Anthropic 发布《Claude 如何标记 AI 生成内容》却未说明具体方法

Anthropic 支持页面：

Anthropic 已签署欧盟《AI 法案》第 50(2) 条《AI 生成内容透明度行为准则》，其身份既是生成式 AI 模型提供商，也是生成式 AI 系统提供商。本文介绍我们计划如何落实这些承诺、标记如何运作，以及其局限性。随着更多信息可用，我们将更新本文并发布更详细的技术指南。

关于生成的纯文本输出：

当受支持的 Claude 模型生成文本时，它会将一种不可感知的水印直接编织进文本本身。你不会看到它，它也不会改变 Claude 回复的含义、质量或可读性。

由于水印是文本的一部分，当文本被复制粘贴到其他地方时，它会随之传播，并且在经过某些编辑后仍可能保留。水印将在模型层面应用，这意味着无论文本来自哪一种 Claude 产品或界面，它都会存在。[…]

我们还在努力让用户及其他第三方能够检测 Claude 嵌入的水印和溯源元数据。检测会检查一段文本或一个文件是否带有受支持的 Claude 标记。如果发现受支持的标记，则表明该内容可能曾由 Claude 处理。

我们将在后续技术文档中分享有关检测机制的详细信息。

Anthropic 声称此举是为了遵守一项欧盟法律，但同时表示：“无论 Claude 在世界何处提供服务，标记都将适用于受支持模型的输出。”

这种不透明程度令人愤怒。对于他们称将嵌入到 PDF、SVG、PNG 和 JPEG 等生成文件中的“签名溯源元数据”，我暂且不谈。这很重要，但也很复杂。纯文本则很简单。一段文本字符串，无非是一个字符接着另一个字符。起初我以为 Claude 将开始在可见字符之间“嵌入”不可见字符或字节序列——而这必然会立刻造成问题。

假设我让 Claude 生成短语“Hello world”。（两个词肯定短到不值得进行“AI 生成”检测——此处我使用两个词只是为了让示例更简单。）Claude 创建的字符串会是“ Hello«这里是构成“水印”的不可见字符 » world”。我将其粘贴到网页上。人们看到的会是“Hello world”，但那些不可见字符仍然存在。设有字符数限制的社交媒体平台会显示，该字符串包含的字符多于“Hello world”中 11 个可见字符。

如果“水印”不是由不可见字符构成，而是由可见字符构成，那么这究竟如何符合他们所谓“不可感知”的说法？这个 Reddit 讨论串声称会以这种方式实现——“这是一种隐写术形式，模型会微妙地偏置其用词选择，从而形成日后可以检测到的统计模式。”这又怎么能与“它不会改变含义、质量或可读性”的说法相一致？而任何此类基于语义的检测都会造成误报问题。

如果我要求一个工具为我生成文本或建议文本，我希望该工具能生成它所能给出的最佳措辞。不是为了水印而损害其输出。如果他们真这么做，这应该会成为 Claude 品牌的毒药。

如果你复制我的文本，并手动将其引用到你所写的内容中，会发生什么？现在你的文章里就带有一个水印——无论是不可见字符还是别的什么——这会暗示你的文章是 AI 生成的？即使你根本没用 AI，只是从我这里复制粘贴了文本？荒唐。

显然，如果真要让任何人检测它，他们就必须确切说明在文本中嵌入了什么；但如果他们说明了，任何人都可以直接将其移除。例如，如果有人只是对 Claude 生成的文本进行 OCR，而不是选中并复制它，又会怎样？这一切都愚蠢至极。我从未如此庆幸自己实际上从来没有用 Claude 做过任何事。

★ 2026 年 8 月 11 日，星期二

Anthropic support page:

Anthropic has signed the EU AI Act’s Article 50(2) Code of Practice on Transparency of AI-Generated Content, as a provider of both generative AI models and generative AI systems. This article describes how we’re planning to put those commitments into practice, how marking works, and what its limitations are. We’ll update this article and publish more detailed technical guidance as it becomes available.

Regarding generated plain text output:

When a supported Claude model generates text, it weaves an imperceptible watermark directly into the text itself. You won’t see it, and it doesn’t change the meaning, quality, or readability of Claude’s response.

Because the watermark is part of the text, it will travel with the text when it’s copied and pasted elsewhere, and may persist through some editing. Watermarking will be applied at the model level, which means it will be present no matter which Claude product or surface the text comes from. [...]

We’re also working to enable users and other third parties to detect Claude’s embedded watermarks and provenance metadata. Detection checks whether a piece of text or a file carries a supported Claude mark. If a supported mark is found, it indicates that the content may have been processed by Claude.

We’ll share details on detection mechanisms in forthcoming technical documentation.

Anthropic is claiming this is for compliance with an EU law but that “Marking will apply to output from supported models wherever Claude is offered, worldwide.”

This is infuriatingly opaque. I won’t speak to the “signed provenance metadata” they say they’ll be embedding in generated files like PDFs, SVGs, PNGs, and JPEGs. That’s important but it’s complex. Plain text is simple. A string of text is just one character followed by another. At first I presumed that Claude is going to start “embedding” invisible characters/byte sequences between visible characters — which would unavoidably cause immediate problems.

Let’s say I have Claude generate the phrase “Hello world”. (Two words is surely too short to bother with “AI generated” detection — I’m using a two-word phrase here for simplicity in the example.) Claude creates the string of characters “ Hello«invisible characters that comprise a “watermark” here » world”. I paste this on the web. Humans who look at it will see “Hello world” but the invisible characters are there. Character-count-limited social media platforms will show that the string contains more than the 11 visible characters in “Hello world”.

If the “watermark” is not comprised of invisible characters but rather visible ones, how in the world does this jibe with their claim that it’s “imperceptible”? This Reddit thread claims that’s how it will work — “It’s a form of steganography where the model subtly biases its word choice to create a statistical pattern that can be detected later.” How in the world can that be squared with “it doesn’t change the meaning, quality, or readability”? And any sort of semantic detection like this is going to cause false-positive problems.

If I ask a tool to generate text or suggest text for me, I expect that tool to generate the best possible word choices it can. Not corrupt its output for the sake of watermarking. If they go through with this, this ought to be poison to the Claude brand.

What happens if you copy my text and quote it in something you write, by hand? Now you’ve got a watermark in your prose, invisible characters or something, that implicates that your prose is AI-generated? Even though you didn’t use AI and just copy and pasted text from me? Madness.

They obviously need to explain exactly what they’re embedding in text if anyone is actually going to detect it, but if they explain it, anyone can simply remove it. And what happens, for example, if someone just OCRs Claude-generated text rather than selects and copies it? This is all so stupid. I have never been happier that I’ve never actually used Claude for anything.

★ Tuesday, 11 August 2026