Webinar on the technical labelling of AI content

A dialogue between law and technology
Applicable from 2 August 2026, Article 50(2) of the AI Act requires providers of generative AI systems to label the content they produce, whether it be text, images, audio, or video. The obligation is legal, but its implementation is primarily technical.
This is the dialogue that the webinar aimed to foster. Betty Jeulin presented the legal framework. Anthony Level, co-founder of Label4.ai, and Cyprien Veyrat-Charvillon, Deputy CEO of KeeeX, detailed the labelling solutions: invisible digital watermarking for the former, and secure metadata for the latter.
The legal framework presented by Betty Jeulin
Structured around nine questions, Betty Jeulin’s presentation covered the following points.
A dual obligation. Labelling must be both machine-readable and detectable. The provider must therefore also provide a means of detection that yields results understandable to a human being.
A definition of 'provider' that is broader than you might think. A company that has an AI tool developed for its own employees, or that modifies an existing system with its own data before operating it under its own name, is a provider within the meaning of the AI Act. The same applies to a company that integrates a third-party LLM into a tool it puts into service under its own name.
A responsibility that cannot be delegated. The provider may entrust the labelling to a third-party solution, but they remain liable for it. Penalties can reach 15 million euros or 3% of total worldwide annual turnover.
A tight schedule. For AI systems placed on the market or put into service before 2 August 2026, compliance is required by 2 December 2026 at the latest.
A framework currently being refined. The webinar also addressed the scope of the Code of Practice on the transparency of AI-generated content, the relationship with the deepfake labelling obligation under Article 50(4), and the legal questions that remain open.
The technical aspect
Anthony Level explained how invisible watermarking adapts to each format and what it can currently withstand (compression, screenshots), as well as its limitations. Cyprien Veyrat-Charvillon presented signed metadata and demonstrated the value of the multi-layered labelling recommended by the Code of Practice. When a file loses its metadata, the watermark still allows its origin to be traced.
Watch the webinar replay