Anthropic will begin embedding imperceptible watermarks in text generated by its Claude AI models, starting with versions released on or after August 2. The watermarks, which do not alter readability or meaning, will persist through some edits and can be detected by third-party tools. The company states that heavy editing or paraphrasing may remove the mark, and its presence does not definitively confirm AI use.
The feature applies globally to all text generated by Claude, including outputs accessed via cloud providers. Anthropic has committed to adding watermarking to older models and plans to provide detection tools to external parties. This move aligns with the company’s obligations under the European Union AI Act, which emphasizes transparency in AI-generated content.
The publishing industry has faced growing concerns over undisclosed AI-generated writing. In July, a literary agent withdrew support for the crime novel Call Me, I’ll Hide the Body after allegations of AI authorship, though the author, Jerry Falade, denied using AI. Similarly, Hachette halted publication of Mia Ballard’s Shy Girl in March following AI use allegations, which Ballard attributed to a freelance editor’s unapproved inclusion of AI-generated material.
Anthropic’s watermarking system is designed to help publishers, educators, and institutions investigate cases where AI-generated text may have been misrepresented as human-authored. However, the company acknowledges that the watermark is not foolproof, as determined efforts to alter or remove it could succeed.
The initiative reflects broader industry and regulatory efforts to address the ethical and practical challenges posed by AI-generated content. While watermarking technology is not new, its application to AI text represents a significant step in distinguishing between human and machine-generated writing in professional and creative fields.