The Article 50 Watermarking Grace Period Ends December 2: What It Covers (and What It Never Did)

When the Digital Omnibus went through this summer, it gave one part of Article 50 a four-month grace period. That grace period ends on December 2, 2026.

Most of the confusion back in June was about what the grace period covered. Now the question is simpler: does it apply to you, and if it does, what do you need to have running by December 2? For most companies, the honest answer is that it doesn't apply at all. For the ones it does apply to, it's a contained engineering job.

What ends on December 2

Article 50(2) requires providers of generative AI systems to mark what their systems produce in a machine-readable way, so software can detect that the content was artificially generated or manipulated. That covers synthetic audio, images, video and text.

The Omnibus gave generative AI systems that were already placed on the market before August 2, 2026 until December 2, 2026 to comply with that one requirement. That's the whole carve-out: one sub-paragraph, one group of systems, four months. On December 2 it closes, and machine-readable marking applies to those systems the same way it already applies to everything launched since August.

Who is actually affected

Three things all have to be true for December 2 to be your deadline:

• You're a provider of the system, meaning you built it or put it on the market under your name. Marking is a provider obligation under Article 50(2).
• The system generates synthetic content, such as images, audio, video or text.
• It was on the market before August 2, 2026. Anything launched on or after August 2 never had the grace period. It has needed marking from day one.

If you build on someone else's model, there's a split worth knowing. A model provider marking its own outputs helps you, but whether you count as the provider of your own generative system depends on what you've built and how you market it. If you sell a product that generates content under your brand, assume the obligation is yours until you've confirmed otherwise. Our provider vs deployer guide walks through where that line falls.

What never had a grace period

This is still the part people get wrong. Nothing else in Article 50 was deferred. These have applied since August 2, 2026:

• AI interaction disclosure (Article 50(1)): telling people they're dealing with an AI system, such as a chatbot or voice agent.
• Emotion recognition and biometric categorisation notices (Article 50(3)): informing people when they're subject to these systems.
• Deepfake labelling (Article 50(4)): disclosing that image, audio or video was artificially generated or manipulated.
• Public-interest text labelling (Article 50(4)): disclosing AI-generated text published to inform the public, unless a human has reviewed it and someone holds editorial responsibility.

If a team read "grace period until December" in June and parked all of Article 50, those disclosures have been overdue since August. They're also the quick ones. A chatbot disclosure is an afternoon of work, so if one is missing, fix that before you think about watermarks.

What "marking" means in practice

Article 50(2) says the marking has to be effective, interoperable, robust and reliable, as far as is technically feasible. Two documents now tell you what that looks like:

• The Commission's final Guidelines on transparency obligations, published July 20, 2026.
• The Code of Practice on marking and labelling, which the Commission and the AI Board confirmed as adequate in July.

The main point in both: no single technique is enough on its own. The Commission and the AI Board concluded that nothing currently meets all four requirements at once, so the expected approach is layered. In general that means digitally signed metadata plus an imperceptible watermark, with detection tools made available to check them. Simpler expectations apply where marking isn't practical, for example outputs that stay inside a closed environment or plain text that can't carry embedded metadata.

The Code is voluntary. Following it is still the most straightforward way to show a regulator that your marking is good enough, because you're pointing to a standard they've already accepted instead of arguing for your own.

What about content you already generated?

According to the final guidelines, for images, audio and video the date that matters is when the content was generated, not when it's published. Content generated before August 2, 2026 doesn't need to be marked retroactively. You don't have to go back through an archive of old images.

Text published on matters of public interest works differently. The relevant date is publication. Text generated before August 2 but published after it still needs the disclosure, unless the editorial-review exception applies.

What to have in place by December 2

If you're in the affected group, this is the short version:

• Confirm you're in scope. Provider, generative output, live before August 2. If any one is false, December 2 isn't your date.
• Check your disclosures first. If any Article 50 disclosure is still missing, that's your oldest gap, so close it before the marking work.
• Ship layered marking on new outputs. Signed metadata plus a watermark, following the Code of Practice. If you rely on a model provider's marking, get it in writing and check it survives your own pipeline (resizing, compression and format changes can strip metadata).
• Write down what you did and why. Which technique, which content types, any cases where marking isn't feasible and your reasoning. If a regulator asks, that document is your answer.

December 2 isn't a cliff for most companies, and for the ones it does affect, the scope is narrow. The work is a defined change to how new content gets marked, plus a short record of how you did it.