It may soon become much harder to have Claude serve as a “ghostwriter” for everything from novels to homework.
Anthropic announced on Monday local time that new Claude models will directly embed an “imperceptible watermark” in AI-generated text. Anthropic said the watermark will not alter the meaning or readability of the text. It will remain even if users copy and paste the text, and “may still be present after partial editing.”
A watermark is a faint pattern or text embedded in a photograph or digital document, typically used to establish ownership and prevent content from being copied or misappropriated.
Anthropic said the move is part of its commitment to further improve transparency under the European Union’s Artificial Intelligence Act. New Claude models released on or after August 2 will support the marking feature from launch, and Anthropic is also working to add the feature to older models. The watermark will apply to content generated by Claude worldwide, including content produced when Claude is accessed through cloud service providers. Anthropic also plans to provide third parties with tools to detect these watermarks.
AI text watermarking technology could become a major turning point for the publishing industry. In recent years, the industry has been working to address the challenges posed by AI-generated content.
Last month, a literary agent ultimately withdrew representation of the popular crime novel Call Me, I'll Hide the Body after someone questioned whether its author may have used AI. The book had previously attracted rights bids from 14 publishers and had already been sold, after which the agent apologized for the incident. Author Jerry Falade denied using AI to write the novel.
Earlier this year, Hachette also pulled author Mia Ballard’s horror novel Shy Girl after it was accused of containing AI-generated content. Ballard told The New York Times that she did not use AI to write the book, and said a freelance editor had added AI-generated content without informing her directly.
Claude’s new watermarking feature could provide publishers, schools, and universities with another way to investigate similar incidents, while making it harder for people to pass off AI-generated text as their own work.
However, the technology still has some vulnerabilities. Anthropic said that substantially modifying, rewriting, or translating Claude-generated content, or mixing it with other text, could make the watermark undetectable. In addition, even detecting a watermark cannot prove that Claude was originally the author of the text, because simply using Claude for proofreading or translation could also leave behind a watermark trace.
Anthropic is the second major AI lab to introduce text watermarking technology. In 2024, Google DeepMind, a subsidiary of Google, announced that it would use SynthID to watermark text and video generated by the Gemini app and web version. The company had previously introduced watermarking for AI-generated images in 2023.