Anthropic embeds an imperceptible watermark in Claude text, giving publishers a tool to detect AI-generated writing. The feature, tied to the EU AI Act, has limits: heavy editing or translation can erase the mark.
Anthropic will watermark text generated by its Claude models, embedding a pattern that survives copy-paste and some editing without altering readability. The move, announced Monday, applies to Claude models launched on or after August 2 and covers content produced worldwide, including through cloud providers. Anthropic said it plans to give third parties tools to detect the marks.
The watermark is part of the AI lab's commitments under the European Union AI Act. It is designed to be imperceptible to human readers but travels with the text when copied, making it harder to pass off AI-generated writing as original work. Older models will get the feature later, the company said.
The publishing industry has been wrestling with undisclosed AI use. In June, a book agent dropped support for the crime novel "Call Me, I'll Hide the Body" after questions about the author's use of AI. A bidding war had drawn 14 publishers before the agent apologized. The author, Jerry Falade, denied using AI. Earlier this year, Hachette pulled Mia Ballard's horror novel "Shy Girl" over AI-generated content. Ballard told the New York Times a freelance editor introduced AI material without her knowledge.
Anthropic acknowledged limits. Heavy editing, paraphrasing, translation, or mixing Claude's output with other writing can make the watermark undetectable, the company said. And a watermark does not prove Claude originally authored the text, because even proofreading with the AI can leave a mark.
Anthropic is the second major lab to introduce text watermarking. Google DeepMind began watermarking text and video from its Gemini app in 2024, using SynthID technology, after releasing an image version in 2023.
Publishers, schools, and universities now have a new tool for investigating suspected AI-generated work, though the loopholes mean determined users can still bypass detection.
Drafted by a large language model from the source reporting linked above, then screened by automated publishing checks. It is not read by a journalist before publication. Some articles cite our Alpha Score. Verify prices and figures against the original source. Educational coverage, not personalized advice.