AI Content: What Labeling Requirements Mean for Businesses

GAIGI Team

— 4 min de lecture
Partager

On March 5, 2026, the European Commission published a draft code of practice on the labeling and marking of AI-generated content. The text comes ahead of the August 2, 2026, implementation of the transparency requirements under Article 50 of the AI Act. Its main contribution is both technical and organizational: “marking” does not simply mean displaying a notice, but rather making the content’s origin detectable and organizing information for the public. Companies must therefore distinguish between the responsibilities of the system provider, the person deploying the system, and the publication chain.

Updated September 10, 2026. Regulation (EU) 2026/1744 amending the AI Act was published in the Official Journal on July 24, 2026, and entered into force on July 27. The new deadlines for high-risk systems are now set for December 2, 2027 (Annex III) and August 2, 2028 (Annex I). Full details in our reference article.

The Essentials in 30 Seconds

  • The second draft of the code, released on March 5, 2026 following an initial draft published in December 2025, supports the transparency requirements of Article 50.
  • These requirements will take effect on August 2, 2026.
  • The labeling of content generated or manipulated by AI must be machine-readable.
  • The draft describes a two-tiered approach: secure metadata and watermarking.
  • Content providers are required to label hyper-fabrications and certain texts relating to topics of public interest.

What Happened

The Commission is facilitating the development of a voluntary code to help providers and implementers prepare for labeling and tagging requirements. The document published on March 5, 2026 presents proposed solutions for implementing the transparency requirements. It does not replace Article 50: the key legal date to remember is August 2, 2026, the date on which the transparency rules for AI-generated content take effect.

The scope includes two issues that are often confused. The first is labeling by the provider of a generative AI system: the generated or manipulated content must be machine-identifiable. The second is labeling intended for the recipients of the content. The project specifically targets deepfakes and text publications on topics of public interest produced by deployers of AI systems. The same content may therefore require a technical element attached to the file or feed, as well as visible information depending on the context of distribution.

This distinction explains why a simple mention in an editorial policy is insufficient. A charter can guide an employee’s behavior; it does not constitute a detection mechanism. Conversely, metadata embedded in a file may be lost, obscured, or misunderstood by a person. A compliance strategy must integrate both dimensions.

Marking Means Designing a Technical Chain

The project describes a two-tiered approach. The first level is that of secure metadata. It allows structured information about the content’s origin to be associated with the content. The second is watermarking, which aims to embed a signal linked to the content itself. The text also mentions, on an optional basis, digital fingerprinting and logging, as well as detection and verification protocols. It encourages the use of open standards for tagging AI-generated content.

For management, these terms must be translated into concrete decisions. It is essential to determine at what stage the content is tagged: upon generation, export, publication, or at each of these stages. It is also necessary to identify the types of content to which the measure applies: images, audio, video, text, or composite content. Finally, it is necessary to test how the labeling behaves when a file is converted, cropped, compressed, embedded in a presentation, or republished by a third party. A technical solution is only useful if the company understands its scope and limitations.

Machine-readability also requires a verification architecture. Who will be able to read the signal? Which tool retains proof that content was generated? What information can be disclosed without revealing personal data, trade secrets, or security details? At this stage, the project does not specify a single metadata format or impose a watermarking algorithm. This is yet another reason to prioritize reversible, documented, and interoperable choices rather than a haphazard collection of tools without governance.

Tagging Means Informing at the Right Time

The second component is the design of visible information. For implementers, the project addresses the labeling of hyper-edits and published texts on topics of public interest. It also addresses the design and placement of icons, labels, and explanatory notes. The stated goal is to ensure a minimum level of consistency while allowing signatories to choose solutions tailored to their needs.

The practical challenge is placement. A notice relegated to the terms and conditions, displayed after a video is shared, or at the bottom of an inaccessible page does not have the same impact as a notice visible at the point of consumption. Product and communications teams must define the channels, formats, language, timing, and person responsible for this information. They must also anticipate cases where content passes through a campaign management tool, an agency, or a platform that they do not own.

The draft mentions specific rules for artistic, creative, satirical, or fictional works, as well as for text-based publications subject to human or editorial review. These categories should not be used as informal exemptions. On the contrary, they require a precise analysis of the purpose, the editorial process, and the method of distribution. When a company claims to exercise human or editorial control, it must be able to demonstrate what that control actually entails.

Shared Responsibility Along the Chain

A provider of a generative system does not always control the final distribution; a deployer does not always have control over the watermarking features included in the acquired system. This is why contracts and procurement processes must now address operational questions: What metadata or watermarking features are available? Can they be enabled by default? How are outputs logged? How will updates be communicated?

What to Do Now

  • Map Outputs — identify images, audio, video, text, and composite content generated or manipulated by AI.
  • Test the tagging — verify metadata, watermarking, file conversion, and detection capabilities in live streams.
  • Design the labeling — specify the locations, labels, and responsible parties for hyper-manipulations and the relevant texts.
  • Establish contractual terms — request from providers their technical capabilities, limitations, and the information necessary for implementation.
  • Document exceptions — for each editorial, creative, or satirical treatment, maintain the rationale and applicable process.

Sources

Lire la suite