Anthropic said content generated by its Claude chatbot will carry machine-readable markers worldwide starting Aug. 2, 2026, including invisible watermarking in text and signed provenance records attached to files.
The move is one of the first major responses from a large AI company after the transparency obligations in Article 50 of the European Union’s AI Act took effect. Anthropic said the marking system will apply across Claude products and access channels globally.
Two marking systems will be used
According to Anthropic’s documentation, Claude’s marking setup has two parts.
The first is embedded text watermarking. Anthropic said it weaves hard-to-notice signals directly into generated text without changing the meaning, quality, or readability of the response. The marker is meant to remain with the content when the text is copied and pasted elsewhere. Because the watermark is applied at the model level, it will work across Claude products and interfaces.
The second is signed provenance metadata for file types such as SVG, PNG, and JPG. This uses the open standard from the Coalition for Content Provenance and Authenticity, or C2PA, to attach a digital signature showing that Claude handled the file and to help detect whether the file was later modified.
Anthropic said the mechanism will cover the Claude app, its API platform, Claude Code, and services delivered through cloud partners including AWS, Google Cloud, and Microsoft Foundry.
New models will support it at launch
Anthropic said models released after Aug. 2, 2026 will support the markers from the moment they go live. Older models are also being retrofitted. The company said it will publish technical documentation for its detection methods so users and third parties can verify whether content came through Claude.
Linked to Article 50 of the EU AI Act
Anthropic’s announcement is tied to Article 50 of the EU AI Act. That provision took effect on Aug. 2, 2026 and requires AI providers to place machine-readable markers on generated or manipulated synthetic content so downstream users, platforms, and regulators can identify AI-produced material.
The source material says the EU’s AI Omnibus provisional agreement, passed in May this year, gives existing systems a grace period until Dec. 2. The European Commission has also introduced a voluntary Code of Practice on AI-generated content transparency, encouraging companies to use open standards and common labeling icons. Anthropic is listed among the signatories.
Rather than limiting the change to Europe, Anthropic chose to apply the policy globally.
A marker does not prove original authorship
Anthropic also said the system is not a foolproof detector.
A detected marker only shows that content was processed by Claude. It does not prove Claude was the original author, since users may rely on the model for proofreading, translation, or summarizing text written by people.
The absence of a marker also does not prove human authorship. Outputs from older models, heavily rewritten passages, and very short text may not carry markers. For images, metadata can disappear after a screenshot or a format conversion.
Outside criticism points to easy workarounds
Tech outlet The Register said prior research has already shown image watermarks can be removed. It also said optical character recognition tools can get around text watermarking, and that ready-made tools for stripping C2PA metadata are already available online.
Its report added that the policy could make open-source models without watermarking more attractive to some users, because some people may not want readers to know AI was used to write the content.
Transparency controls are becoming a competitive front
Even with technical gaps, AI content marking is moving into the mainstream. For spam content operations, academic plagiarism, and deepfake fraud, watermarking at least raises the cost of manipulation. For ordinary users, it adds one more clue when trying to trace where an article or image came from.