+90 312 219 54 51[email protected]Koç Towers, Söğütözü · Ankara
Banko Law Firm & Mediation Speak to a lawyer
Article · Technology, AI & Cybersecurity

Content Marking for Transparency under the EU AI Act: Anthropic's Implementation

Banko Law Firm & Mediation2 min read

The artificial intelligence company Anthropic has announced that it has signed the Code of Practice prepared under Article 50 of the European Union Artificial Intelligence Act (EU AI Act), which concerns the transparency of AI-generated content.

According to the company's statement, digital marking methods will be used to make content generated or processed through Claude models technically detectable. New Claude models offered on the EU market on or after 2 August 2026 are expected to support machine-readable content marking from launch, and work is reported to be under way to add these features to models released earlier.

Under the implementation, hidden mathematical watermarks that do not change the meaning, quality or readability of the content will be embedded in text generated by Claude. For supported image and file formats, digitally signed metadata on the origin of the content will be used. For visual content, the mechanism is planned to rely on C2PA (Coalition for Content Provenance and Authenticity), an industry standard for verifying the source and history of content.

Anthropic has stated that the implementation will not be limited to the European Union and will apply in other regions where Claude services are offered. The marking mechanism is expected to apply to the Claude API, consumer chat services and other Claude products, as well as where Claude models are accessed through cloud services such as AWS, Google Cloud and Microsoft Foundry.

It was emphasised, however, that these technical marking methods are not a detection mechanism that conclusively establishes that content was produced entirely by AI. For example, a marker may also be present where a human-written text is processed through Claude for translation, summarisation or proofreading. The presence of a marker alone will therefore not mean that the entire content was created by AI.

On the other hand, extensive rewriting, heavy editorial intervention, translation into another language or certain technical transformations may reduce the detectability of text watermarks.

Similarly, taking a screenshot of an image or changing its file format may cause the provenance metadata to be lost.

Anthropic also announced that it is developing tools enabling users and third parties to verify marks in content generated or processed by Claude and that technical documentation will be published. Developers and organisations building on Claude should separately assess the transparency obligations that may apply to them under Article 50 of the EU AI Act.

This development is one of the practical examples of a transparency approach that seeks to reveal, through technical means, the origin of content and the involvement of AI in generative AI systems.

Sources · Kaynaklar

  • AB Yapay Zekâ Yasası (EU AI Act) m.50
  • Anthropic açıklaması
This article is for general information only and does not constitute legal advice. The English text is a translation; in case of discrepancy, the Turkish text prevails.