Anthropic plans to embed watermarks in text and digital files generated by its Claude AI models, initially as part of its effort to comply with the European Union’s AI transparency requirements. The company says future models launched in the EU will include these markings, while already-released models will be updated during the transition period. Anthropic also says the marking system will ultimately apply to supported Claude models worldwide, regardless of where Claude is being used.
For text, Anthropic says the watermark will be imperceptible and embedded directly into the generated text, without changing its meaning, quality, or readability. The company says the mark can travel with text when it is copied and pasted and may survive some editing. For generated files, Anthropic plans to use digitally signed provenance metadata based on the C2PA standard where supported. The markings will cover Claude through its platform and products, including Claude Code and Claude Cowork, as well as third-party cloud providers offering Anthropic models.
The move is intended to make it easier to determine the origin and provenance of AI-generated material, but the technology has important limitations. The Register notes that it is not yet clear how difficult Anthropic's text watermark will be to remove, while existing provenance metadata can also be stripped using available tools. Anthropic itself acknowledges that detecting a watermark would not conclusively prove Claude created a piece of content, while the absence of a watermark would not prove that AI was not involved.
The development reflects a broader push toward AI transparency and content labeling in Europe. Under the EU's AI Act, providers face requirements to make AI-generated or manipulated content identifiable and to disclose when people are interacting with AI systems. Anthropic's approach could help establish a technical standard for tracing AI content, but it also raises questions about whether watermarks can remain reliable once content is edited, reformatted, or passed through other AI systems. Ultimately, the initiative shows that the challenge is shifting from simply generating AI content to establishing trust, provenance, and accountability for that content.
