Claude began to secretly stamp all output: Anthropic launched text watermarks, and AI-generated content will have nowhere to hide. Anthropic announced that it would add machine-readable tags to the output content of Claude. This mechanism consists of two layers: an invisible watermark embedded in the text, and a digital signature metadata attached to the generated file.

Claude began to secretly stamp all output: Anthropic launched text watermarks, and AI-generated content will have nowhere to hide. Anthropic announced that it would add machine-readable tags to the output content of Claude. This mechanism consists of two layers: an invisible watermark embedded in the text, and a digital signature metadata attached to the generated file. Starting from August 2, all newly released Claude models have begun to implement this marking scheme. This is the transparency code of conduct signed by Anthropic in order to comply with Article 50 of the EU AI Act, but the scope of implementation is not limited to the EU - it will be effective everywhere in the world where Claude is used. Specifically, the first layer is a text watermark. When Claude generates text, it will weave invisible watermarks into the text without affecting the reading experience and content quality. The characteristic of this watermark is that it "follows the text" - if you copy and paste a paragraph written by Claude into an email, document or blog, the watermark will still exist and may remain after a certain degree of editing and modification. The second layer is file metadata. When Claude generates SVG, PNG, JPG and other files, signature metadata that complies with C2PA standards will be attached. C2PA is an open content traceability protocol jointly promoted by Adobe, Microsoft, Google and other companies. The image generation tools of OpenAI and Google are already using this set of standards. The coverage is quite broad. API, Claude official website, Claude Code, Claude Cowork, Claude Tag, all product lines are applicable. When calling Claude through AWS, Google Cloud or Microsoft Foundry, the text watermark also takes effect, but the file metadata depends on the functional support of each cloud platform. Prior to this, Claude had never publicly deployed text watermarks. OpenAI has previously developed a text watermarking solution, but it has never been launched due to various considerations. Anthropic implements watermarking directly at the text level this time, which is a relatively advanced step among large model manufacturers. For people who write with Claude, a direct impact is: emails, reports, and articles you draft with Claude will all carry watermarks that can be detected by machines. Anthropic said that supporting testing tools are being developed, and third parties can also perform testing in the future. But the restrictions are also very practical. Detecting the watermark can only indicate that the content "may have been processed by Claude". It cannot confirm that Claude is the original author, because many people use it to polish, translate, and summarize existing content. The opposite is also true, no water is detected