Anthropic宣布将为Claude模型实现文本水印功能，以便识别AI生成的内容[1]。这一举措旨在符合欧盟AI法案的要求，该法案要求AI服务提供商对生成的内容进行标记[1]。
该水印技术基于Google DeepMind的SynthID-Text方法，源自2022年Scott Aaronson的相关提议[1]。与传统水印不同，这项技术通过改变词语选择的随机数源来实现标识，而不会影响输出质量[1]。Anthropic强调，水印对模型的运行速度、成本以及输出质量均无实质性影响，也不会产生额外的token消耗[1]。
就隐私保护而言，该水印设计无法追踪到具体的用户、组织或聊天记录[1]。Anthropic计划推出水印检测API供使用者查证[1]。对于图片等非文本文件，公司将采用C2PA开放标准来附加内容凭证[1]。根据计划，8月2日前发布的Claude模型将在未来几个月内逐步获得水印功能[1]。Anthropic等约190家企业在欧盟AI透明度行为准则上签署了承诺[1]。
Anthropic has announced that its Claude models will implement text watermarking to identify content generated by artificial intelligence [1]. The watermarking approach, based on Google DeepMind's SynthID-Text technology and rooted in a 2022 proposal by Scott Aaronson, alters the random number source for word selection without compromising output quality [1]. This implementation aligns with the European Union AI Act's requirement that AI providers mark AI-generated content, with Anthropic and approximately 190 other companies having signed the EU's AI transparency code of conduct in July 2026 [1].
The watermarking system introduces no material impact on model speed, operational costs, or output quality, and generates no additional tokens [1]. Critically, the watermark cannot be traced to specific users, organizations, or chat records, addressing privacy concerns [1]. Anthropic plans to release a watermark detection API and intends to use the C2PA open standard to append content credentials to non-text files such as images [1]. Claude models released before August 2 will receive watermarking capabilities gradually over the coming months [1].