Anthropic adds an invisible watermark to the text generated by Claude to improve the traceability of AI content. According to news on August 11, Anthropic announced that it will add "undetectable watermark" technology to the new Claude model to identify the source of AI-generated text.

Anthropic adds an invisible watermark to the text generated by Claude to improve the traceability of AI content. On August 11, Anthropic announced that it will add "undetectable watermark" technology to the new Claude model to identify the source of AI-generated text. This watermark will not affect the meaning of the text or the reading experience. It can be retained with the content during the copy and paste process, and may continue to exist after some editing operations. Anthropic said that this feature is one of its measures to promote AI transparency in accordance with the EU Artificial Intelligence Act. Starting from the Claude model released on August 2 and later, the generated content will support mark recognition, covering text generated by Claude globally, including versions called through cloud service providers. The company also plans to make watermark detection tools available to third parties and is considering extending the capability to older models. This technological upgrade is believed to have an impact on areas that are highly dependent on text originality, such as publishing and education. In recent years, the issue of AI-generated content passing off as human creation has continued to cause controversy, and some publishing projects have been questioned for allegedly using AI text. Claude text watermark can provide publishers, schools and universities with new content traceability methods to help identify undisclosed AI-generated works. However, Anthropic also pointed out that watermarks are not absolutely reliable. Large-scale editing, rewriting, translation, or mixing of AI-generated text with other content may result in watermarks that cannot be detected. In addition, detecting watermarks can only prove that the content was generated by Claude, but it cannot directly prove that Claude is the sole author of the text, because auxiliary operations such as proofreading and translation may also leave marks. Anthropic is not the first company to explore AI text marking technology. In 2024, Google DeepMind has announced that it will add watermarks to the text and videos generated by Gemini through SynthID technology, which has previously been applied to the field of AI images. With the popularity of generative AI applications, AI content identification and source tracking are becoming important directions for the development of industry transparency. via AI News (author: AI Base)