Anthropic’s Claude Will Add Watermarks to AI-Generated Text and Files

Anthropic has announced that its AI model, Claude, will begin watermarking text and files generated by the system. This new feature is intended to comply with EU regulations requiring transparency in AI use. The EU's Code of Practice on Transparency of AI-generated Content mandates that companies informing customers when they are interacting with AI and include watermarks on AI-generated content.

Additionally, consumers need to be aware of deepfakes and emotion-recognition and biometric-categorization tools. Watermarking is a process that has been used for centuries to authenticate documents, and AI systems can embed markers into text that remain even if the content is copied and pasted across different platforms. Claude models launched on or after August 2nd will be equipped with this watermarking capability, which includes text generated through the API, Claude, Claude Code, Claude Cowork, and Claude Tag.

The company has not yet provided specifics on how to detect these watermarks, but they will be present as embedded signals within the text that can be identified by machine systems. Even after copying and pasting or human editing, these watermarks will remain intact. Furthermore, image files created by Claude will have a signed provenance metadata, providing information about the image's origin, creator, and any modifications.

This cryptographic signature will break if the image is tampered with, alerting readers to potential alterations. Anthropic has cautioned that the presence of a watermark doesn't necessarily confirm that the content was created by Claude, nor does its absence guarantee Claude's involvement. For instance, if a writer submits their work to Claude for rewriting, the watermarked output would still contain the original writer's ideas.

Similarly, if someone takes Claude-generated content, edits it, and combines it with other text, the resulting document may still bear a watermark.