Anthropic’s Claude will begin adding watermarks to text, images and files generated by its family of AI models, allowing consumers to verify whether content originated from a human or from artificial intelligence. The change enables the company to comply with European Union regulations governing the transparency of AI use.
The EU’s Code of Practice on Transparency of AI-generated Content requires companies that provide and deploy AI systems to inform customers when they are interacting with AI and to include watermarks on AI content. Consumers must also be told when they are exposed to deepfakes and to emotion-recognition and biometric-categorisation tools.
Watermarking is an authentication process that originated in Italy in the 13th century, though the technique works entirely differently for AI content. An AI system can automatically embed markers in text that indicate the content came from AI rather than a person. These signals may include altered spacing between words and numbers, word sequencing and other invisible markers that remain with the text even when it is copied and pasted across platforms. Readers cannot see these markers, but computer systems can.
Which Claude content will be watermarked
Anthropic said Claude models launched on or after 2 August would include watermarking. That covers content created through the API, Claude, Claude Code, Claude Cowork and Claude Tag. The measure will apply to all Claude-generated content wherever the AI system is offered, not only within the EU.
Claude will embed watermarks in text and in files, including .svg, .png and .jpg formats. With text, the signals are placed within letters and words, invisible to readers but detectable by machine systems. The markers will persist even if the text is copied and pasted from editors such as Windows Notepad or macOS TextEdit, and may not be removed through human editing.
“Watermarking will be applied at the model level, which means it will be present no matter which Claude product or surface the text comes from,” the company said.
How image provenance works
Image files created by Claude will carry signed provenance metadata, including information about where the image came from, who created it and whether it has been modified. If someone attempts to tamper with the file, for example to conceal that it was generated by AI, the cryptographic signature will break, revealing to the reader that the content has been altered. Anthropic said it would share details on detecting the watermarks in future.
AI transparency is a growing trend across the technology sector. Substack has partnered with Pangram to let readers know how much of a post was generated by AI. Suno recently announced changes to help listeners identify AI-created songs. LinkedIn users who suspect low-quality AI content can flag it, while Spotify’s new AI Persona feature allows listeners and creators to identify music made with AI.
Anthropic included a caveat in its announcement, noting that the presence of watermarking does not necessarily guarantee detection in every case. Claude models launched on or after 2 August will carry the new watermarking.
