The marking system does not appear as a visible logo or label in the text produced by Claude. Anthropic says generated text will carry an embedded watermark designed to be detected by machines rather than noticed by readers. Generated files, meanwhile, will include digitally signed provenance metadata where supported.
The marking is applied at the model level. Anthropic says Claude models launched in the European Union on or after Aug. 2, 2026 will support machine-readable marking from launch. The company also plans to add marking support to older models during the applicable transition period.
Read also : Project Panama: How Anthropic bought and destroyed millions of books to train AI
The marks are intended to make it possible to determine whether content was generated by Claude without changing its appearance or meaning for the person reading it. Anthropic is also developing detection tools and says it plans to publish technical documentation to allow users and third parties to identify the marks.
Why Anthropic is doing it
The change follows the implementation of Article 50 of the EU AI Act, whose transparency requirements began applying on Aug. 2.
Under the rules, providers of generative AI systems must add machine-readable marks that enable AI-generated or manipulated content to be detected. The European Commission says the measures are intended to reduce the risks of deception and manipulation and help people distinguish AI-generated material from other content.
The European Commission’s Code of Practice on Transparency of AI-Generated Content sets out technical measures for providers, including marking AI-generated audio, images, video and text in a machine-readable format. The code itself is voluntary, but the underlying transparency requirements in Article 50 are legal obligations.
Read also : White House to hold AI safety talks with Meta, OpenAI, Google and Anthropic
The rules do not require previously generated content to be marked retroactively. The European Commission says content generated before Aug. 2 does not have to be labelled under the new rules. For generative AI systems placed on the market before that date, the EU has provided a limited transition period for the marking requirement, running until Dec. 2, 2026.
The mark is not a visible “AI” label
Anthropic’s approach is different from simply placing a notice saying that a text was produced by AI.
For Claude’s text, the company describes the system as an imperceptible watermark embedded directly in the output. For generated files, it uses signed provenance information where the relevant format supports it.
This means a person reading a Claude-generated paragraph would not necessarily see anything indicating that it came from an AI system. The information is instead intended to be detectable through appropriate technical tools.
Anthropic has also acknowledged that marking is not the same as providing definitive proof of authorship. Its system is designed to help identify Claude-generated material, but detection can have limitations, particularly when content undergoes substantial changes.
A broader EU push for identifiable AI content
Anthropic’s move comes as the EU begins enforcing a wider set of transparency requirements for AI systems.
From Aug. 2, providers covered by Article 50 must use machine-readable marks for AI-generated or manipulated content. Separate requirements apply to certain deployers, including obligations to disclose deepfakes and certain AI-generated text published on matters of public interest without human review or editorial control.
The European Commission says the purpose is to make it easier for people to recognize when they are interacting with AI or encountering AI-generated material, particularly in situations where synthetic content could otherwise be mistaken for human-created or authentic content.
Read also : OpenAI expands review after identifying additional AI agent containment breaches
The Commission’s rules also distinguish between marking and labelling. Machine-readable marks are aimed at enabling technical detection of AI-generated or manipulated content, while visible disclosures and labels are required in certain circumstances for people who encounter specified types of AI-generated material.
For Anthropic, that means Claude’s outputs are now being built with a technical signal that can identify their AI origin, while the broader European framework establishes when AI-generated material must also be disclosed to people.
