
AI companies have every incentive to at least solve the problem internally
That could change in the future, and AI companies have every incentive to at least solve the problem internally. Early models were trained on text indiscriminately scraped from the internet, and as human writing is crowded out by its synthetic counterpart, they could risk a “model collapse” scenario that would degrade the accuracy of outputs.
For now, though, fandom communities are still mostly relying on vibes. Most fanfics aren’t judged by a tool like the AO3 skin, but by “tells” that could include anything from specific sentence structures — like the notorious “it’s not X, it’s Y” — to overuse of flowery metaphors. (At least nobody in fandom, so far, has benches becoming men.) But we have to remember that AI often Writes Like That because it was trained on stuff that real people have written. It’s trying to replicate us. I’m not bold enough to share my own AO3 bookmarks, but I’ve definitely read some overly grandiloquent fanfics in the pre-ChatGPT internet days that wouldn’t pass this dubious sniff test.
The best solution for distinguishing AI works on AO3 is already available: the site’s robust tagging system. A “Created Using Generative AI” tag exists, and many authors do include it to disclose the use of tools like Claude. That requires honest transparency, though, and there’s little incentive for honesty given the backlash. It’s also worth remembering that fanfiction is supposed to be a hobby, not a regulated industry.
With these efforts to prevent AI from taking eyeballs away from genuine human-driven creativity, authors who don’t conform to what’s deemed to be an acceptable quality of writing may become innocent victims of the ongoing witch hunt. At least one writer has already been caught up in this because another person they trusted to edit their fic did so using Claude. So if the next fanfic you read feels a little robotic, just bear in mind that it might not actually be the product of a robot.