Munich: Europe spent the first half of 2026 arguing about how much of the AI Act to postpone, and it settled on an answer that pleases nobody consistently. The Digital Omnibus pushed the heaviest obligations well into 2027 and 2028. It left one deadline almost untouched, and that deadline now arrives in under three months.
AI watermarking under Article 50 is the obligation that survived the delay. Providers of generative systems must mark synthetic output in a machine-readable format so that downstream software can tell whether a human or a model produced it. Systems already on the market before 2 August 2026 received a grace period that expires on 2 December 2026. After that date, legacy deployments carry the same duty as anything launched tomorrow.
Why the transparency clock kept ticking
The postponement debate focused on high-risk classification. Regulation (EU) 2026/1744 entered into force on 27 July 2026 and moved stand-alone Annex III obligations to 2 December 2027 and product-embedded Annex I obligations to 2 August 2028. Legislators justified the shift on the grounds that harmonised standards and conformity assessment infrastructure simply did not exist yet.
Article 50 escaped that logic because it demands no conformity assessment, no notified body and no risk classification. It asks providers to attach a marker. The transparency provisions sit closer to labelling law than to product safety law, and legislators concluded that industry could meet them without an institutional scaffold that does not yet exist.
That reasoning holds up for new systems. It holds up considerably less well for the retrofit cases the grace period was designed to cover.
Retrofitting is the hard part
Marking new output is straightforward. A provider signs metadata at generation time, embeds an imperceptible watermark in the pixel or waveform data, and ships both. The C2PA specification already handles the metadata side, and several major image and audio models emit compliant credentials by default.
The difficulty lies in systems built before anyone specified a marker. A model deployed in 2024 inside an enterprise document pipeline produces text and images through code paths that never anticipated a signing step. Adding one means touching inference infrastructure, storage formats and every downstream consumer that validates file integrity. Providers with three months left are discovering that the work is engineering, not compliance paperwork.
Robustness compounds the problem. The regulation expects markers to survive ordinary handling, which means compression, cropping, format conversion and re-encoding. Metadata alone fails that test immediately, because almost every social platform strips metadata on upload. Imperceptible watermarking survives more transformations but degrades under aggressive editing, and no technique currently withstands a determined adversary. The obligation therefore targets accidental laundering of provenance rather than deliberate removal, which is a reasonable scope for a first attempt and a modest one.
Text remains the weakest case. Watermarking generated prose relies on biasing token selection in patterns a detector can recognise later, and the signal collapses in short passages or after light paraphrasing. Providers can sign a document container, but the moment a user copies a paragraph into an email, the marker vanishes. Regulators know this. The omnibus negotiations did not solve it, and no amount of enforcement will.
The asymmetry that results deserves attention. Europe delayed the rules that would have constrained how AI systems make consequential decisions about credit, employment and public services. It kept the rule that labels the output. Provenance matters, and December’s deadline will improve it at the margins. But labelling synthetic content and governing automated decisions solve different problems, and the Union has now sequenced the easier one first while the harder one waits another fifteen months.
