Anthropic Adopts EU Practices on Labeling AI-Generated Content

Starting August 2, 2026, the EU AI Act’s transparency requirements for AI-generated content entered into force. Anthropic, the developer of Claude, has formally signed the Article 50(2) Code of Practice on Transparency of AI-Generated Content, which we analyzed in our earlier publication on the EU Code of Practice, and has published its full implementation roadmap. Here is what businesses need to know.

AI-Generated Content labeling framework

Article 50 of the EU AI Act imposes two structurally distinct obligations:

• The provider obligation under Article 50(2) requires AI developers to mark all synthetic outputs (text, image, audio, video) in a machine-readable format – with no restriction to content of public interest. 
• The deployer obligation under Article 50(4) is narrower: businesses that publish AI-generated text must visibly label it only if it concerns “matters of public interest.” Business communications, internal documents, technical output, or commercial copy carry no such visible labeling requirement.

Anthropic signed the Code as a provider, so Claude’s watermarks will be present on all generated text – regardless of its purpose and regardless of where in the world it is produced. This goes well beyond the EU’s deployer-level visible labeling obligation and extends beyond the EU’s territorial scope.

What Anthropic Has Committed To

The commitments cover all Claude products – the API, Claude.ai, Claude Code, Claude Cowork, and Claude Tag – applied worldwide. Models launched on or after August 2, 2026 support marking from day one. Models released before that date have until December 2, 2026 under the EU’s AI Omnibus grace period. Signatories may rely on the Code to demonstrate compliance uniformly across all EU Member States, though the Code is not a conclusive safe harbor, and national authorities retain the right to scrutinize actual implementation.

How the Marking Works

Anthropic uses two complementary techniques:

• Embedded text watermarks are woven imperceptibly into all generated text at the model level, persisting through copying For supported file formats (.svg, .png, .jpg), Claude attaches cryptographically signed C2PA provenance metadata, enabling third parties to verify AI origin and detect tampering.
• Detection of both mark types will be available free of charge, with full interoperability solutions due by February 2, 2027. Anthropic has also committed not to develop or distribute tools that strip or circumvent its own markings.

What This Means for Businesses Deploying Claude

A detected Claude mark confirms that content was processed by AI. Claude may have been used to edit, translate, or reformat content that originated elsewhere. Conversely, the absence of a mark does not rule out AI involvement: marks may be missing on pre-marking model outputs, heavily edited text, or very short passages. 

Per the EU Commission’s July 20, 2026 final Guidelines, AI-generated translations now fall within the “standard editing” exemption. When Claude is accessed via cloud integrations (AWS, Google Cloud, Microsoft Foundry), signed file metadata may also not be available on all platforms.

Anthropic’s commitments cover the provider-level obligations under Article 50(2). They do not discharge deployer obligations. Under Article 50(4), deployers must visibly label AI-generated text published for public interest purposes – an exception applies only where content undergoes genuine human editorial review with a named person holding editorial responsibility (pro-forma review does not qualify).

Article 50 has extraterritorial reach: any organization whose outputs are used in the EU falls within scope, regardless of where it is established. Non-compliance risks fines of up to EUR 15 million or 3% of worldwide annual turnover. Each business using Claude must independently assess its own obligations.

REVERA’s Arbitration & IT Disputes practice is ready to assist in assessing your obligations under the EU AI Act in connection with AI-generated content, including compliance structuring for businesses using generative AI tools.