Claude 将为 AI 生成的文本和图像添加不可见水印

Anthropic 承诺将开始为 Claude 生成的文本和图像添加机器可读数据，以遵守欧洲有关 AI 透明度的规定。Anthropic 在新的 Claude 支持页面中表示：“生成的文本将携带嵌入式水印，而生成的文件将在支持的情况下包含经过数字签名的溯源元数据。”这些变化对人眼不可见，但将让用户和在线平台更容易检测内容是否由 Claude 模型生成。

这些更新是一项未来承诺，而非立即生效的措施。欧盟《AI 法案》下新的AI 标识与透明度义务于 8 月 2 日生效，其中规定：在该日期之前已推出的现有 AI 产品可享有四个月的合规宽限期。因此，Anthropic 表示，新的 Claude 模型将在发布首日就为 AI 生成内容添加标记，但对现有模型的支持仍在推进中。

机器可读标记将全球应用于受支持的 Claude 模型，包括 Claude Platform（API）、Claude、Claude Code、Claude Cowork 和 Claude Tag。Anthropic 正在使用两种不同的标记技术：对于由 Claude 处理的图像，将对受支持文件应用 C2PA——这是一项已被 Adobe、OpenAI 和 Google 采用的溯源元数据标准；但对于 Claude 生成文本的标记方式，Anthropic 披露的细节则少得多。

Anthropic 表示，一种“不可感知水印”会被直接编织进 Claude 模型生成的文本中，不会改变聊天机器人的回答在含义、质量或可读性上的表现。Anthropic 没有说明该水印系统的名称，但表示，当用户通过 AWS、Google Cloud 或 Microsoft Foundry 访问 Claude 模型时，也将应用这些文本水印。

Claude 支持页面称：“由于水印是文本的一部分，在文本被复制并粘贴到其他位置时，它也会随之传播，并且可能在经过部分编辑后依然保留。”“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。”

Anthropic 还在开发让用户及其他第三方检测 Claude 生成内容中嵌入的水印和溯源元数据的能力，并表示将在后续技术文档中分享该检测系统的详细信息。目前已有多种用于检测 C2PA 元数据的工具，包括 Google 的 Gemini 聊天机器人，但尚不清楚这些工具是否能识别 Claude 生成文件。我已向 Anthropic 询问澄清。

这是朝着在各类在线平台清晰标识 AI 生成文本和图像迈出的又一步，对于希望避免接触此类内容的人而言也可能是一项利好。同人小说读者已经在构建较为初步的检测系统，用于标记 Claude 工具是否被用于 AO3 同人作品；但这类标记系统的应用范围可能广泛得多——前提是它们确实有效。

众所周知，C2PA 数据很容易被移除，有时甚至会在承载这些数据的媒体上传至在线平台时被意外删除；Anthropic 的文本水印方案究竟有多稳健也尚不明确。即使 Anthropic 自己也承认，这些标记系统远非万无一失，任何缺少可检测标记的内容仍有可能来自生成式 AI 模型。

Anthropic has pledged to start marking Claude-generated text and images with machine-readable data, in an effort to comply with European rules for AI transparency. “Generated text will carry embedded watermarks, and generated files will include digitally signed provenance metadata where supported,” Anthropic says on a new Claude support page. The changes are invisible to human eyes, but will make it easier for people and online platforms to detect if content was generated by Claude models.

These updates are a future commitment rather than something that will go into effect immediately. New AI labeling and transparency obligations under the EU’s AI Act, which came into effect on August 2nd, include a four month compliance grace period for existing AI products that launched prior to that date. As such, Anthropic says new Claude models will mark AI-generated content from day one upon release, but support for its existing models is a work in progress.

The machine-readable marks will be applied globally to supported Claude models, including Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag. Two different marking techniques are being used: For images processed by Claude, C2PA — a provenance metadata standard already embraced by Adobe, OpenAI, and Google — will be applied to supported files, but the process for marking Claude-generated text is much lighter on details.

According to Anthropic, an “imperceptible watermark” is woven directly into the text generated by Claude models without changing the meaning, quality, or readability of the chatbot’s response. Anthropic doesn’t name this watermarking system, but says those text watermarks will also be applied when Claude models are accessed through AWS, Google Cloud, or Microsoft Foundry.

“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,” Anthropic says on the Claude support page. “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.”

Anthropic is also working to enable users and other third parties to detect watermarks and provenance metadata embedded into Claude-generated content, and says it’ll share details on this detection system in upcoming technical documentation. There are several tools already available that are designed to detect C2PA metadata, including Google’s Gemini chatbot, but it isn’t clear if those will work with Claude-generated files. I’ve asked Anthropic for clarification.

This is another step toward AI-generated text and images being clearly tagged across online platforms, and a potential win for folks who want to avoid consuming such content. Fanfiction readers have already been building more rudimentary detection systems to flag when Claude tools have been used in AO3 fanworks, but these marking systems can be applied far more broadly — if they work, that is.

C2PA data is known to be easily stripped out, sometimes even accidently when the media carrying it is uploaded to online platforms, and it’s unclear how robust Anthropic’s text watermarking solution is. Even the company itself is hedging that these marking systems are far from infallible, and that any content that lacks detectable marks could still originate from generative AI models.