EU's AI Regulation Mandates Digital Watermarking for Synthetic Content

Article 50 of the EU's AI Regulation (AI Act) comes into force on August 2, 2026, requiring AI systems to affix machine-detectable identification marks to synthetic content they generate. In response, major frontier model developers and providers are implementing digital watermarking technology.

The purpose of marking requirements for synthetic content is to distinguish AI-generated text, images, audio, and other materials from human-created content. Against the backdrop of mounting concerns about fake news and deepfakes, efforts to ensure transparency of content distributed within the EU have led to this legislation. The AI Act is a comprehensive regulatory framework for AI established in this context, with Article 50 being a provision particularly focused on content transparency.

The approach adopted by major operators is a technique called "statistical watermarking." This technology embeds statistical patterns into the natural language generation process without affecting the quality or performance of the output text. In other words, from the user's perspective, usability remains unchanged, while the system can detect watermarks on the machine side.

Meanwhile, the open-source community has responded quickly to this development. Open-source AI models, by their nature, allow anyone to modify their internals, making it easier for techniques to remove or circumvent watermarks. This has raised concerns about the effectiveness of compliance and vulnerabilities in the watermarking technology itself.

One reason this issue is important is that digital watermarking is now legally mandated as an obligation. Previously, digital watermarking remained largely a voluntary initiative, but with its codification into EU regulation, all AI operators entering the European market face pressure to comply. For large-scale operators with global operations, this necessitates embedding compliance into product design from the outset.

Regarding the application to open-source models, discussions are expected to continue. Unlike closed systems, ensuring the effectiveness of watermarking in open-source environments is technically challenging, and detailed examination of regulatory scope and implementation methods is required. How a consensus is reached between EU authorities and the open-source community will be a key point of attention going forward.

It can be argued that AI-generated content identification technology is becoming indispensable infrastructure for maintaining a trustworthy information environment. The EU regulation may serve as a model case influencing regulatory discussions in other regions. How the proliferation of digital watermarking and competing technologies to circumvent it develops is deeply intertwined with the broader question of generative AI's trustworthiness.