Anthropic embeds invisible watermarks in all Claude-generated text to comply with EU AI Act

Watermarking system announced

Anthropic announced on August 11 that its Claude AI models will embed invisible watermarks in all generated text, complying with the EU AI Act's Transparency Code (Article 50(2)) that took effect on August 2. Models released after August 2 already include the watermarking technology, which embeds a signal directly into the sequence of tokens the model outputs. The watermark is imperceptible to humans and does not alter the meaning, quality, or readability of the text.

Because the watermark is part of the text, it will travel with the text when it's copied and pasted elsewhere, and may persist through some 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.

Global scope and product coverage

The measure applies globally, not only within the European Union. Watermarking will be present across Claude, the Claude platform API, Claude Code, Claude Cowork, and Claude Tag, as well as in outputs from Claude models accessed through third-party platforms including AWS, Google Cloud, and Microsoft Foundry. For generated files, Anthropic will attach digitally signed provenance metadata using the C2PA open standard, the same standard used by Adobe, Microsoft, and much of the photography industry. Older models will receive the watermarking progressively, with the EU setting a deadline of December 2 for compliance, which Anthropic plans to meet.

Limitations and caveats

Anthropic has been explicit about the system's limitations. A detected watermark indicates that content may have been processed by Claude but is not conclusive proof that Claude authored it. Users who employ Claude to edit, translate, summarize, or convert human-written text may find the resulting output carries the Claude mark even though the underlying ideas originated from another source.

A detected mark provides a signal that content was processed by Claude, but is not fully conclusive.

Conversely, the absence of a watermark does not prove a text was not generated or processed by AI. The mark can disappear if text is heavily edited, paraphrased, translated, mixed with other writing, or if the passage is too short to leave a reliable signal. File metadata can be lost through format conversion, screenshots, re-saving, or generation on a platform that does not support the marking.

Regulatory context and industry landscape

The watermarking responds to Article 50 of the EU AI Act, which requires synthetic content to be automatically detectable and deepfakes and certain texts on matters of public interest to be properly labeled. Anthropic signed the EU's Code of Good Practices on Transparency of AI-Generated Content. Other companies including OpenAI, Google, Meta, Microsoft, Black Forest Labs, and Synthesia have also committed to the EU code. Google's SynthID system, which serves a similar function, has been adopted by Apple, ElevenLabs, Kakao, Nvidia, and OpenAI. Separately, AI music platform Suno announced it will mark tracks created on its platform, and newsletter service Substack partnered with Pangram to flag AI-generated content, with Substack CEO Chris Best calling out "Claudefishing," a term for people using AI to generate content.

User reactions and what comes next

The announcement has drawn criticism on social media platform X, where some users expressed opposition and others said they would stop using Claude. Concerns center on the risk that text merely edited or refined through Claude could be falsely flagged as AI-generated. Anthropic has not disclosed the technical mechanism for embedding the watermark, presumably to make removal more difficult. The company said it is working to enable users and other entities to detect the marks and metadata provenance, with a timeline of a few weeks for publishing detection details.