August 11th news: Anthropic announced that it will add "imperceptible watermarking" technology to its new Claude model, used to identify the source of AI-generated text. The watermark does not affect the meaning or reading experience of the text and can be retained with the content during copy-pasting, and may still exist after some editing operations.
Anthropic stated that this feature is a measure to promote AI transparency according to the EU's Artificial Intelligence Act. Starting from the Claude models released on August 2nd and later, generated content will support marker recognition, covering texts generated globally through Claude, including versions accessed via cloud service providers. The company also plans to provide third parties with watermark detection tools and is considering expanding this capability to older model versions.
This technical upgrade is considered to potentially affect fields such as publishing and education, which highly rely on text originality. In recent years, the issue of AI-generated content impersonating human creation has continued to spark controversy, with some publishing projects facing scrutiny due to suspected use of AI text. Claude text watermarking offers publishers, schools, and universities a new means of content tracing, helping identify AI-generated works that are not disclosed.
However, Anthropic also pointed out that the watermark is not absolutely reliable. Large-scale editing, rewriting, translation, or mixing with other content may lead to the watermark being undetectable. In addition, detecting a watermark can only prove that the content was generated with the participation of Claude, but cannot directly prove that Claude is the sole author of the text, as auxiliary operations such as proofreading and translation may also leave marks.
Anthropic is not the first company to explore AI text identification technology. In 2024, Google DeepMind had already announced that it would add watermarks to texts and videos generated by Gemini using SynthID technology, which had previously been applied in the field of AI images. With the widespread adoption of generative AI, AI content identification and source tracking have become an important direction for industry transparency development.
