AI watermarks are now mandatory. Enforcement is still around the corner

Thesis: the obligation to label synthetic content will not hold up without detectability and penalties. The four-month transition period for older models shows that the EU calendar is a political compromise, not a technical one.

Article 50 of the EU artificial intelligence regulation has applied since 2 August 2026. It requires providers of systems that generate synthetic audio, image, video or text to label their outputs in a machine-readable format and to make them detectable as artificially generated or manipulated. The labels must be effective, interoperable, robust and reliable, "to the extent technically feasible".

That caveat is fair to engineers. No watermark survives aggressive compression, cropping or a text paraphrase run through another model. In practice, though, it shifts the burden of proof onto the regulator. The provider does not have to show that a label works in every case. The supervisory authority has to show that it failed. With hundreds of models and tens of thousands of generators, today's watchdogs do not have the staffing to carry that out.

The bigger problem is the calendar. Providers of systems, including general-purpose models, that placed content generators on the market before 2 August 2026 have until 2 December 2026 to meet the requirements of Article 50(2). Put differently: the entire fleet of models from 2023 to 2025 operates legally without watermarks for several more months. On the same day, 2 December, the prohibitions on systems generating non-consensual intimate content and material depicting sexual abuse of children only start to apply. The schedule works out so that we first legalise the delay and only then ban the most dangerous category.

This is not a complaint aimed solely at Brussels. The whole AI Act enters into force in stages, and its final part took effect on 2 August 2026. A year earlier, in August 2025, the rules on general-purpose models kicked in. That staggered timetable was the price of a political agreement. Today we are collecting its consequences in the form of questions about who measures compliance, and with what.

The industry is not waiting idle. In music, the RIAA and IFPI are pushing voluntary AI labels for tracks on streaming services. Deezer released a free AI detector for other people's playlists, and Sony developed a technique for estimating how much protected material went into a generated track. The market can produce a standard faster than the legislator. But a voluntary standard, without penalties and without a public register, is easy to sidestep for those with the most to lose from the truth coming out.

So the real test of Article 50 will not be settled on 2 August or 2 December. It will be settled when the first supervisory authority fines a provider for a label that "formally exists" but cannot be detected after a single file conversion. Without such a precedent, watermarks will remain what they are today: a declaration on paper, easy on the eye and inconvenient for fraudsters.