Anthropic 为所有 Claude 输出全球添加水印，称"可能经部分编辑后仍留存"

Anthropic 已签署欧盟《人工智能生成内容透明度行为准则》。从 2026 年 8 月起，新的 Claude 模型将在文本中嵌入水印，并为文件附上签名的来源元数据。

自 2026 年 8 月 2 日起在欧盟推出的 Claude 模型将自带这些标签。这一要求也不会仅限于欧盟境内，将在全球范围内适用于所有 Claude 产品，包括 API、Claude、Claude Code、Claude Cowork 和 Claude Tag。

生成的文本将携带嵌入水印：https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content，而生成的文件将附有数字签名的来源元数据。现有模型在法律上享有过渡期，但 Anthropic 表示已经在着手改造它们。广告

公司还计划发布验证工具，以便用户和第三方检查这些标签，但尚未说明具体时间。Anthropic 表示，将 Claude 集成到产品中的开发者必须确定他们的服务适用哪些第 50 条要求。广告 DEC_D_Incontent-1

Anthropic 计划使用两种类型的标签。据公司称，Claude 生成的文本将带有一种不会影响其含义、质量或可读性的隐形水印。该水印在复制和粘贴过程中仍然有效，并且“可能在某些编辑操作后仍然存在”。水印直接在模型层面应用，因此使用哪款 Claude 产品都无关紧要。

支持的文件，包括 .svg、.png 和 .jpg 图像，将附带基于开放 C2PA 标准的签名来源元数据：https://the-decoder.com/openai-supports-c2pa-standard-for-ai-images-and-publishes-classifier/，由内容来源与真实性联盟（Coalition for Content Provenance and Authenticity）开发。该签名表明文件由 Claude 处理，并可揭示后续篡改。文本水印也应可通过 AWS、Google Cloud 和 Microsoft Foundry 等云合作伙伴使用，尽管这些平台可能不支持签名元数据。广告

Anthropic 对其限制持诚实态度。检测到水印并不意味着内容确实由 Claude 撰写。人们经常使用 Claude 进行校对、翻译或总结，因此输出内容可能带有水印，即使思想本身来源于人类。

没有水印也无法澄清问题。该模型可能在水印推出之前就已发布，文本可能经过大量编辑或翻译，篇幅可能过短而无法进行可靠检测，或者元数据在格式转换或截图过程中被移除。广告 DEC_D_Incontent-2

真正的考验是这些水印在编辑、重新格式化和翻译后能维持多好。如果它们能保持稳定，检查已知水印应该比像 Pangram 这样的工具更可靠，因为 Pangram 的专有检测方法不会显示触发结果的原因：https://the-decoder.com/pangram-ceo-says-language-models-give-themselves-away-by-making-the-same-arguments/。第三方检测器可以增加对 Anthropic 水印的支持，从而提供更可靠的信号。广告

Anthropic 并非孤军作战。Google Deepmind 开源了其 SynthID 水印系统：https://the-decoder.com/google-deepmind-open-sources-synthid-text-watermarking/，并将其集成到 Gemini 模型中。SynthID 在令牌预测过程中轻微调整概率值，以在不降低文本质量的情况下生成水印。它适用于多种语言，但对于生成后被编辑的文本效果不佳。

OpenAI 已经拥有一个准确率达 99.9% 的文本检测器：https://the-decoder.com/openai-has-a-highly-accurate-chatgpt-text-detector-but-wont-release-it-for-now/，但大约两年来仍未发布。原因包括用户可以通过翻译或重写轻易绕过检测、存在对某些群体的污名化风险，以及可能担心公开检测器会损害 OpenAI 自身业务。

这种风险在教育领域尤其严重，因为不可靠的检测器可能导致错误的作弊指控：https://the-decoder.com/us-court-ruling-backs-schools-right-to-penalize-students-for-ai-cheating/。与此同时，人们也有充分理由希望知道AI何时以及使用了多少。研究表明，过度依赖AI工具可能削弱批判性思维和写作能力：https://the-decoder.com/grades-dropped-from-96-to-48-percent-when-a-brown-professor-made-students-take-the-exam-without-ai/，尤其是在那些把AI当作捷径而非学习辅助工具的学生中：https://the-decoder.com/95-of-uk-students-now-use-ai-and-their-experiences-couldnt-be-more-divided/。问题不仅限于学术领域，诈骗分子现在也在美国大学招收虚假学生，并利用AI轻松完成课程作业并领取经济援助：https://the-decoder.com/scammers-are-enrolling-fake-students-at-us-community-colleges-and-using-ai-to-collect-financial-aid/。

Anthropic的决定也可能影响其业务。Claude在知识工作中很受欢迎，尤其是在学校和大学学生中，因为即使是较早的模型也能生成相当自然的文章。随着学校和大学已经在学术工作中争论AI的使用：https://the-decoder.com/one-of-the-worlds-top-law-schools-draws-a-hard-line-against-ai-in-legal-education/，更可靠的检测可能会使Claude对这些用户的吸引力降低。

保持对AI的了解。内容清晰、有用，无废话。

关注The Decoder获取AI新闻、背景故事和专家分析。

解码器：https://the-decoder.com/

Anthropic has signed the EU AI Act Code of Practice on transparency for AI-generated content. Starting in August 2026, new Claude models will embed watermarks in text and attach signed provenance metadata to files.

