Executive Summary

The EU Commission has published a binding code of conduct for labeling AI-generated content. The regulatory framework, developed by six experts with over 180 stakeholders, specifies the transparency requirements of the AI Regulation from August 2, 2024. Deepfakes, manipulated, and artificially generated texts in the public interest must be clearly marked. Interactions with chatbots must be machine-detectable. The code is based on voluntariness but enables participating companies to officially demonstrate their compliance with the legal AI Act obligations.

People

  • Stefan Krempl (Author, Heise News)

Topics

  • Artificial Intelligence
  • EU Regulation
  • Transparency and Labeling
  • Deepfakes and Manipulation
  • Digital Authentication

Clarus Lead

The code establishes binding technical standards for detecting AI-generated content in the EU for the first time – a decisive step in combating misinformation and manipulation in real-time. With a standardized EU label and multilayered machine-readable markings, the Commission establishes a practical compliance mechanism that provides companies with a legal incentive: those who participate can demonstrate that they meet AI Act requirements. This could increase market acceptance of AI systems, while simultaneously providing media, fact-checkers, and authorities with free access to detection tools.

Detailed Summary

The code is divided into two core areas of responsibility. Developers and providers of generative AI systems must implement technically multilayered markings: tamper-proof digital metadata with digital signatures and timestamps combined with invisible watermarks. For texts exceeding 200 characters that do not carry conventional metadata, watermarks are mandatory; detection tools may initially be restricted to verified experts. Additionally, optional fingerprints and logging solutions are possible, provided data protection is maintained. Providers may not distribute tools designed to circumvent these mechanisms.

Users and operators must unambiguously label deepfakes and non-editorially reviewed AI-generated texts in the public interest – using the new standardized EU label "AI + GENERATED" or "AI + MODIFIED" depending on the degree of modification. The label must be placed visibly from the first user encounter; for videos, it must be repeated at regular intervals. For audio deepfakes, acoustic warnings apply at the beginning and during longer sequences. Artistic, satirical, and creative works are subject to more flexible rules to avoid compromising artistic enjoyment. Traditional media services can refer to their own editorial control processes. A task force will continuously adapt the code to technological advances and develop an interactive label layer that provides users with detailed source data with a click.

Key Statements

  • Two-tier regulatory framework with technical standards for developers and labeling obligations for users
  • Standardized EU label as a uniform identification mark for AI content from August 2024
  • Free access to detection tools for supervisory authorities, media, and researchers is mandatory
  • Voluntary code provides compliance proof for AI Act requirements – incentive for company participation
  • Flexible rules for artistic content; traditional media can use established editorial processes

Critical Questions

  1. Evidence/Data Quality: What empirical tests has the Commission conducted with the two machine-readable layers (metadata + watermarks) to validate their effectiveness against technical circumvention attempts?

  2. Data Quality (Watermark Robustness): The code acknowledges that "to the current state of the art, no single method is sufficient" – how robust are watermarks against compression, cropping, and conversions in practice?

  3. Conflicts of Interest: Who finances and controls the task force that adapts the code and develops the interactive label layer? Are AI companies represented on this task force?

  4. Causality/Alternatives: Can a standardized label alone combat misinformation, or are additional measures needed (e.g., platform algorithms, media literacy)?

  5. Feasibility: How will small and medium-sized providers finance and implement two-tier marking systems – are there transition support measures?

  6. Side Effects: Could the labeling requirement for artistic works with more flexible rules lead to misuse if users can no longer distinguish whether they are viewing artwork or misinformation?

  7. Enforcement: What sanctions apply to providers who refuse to sign the code or ignore it – are they stronger than merely failing to receive AI Act compliance recognition?

Sources

Primary Source: EU Commission presents code of conduct for labeling AI content – Heise News, Stefan Krempl

Verification Status: ✓ 2024

This text was created with the support of an AI model. Editorial Responsibility: clarus.news | Fact-check: 2024