Anthropic Introduces an Invisible Watermark for Claude

Ixbt.com reports .

The new technology took effect immediately in all Claude models launched after August 2 this year. Developers are also actively working to add the marking system to older models. Importantly, the watermark will be automatically applied not only to texts created directly on the platform but also to all texts generated through cloud providers.

New Opportunities for Education and Publishing

Anthropic says it took this step to meet its transparency obligations under the European Union’s Artificial Intelligence Act (EU AI Act). For schools, universities, and major publishers, such marking could provide an additional tool for verifying the authenticity of work submitted by students and authors. This is particularly relevant because disputes related to artificial intelligence have recently increased in literary circles.

In particular, in July this year, a prominent literary agency refused to support the novel “Call Me, I'll Hide the Body” because of questions about whether the author had potentially used artificial intelligence. Hachette also withdrew the horror novel “Shy Girl” from publication amid similar allegations. In both cases, the authors firmly insisted that they had not used artificial intelligence to create the texts, but an atmosphere of distrust remained.

Technology Limitations and Future Plans

Nevertheless, experts warn that the watermark may not always identify a text’s origin with 100% accuracy. Anthropic acknowledges that intensive editing, paraphrasing, translation, or mixing a Claude output with other texts can remove the watermark. Moreover, detecting a watermark does not prove that the entire work was written by artificial intelligence, as it may also appear during text editing or translation.

In the future, Anthropic plans to provide third-party organizations with specialized tools that will allow them to detect these marks using readers and software. Notably, Anthropic is the second major company on the market to adopt such technology. Before this, Google DeepMind announced in 2024 that it had introduced the SynthID system for texts and videos generated through its Gemini model.