EU Releases Code of Practice for AI Content Labelling Effective June 10

The European Union has released the Code of Practice (CoP) under Article 50 of the EU AI Act, establishing mandatory transparency requirements for the labelling of AI-generated content. The guidelines differentiate between obligations for AI Providers (those building systems) and AI Deployers (those using systems to create and publish content).

Obligations for AI Deployers

AI deployers and platforms are now obligated to label AI-generated content, specifically targeting deepfakes and published text on matters of public interest.

• Deepfakes: AI-generated or manipulated images, audio, or video that resemble existing persons, objects, or entities must be clearly and visibly labelled.
• Public Interest Text: AI-generated text published without human review or editorial control must be disclosed.
• Disclosure Timing: Labels must appear at the latest during the first interaction or exposure. Late disclosure constitutes non-compliance.
• Visual and Audio Standards: For visual content, icons must be placed in an appropriate area (e.g., top right corner) and displayed at the beginning and at regular intervals. For audio, a spoken disclaimer in plain language must be provided at the start, with reminders for long-form content.

Obligations for AI Providers

To combat misinformation and fraud, the Code mandates a two-layer approach for marking AI-generated content:

1. Digitally Signed Metadata: Providers must record whether content is AI-generated or manipulated using secure, tamper-evident, and timestamped metadata.
2. Imperceptible Watermarking: Mandatory for all content except very short text (fewer than 200 tokens). This can be implemented via post-hoc watermarking or model-level watermarking during inference.

Providers are also required to ensure that existing metadata markings are not intentionally altered or removed during reprocessing.

Compliance and Accessibility

All labelling must comply with the European Accessibility Act and the Web Accessibility Directive, including requirements for audio descriptions, high-contrast icons, and screen-reader compatibility. Furthermore, by February 2, 2027, AI providers must implement at least one of four prescribed interoperability approaches for watermark detection.

Deployers should immediately review their content workflows to ensure that AI-generated assets meet the new disclosure and placement standards.

Source: Lowdown: The EU's new rules for labelling AI-generated content
Domain: medianama.com