Anthropic rolls out imperceptible watermark for AI-generated text

Anthropic is rolling out a new feature for its Claude AI models that embeds an imperceptible watermark directly into generated text. This development aims to increase transparency and make it more difficult for individuals to pass off AI-created content as their own, a growing concern in industries like publishing and education.

The watermark, which does not alter the text's meaning or readability, is designed to travel with the content even when copied and pasted, and may persist through some editing. This move is part of Anthropic's commitment to transparency, particularly in light of the European Union AI Act. Models launched on or after August 2 will support this feature, with plans to extend it to older versions. Anthropic also intends to provide tools for third parties to detect these watermarks.

The publishing industry has been grappling with allegations of AI-generated content. Recent cases include the crime novel "Call Me, I'll Hide the Body" and the horror novel "Shy Girl," where authors denied using AI, but concerns were raised about AI-generated material being introduced into the writing process.

While the watermarking offers a new investigative tool for publishers and educational institutions, Anthropic acknowledges potential loopholes. Heavy editing, paraphrasing, translation, or mixing Claude's output with other writing could make the watermark undetectable. Furthermore, the presence of a watermark does not definitively prove original authorship by Claude, as even AI-assisted proofreading can leave a mark.

Anthropic is not the first major AI lab to implement such technology. Google DeepMind announced in 2024 that it was watermarking text and videos generated by its Gemini models using its SynthID technology, following an image watermarking feature released in 2023.