- Published on
Anthropic to watermark Claude-generated text to prevent AI content masquerading as original
- Authors
- Name
- DeepHorus
On Monday local time, Anthropic announced that its new Claude models will embed an "imperceptible watermark" in AI-generated text to prevent AI content from being passed off as original. The company said this watermark will not alter the meaning or readability of the text, and it will be retained even if users copy or paste it, and it may still exist after partial editing.
Anthropic said this move is part of its commitment to transparency under the European Union's Artificial Intelligence Act. New Claude models released on or after August 2 will support this feature, with older models to be added later. The watermark applies to content generated by Claude worldwide, including scenarios where it is used through cloud service providers. Anthropic plans to provide third parties with tools to detect the watermark.
Text watermarks are faint marks embedded in documents to prove authorship. AI text watermarks need to be tamper-proof while maintaining readability. Anthropic said its watermarks remain detectable after copying, pasting, or light editing, providing a technical means to trace AI content.
This technology could become an important turning point for the publishing industry. In recent years, publishing has been grappling with the challenges of AI-generated content. Last month, a literary agent withdrew representation after questioning whether the author of the crime novel "Call Me, I'll Hide the Body" had used AI. The book had previously attracted bidding for copyright from 14 publishers.
Earlier this year, Hachette pulled "Shy Girl" from shelves after it was accused of containing AI content. The author, Mia Ballard, denied using AI, saying a freelance editor had added it without authorization. These cases highlight the publishing industry's need for AI text detection tools.
Claude's watermarking provides a new tool for publishers and schools to investigate AI use, and makes it harder for AI text to be passed off as original. But the technology still has loopholes: substantial rewriting, translation, or mixing with other text could render the watermark ineffective. Moreover, detecting a watermark does not prove that Claude is the original author, because even using it solely for proofreading or translation would leave traces.
Anthropic is the second major AI lab to roll out text watermarking. In 2024, Google DeepMind announced the use of SynthID to watermark Gemini text and video, having previously offered the feature for AI images. As major companies make moves, text watermarking may become a standard feature of generative AI.
