Anthropic said on August 11, 2026 that it will embed an invisible watermark in the text and files generated by its Claude models. The move is driven by the EU AI Act's transparency code, in force since August 2, 2026, which requires marking content created or edited by AI. The company will apply the watermark globally, not only in Europe.
Key takeaways
- The watermark covers both text and files generated by Claude
- Models released after August 2, 2026 mark output automatically, with support extending to older models
- Watermarking works across Claude API, Claude, Claude Code, Claude Cowork and Claude Tag
- For files, Anthropic uses the open C2PA standard
- Other signatories of the code include Google, Meta, Microsoft, OpenAI and Synthesia
How the watermark works
The watermark is woven directly into the generated text and is meant to survive copy-paste and some editing. That sets it apart from simple metadata, which disappears once content is pasted into another document. For files, Anthropic relies on C2PA, an open standard that attaches signed provenance metadata to a file. Watermarking is applied at the model level and spans the whole product family: the Claude API, the Claude app, Claude Code, Claude Cowork and Claude Tag.
Compliance with the AI Act, but not only in the EU
The immediate trigger is the AI Act transparency code, effective August 2, 2026. It requires model providers to mark AI-generated or edited content in a detectable way. Anthropic is not limiting the watermark to the EU market — it will reach users worldwide. The company is not alone: Black Forest Labs, Google, Meta, Microsoft, OpenAI and Synthesia have also signed the code. Platforms beyond pure AI, such as Suno and Substack, announced similar steps.
Why it matters
Watermarking text is harder than watermarking images or audio — text can be rewritten, shortened or passed through another model, which strips the signal. Anthropic's move shows that EU regulation is reshaping products globally, because it is cheaper for firms to ship one standard everywhere than to split the market. The open question is how robust the watermark is in practice and how much editing is enough to remove it. That determines whether the mechanism is useful for detecting AI or merely a formal box-tick.
What's next
- Anthropic is to extend watermarking to models released before August 2, 2026
- The watermark's resilience to editing and paraphrasing remains unconfirmed and will be tested in practice
- Other signatories — Google, OpenAI, Meta — must deploy their own mechanisms, allowing a comparison of approaches
