Anthropic Adding Invisible Watermarks to Claude-Generated Text
- By John K. Waters
- 08/24/26
Anthropic announced it is adding invisible watermarks to text generated by its Claude AI models, offering an early look at how major AI companies may respond to new European rules requiring machine-readable identification of AI-generated content.
The AI company detailed its approach as part of changes designed to comply with transparency requirements under the European Union's AI Act. The measures apply globally to supported Claude models, rather than only to users in Europe.
For text, Anthropic is using a version of SynthID-Text, an open source watermarking approach developed by Google DeepMind. Instead of adding visible labels or hidden characters, the system subtly influences the model's choices as it generates text, creating a statistical pattern that can later be detected.
The process takes advantage of the fact that large language models often have several plausible choices for the next token in a response. The watermarking system can influence those choices in ways that create a detectable signature while preserving the overall meaning of the text.
Anthropic says the watermark has no practical impact on the quality or content of Claude's output and does not increase the cost of using the model.
The company is taking a different approach for images. Claude-processed images will use the Coalition for Content Provenance and Authenticity (C2PA) standard to attach provenance information to supported image files.
The changes come as AI developers face growing pressure to make synthetic content easier to identify.
Article 50 of the EU AI Act requires providers of systems that generate synthetic audio, images, video, or text to ensure their outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. Anthropic has linked its watermarking changes directly to those requirements.
The rules could make watermarking a more common feature across generative AI services as other companies operating in Europe address the same requirements.
But watermarking AI-generated text presents challenges that do not exist in the same way for images or video. Anthropic acknowledges that its watermark is not intended to provide definitive proof that a piece of text was written by Claude. The company has also warned that quoted Claude-generated text could carry the watermark into another document, while text without a detectable watermark should not automatically be considered human-written.
The watermark may survive copying, pasting, and some editing, but more extensive changes to generated text can make detection more difficult.
The approach has also prompted criticism from some Claude users who are concerned about how watermarked text could be interpreted when AI is used for tasks such as editing, translation, or formatting rather than generating an entire document. Anthropic has said the presence of a watermark indicates that text was processed by Claude, not necessarily that Claude was responsible for its authorship.
For more information, read the Anthropic blog.
About the Author
John K. Waters is the editor in chief of a number of Converge360.com sites, with a focus on high-end development, AI and future tech. He's been writing about cutting-edge technologies and culture of Silicon Valley for more than two decades, and he's written more than a dozen books. He also co-scripted the documentary film Silicon Valley: A 100 Year Renaissance, which aired on PBS. He can be reached at [email protected].