Anthropic发布博客文章，详细说明了为Claude聊天机器人部署的文本水印技术如何运作。[1]该水印系统旨在遵守欧盟AI法案的透明度要求，通过在"低风险选择"中嵌入不可见但可检测的模式来标识AI生成内容。[1]根据Anthropic的说法，"对读者来说，带水印的回复与无水印的回复无法区分"。[1]
Anthropic采用了Google DeepMind在2024年提出的SynthID-Text方法来实现这一技术。[1]水印的抗编辑能力有限度：轻度编辑"可能无法完全移除水印"，但"完全重写（每个词都被替换）将可以移除水印"。[1]在代码生成方面，水印效果较弱，因为模型需要生成有效代码，但在代码注释等任意选择的区域仍然可以使用水印。[1]
此外，"其他主要模型开发者已签署相同的行为准则，将实施自己的水印"。[1]
Anthropic has disclosed technical specifics about how watermarks will function in Claude, its conversational AI system, to meet transparency requirements under the European Union's AI Act [1]. The company has adopted Google DeepMind's SynthID-Text approach to embed imperceptible yet detectable patterns into AI-generated text, allowing the system to mark content as machine-produced while remaining visually indistinguishable from unmarked responses to readers [1].
The watermarking technique operates within what Anthropic describes as "low-risk choices"—areas where the model has flexibility in word selection without compromising output quality [1]. While minor edits may not fully remove the watermark, comprehensive rewrites that replace every word can eliminate it entirely [1]. Code generation presents a particular challenge for watermarking, as the model must prioritize producing valid, functional code; however, watermarks can still be applied in flexible areas such as code comments [1].
Anthropic plans to release a watermark detection application programming interface (API) to facilitate identification of marked content [1]. The company noted that other major model developers have signed the same conduct commitments and will implement their own watermarking systems [1], signaling an industry-wide shift toward machine-generated content transparency.