Section 1: Rules for marking and detection of AI-generated and manipulated content applicable to providers of generative AI systems

Objectives

The Code provides voluntary guidance for demonstrating compliance with Article 50(2) and (5) of the European Union (EU) AI Act. It addresses transparency obligations for providers of systems generating synthetic text, images, audio and video. Adherence does not automatically establish legal compliance.

Recitals
The Code identifies misinformation, impersonation, fraud and consumer deception as risks arising from synthetic content. It advocates interoperable technical standards, cooperation across the AI value chain and proportionate implementation for smaller businesses.

Commitments

Commitment 1: Marking of AI-generated or manipulated content

Providers should implement machine-readable identification techniques. The Code generally specifies two complementary layers: digitally signed metadata and imperceptible watermarking.

Free-form text shorter than 200 tokens is treated differently because reliable watermarking remains technically challenging. Optional measures include fingerprinting, logging and additional provenance information identifying the AI system or provider.

Providers should preserve existing markings and discourage their deliberate removal.

Commitment 2: Detection of markings of AI-generated or manipulated content

Providers should offer detection mechanisms allowing users, regulators and other legitimate parties to verify whether content was generated or manipulated by their systems.

Detection results must be understandable and accessible. Privacy safeguards include data minimisation, secure processing and deletion of submitted content after detection.

Commitment 3: Measures to meet the requirements for marking and detection solutions

Technical solutions must satisfy four requirements: effectiveness, reliability, robustness and interoperability.

Providers should evaluate detection accuracy, resilience to content alterations and malicious attacks, and compatibility across systems. The Code establishes a 2 February 2027 milestone for implementing specified interoperability solutions.

Commitment 4: Testing, verification, and compliance

Providers should document compliance processes, undertake testing before deployment and periodically thereafter, monitor deficiencies and implement corrective actions.

The Code also addresses staff training and cooperation with market surveillance authorities. Independent testing and regulatory sandboxes are identified as possible assessment mechanisms.

Glossary

The glossary defines technical concepts including metadata, watermarking, fingerprinting, detection mechanisms and content provenance to support consistent implementation.
Section 2: Rules For Labelling Deep Fakes And AI-Generated And Manipulated Published Text Applicable To Deployers Of AI Systems

Objectives

This section supports compliance with Article 50(4) and (5), requiring disclosure of deepfakes and certain AI-generated publications concerning matters of public interest.

Recitals

The Code emphasises clear, consistent and accessible disclosure to reduce deception and protect public trust. It recognises different communication formats, audience needs and existing legal obligations.

Commitments

Commitment 1: Disclosure of deep fakes and published text

Deployers should use the EU’s optional AI icon or an equivalent label meeting specified design and placement requirements.

Visual labels must be recognisable, while audio-only deepfakes require audible disclosure. Information should be provided no later than first exposure, with accessibility safeguards.

Commitment 2: Internal processes

Deployers should establish proportionate compliance procedures, improve staff awareness, review disclosure practices and cooperate with authorities.

Commitment 3: Disclosure for artistic, creative and similar works

Artistic, fictional and satirical content receives adapted disclosure requirements intended to preserve the presentation and enjoyment of creative works.

Commitment 4: Human review and editorial control for published text

The Code addresses the exemption for AI-generated public-interest text subject to human review or editorial control where an identifiable person or organisation assumes editorial responsibility.