New EU rules to unmask deepfakes and AI text
The European Union has unveiled the first draft of its Code of Practice on Transparency, designed to enforce the AI Act’s requirements for marking synthetic content. This framework establishes technical standards for providers to embed machine-readable watermarks and mandates that deployers clearly label deepfakes for human audiences. With a strategic focus on interoperability, the code aims to build a robust digital provenance chain that protects public discourse against misinformation.
Points clés
- The Code of Practice was developed to facilitate compliance with Article 50 of the EU AI Act regarding content transparency.
- AI providers are required to adopt a multi-layered marking approach, combining metadata, digital signatures, and imperceptible watermarks.
- For text and cases where watermarking fails, providers must use fingerprinting, logging, or hashing for future verification.
- Developers of open-source models are expected to encode structural markings directly within model weights.
- Providers must offer free public APIs or detection tools to allow authorities and users to verify AI-generated content.
- Technical solutions must prove robustness against “adversarial attacks,” such as pitch shifting in audio or mirroring in images.
- Deployers must use a visible “AI” icon to disclose deepfakes and synthetic text at the moment of first exposure.
- Audio deepfakes under 30 seconds require an audible disclaimer at the beginning of the clip.
- An exception to disclosure exists for AI-generated text if it has undergone significant human review or editorial control.
- Small and medium-sized enterprises (SMEs) will be granted “proportionality” to ensure the compliance burden is manageable.
À retenir
So, the EU is finally teaching us how to spot a fake, because apparently, our eyes and ears just aren’t cutting it anymore. We’re moving toward a future where every digital image comes with its own birth certificate and every podcast starts with a robotic “I’m not real” warning. My advice for the average user? Just assume everything on the internet is a hallucination until you see a certified “Human-Reviewed” badge—even then, don’t hold your breath. It’s heartening to know that while AI is busy learning to mimic us perfectly, we’re busy designing the world’s most sophisticated “kick me” signs for lines of code.
Sources
Quiz sur le document: 10 questions