Claude models that launch in the EU on or after August 2, 2026, will ship with this labeling baked in. The requirement won't stop at EU borders either. It'll apply globally across all Claude products, including the API, Claude, Claude Code, Claude Cowork, and Claude Tag.

Generated text will carry embedded watermarks：https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content, while generated files will get digitally signed provenance metadata. Existing models get a transition period under the law, but Anthropic says it's already working on retrofitting them. Ad

The company also plans to release verification tools so users and third parties can check the labels, though it hasn't said when. Developers that integrate Claude into their products must determine which Article 50 requirements apply to their services, according to Anthropic. Ad DEC_D_Incontent-1

Anthropic plans to use two types of labels. Text generated by Claude will carry an invisible watermark that doesn't affect its meaning, quality, or readability, according to the company. The watermark survives copying and pasting and "may persist through some editing." It gets applied at the model level, so it doesn't matter which Claude product you're using.

Supported files, including .svg, .png, and .jpg images, will carry signed provenance metadata based on the open C2PA standard：https://the-decoder.com/openai-supports-c2pa-standard-for-ai-images-and-publishes-classifier/, developed by the Coalition for Content Provenance and Authenticity. The signature indicates that Claude processed the file and can reveal later tampering. Text watermarks should also work through cloud partners such as AWS, Google Cloud, and Microsoft Foundry, though those platforms may not support signed metadata. Ad

Anthropic is upfront about the limits. A detected watermark doesn't mean Claude actually wrote the content. People use Claude all the time for proofreading, translating, or summarizing, so the output might carry a watermark even though the ideas came from a human.

No watermark doesn't clear things up either. The model might have shipped before watermarking rolled out, the text could have been heavily edited or translated, the passage might be too short for reliable detection, or the metadata got stripped through format conversion or a screenshot. Ad DEC_D_Incontent-2

The real test is how well these watermarks survive editing, reformatting, and translation. If they hold up, checking for a known watermark should be more reliable than tools like Pangram, whose proprietary detection methods don't reveal what triggered a result：https://the-decoder.com/pangram-ceo-says-language-models-give-themselves-away-by-making-the-same-arguments/. Third-party detectors could add support for Anthropic's watermark, giving them a more reliable signal. Ad

Anthropic isn't alone here. Google Deepmind open-sourced its SynthID watermarking system：https://the-decoder.com/google-deepmind-open-sources-synthid-text-watermarking/, building it into the Gemini models. SynthID slightly tweaks probability values during token prediction to create a watermark without degrading text quality. It works across languages but struggles with text that's been edited after generation.

OpenAI has been sitting on a text detector with 99.9：https://the-decoder.com/openai-has-a-highly-accurate-chatgpt-text-detector-but-wont-release-it-for-now/ percent accuracy for about two years and still hasn't released it. The reasons include how easily users can beat it through translation or rewriting, the risk of stigmatizing certain groups, and likely worries that a public detector could hurt OpenAI's own business.

That risk is especially serious in education, where unreliable detectors can lead to false cheating allegations：https://the-decoder.com/us-court-ruling-backs-schools-right-to-penalize-students-for-ai-cheating/. At the same time, there are good reasons to want to know when and how much AI was used. Studies show that heavy reliance on AI tools can weaken critical thinking and writing skills：https://the-decoder.com/grades-dropped-from-96-to-48-percent-when-a-brown-professor-made-students-take-the-exam-without-ai/, particularly among students who treat them as a shortcut rather than a learning aid：https://the-decoder.com/95-of-uk-students-now-use-ai-and-their-experiences-couldnt-be-more-divided/. The problem goes beyond academics too, with scammers now enrolling fake students at US colleges and using AI to breeze through coursework and collect financial aid：https://the-decoder.com/scammers-are-enrolling-fake-students-at-us-community-colleges-and-using-ai-to-collect-financial-aid/.

Anthropic's decision could also affect its business. Claude is popular for knowledge work, especially among school and college students, because even older models produce fairly natural prose. With schools and universities already fighting over AI use in academic work：https://the-decoder.com/one-of-the-worlds-top-law-schools-draws-a-hard-line-against-ai-in-legal-education/, more reliable detection could make Claude less appealing to those users.

Stay in the loop on AI. Clear, useful, no fluff.

Follow The Decoder for AI news, background stories and expert analyses.

The Decoder：https://the-decoder.com/