Anthropic will embed a hidden digital signature in every piece of text and file produced by its Claude models, a move designed to satisfy the European Union’s new AI transparency rules. The change takes effect for all models released after the EU AI Act’s Transparency Code came into force on August 2, and it will appear automatically across Claude’s API, Code, Cowork and Tag interfaces.

Why the EU rule matters

The Transparency Code obliges developers to mark AI-generated or edited content so downstream systems can spot its synthetic origin. The rule aims to curb misinformation and give users a clear signal when they are dealing with machine-crafted material. Anthropic’s response—model-level watermarking—places the marker directly into the output rather than tacking it on in a product UI.

Other AI giants are making similar promises. Google, Meta, Microsoft, OpenAI and Black Forest Labs have all said they will meet the EU’s standards. For non-text assets, Anthropic will use the C2PA open standard, a publicly documented method for embedding provenance data in images, video and other media. By following an open spec, Anthropic hopes its markings will work with tools journalists, platforms and fact-checkers already use.

How the watermark works

Anthropic says the watermark is baked into the model’s text-generation process. In practice, the signature becomes part of the token stream the model emits. When a user copies the text into a word processor, an email, or a social-media post, the hidden pattern travels with it. The company notes the watermark can survive modest editing, but it hasn’t disclosed how many characters must change before the mark disappears.

The “Claudefishing” concern

Anthropic’s rollout arrives as alarm grows over the misuse of large language models (LLMs). Substack CEO Chris Best recently coined “Claudefishing” to describe using AI to craft deceptive content that mimics a particular author’s voice. As synthetic text becomes harder to distinguish from human writing, automated detection of provenance becomes a crucial line of defense.

Anthropic isn’t alone in adding safeguards. The music-generation platform Suno and the newsletter service Substack have introduced flagging mechanisms that alert readers when a piece was AI-created. By embedding the watermark at the source, Anthropic aims to make the signal harder to strip and easier for downstream filters to act on.

Key Takeaways

  • Regulatory Compliance: Anthropic is implementing watermarking to meet the EU AI Act’s Transparency Code, which took effect on August 2.
  • Model-Level Integration: Watermarks are applied at the model level, so they persist across all products (API, Claude Code, etc.) and survive copy-paste actions.
  • Standardized File Marking: For files, Anthropic will use the C2PA open standard to ensure interoperability and transparency in digital content.