The Article 50 Watermarking Grace Period Everyone's Misreading

There's a four-month grace period on part of Article 50. It runs until December 2, 2026. And almost everyone I've seen talk about it gets what it covers wrong — in one of two opposite directions, both dangerous.

Let me be precise about what the Digital Omnibus actually changed, because the gap between "what people think it says" and "what it says" is exactly the kind of thing that gets a company a fine.

What the Omnibus actually changed

The Digital Omnibus — the first set of amendments to the AI Act since it was adopted — created a narrow, specific carve-out. Generative AI systems that were already placed on the market before August 2, 2026 get until December 2, 2026 to comply with the machine-readable marking requirement in Article 50(2).

That's it. That's the whole change. One sub-article, one category of systems, four extra months. Everything else about Article 50 stayed exactly where it was: applicable from August 2, 2026.

The two ways people misread it

Misread #1: "Great, we have until December for Article 50." Wrong, and dangerously so. The grace period covers only the technical watermarking of generated content. The obligations a regulator or a user actually notices— telling people they're talking to a chatbot, labelling deepfakes — are not covered. Those hit August 2. If you tell your team "we have until December" and ship an unlabelled chatbot in September, you're non-compliant.

Misread #2: "We have to build watermarking infrastructure by August or we're sunk." Also wrong, in the opposite direction — and it causes teams to panic-prioritise the hardest technical piece over the easy, mandatory ones. If your system is already live before August 2, the machine-readable marking specifically is the one thing you have breathing room on. Spending August scrambling on watermarking while your chatbot still has no AI disclosure is optimising the wrong corner.

What gets the grace period

Exactly one thing: Article 50(2) machine-readable marking of generative AI outputs. This is the requirement to embed markers in the content itself — metadata, watermarks — so that software can detect the content was artificially generated. For systems already on the market before August 2, 2026, this is what moves to December 2, 2026.

What does NOT get the grace period

Everything else in Article 50 applies from August 2, 2026 with no deferral:

• AI interaction disclosure (Article 50(1)). Telling users they're interacting with an AI system — chatbots, voice assistants, support agents. No grace period.
• Deepfake labelling (Article 50(4)). Disclosing that image, audio, or video content was artificially generated or manipulated. No grace period.
• Emotion recognition & biometric categorisation notices (Article 50(3)). Informing people when they're subject to these systems. No grace period.
• Public-interest text labelling. Marking AI-generated text published to inform the public, absent human editorial review. No grace period.

These are also, notably, the cheaper and faster obligations to meet. Adding a disclosure line to a chatbot is an afternoon. The thing with the grace period — building robust watermarking into a content pipeline — is the genuinely hard part. The Omnibus gave relief on the hard part for existing systems, not on the easy parts.

Who actually qualifies

The grace period is narrower than "everyone gets until December." To benefit, your generative AI system has to have been placed on the market before August 2, 2026.

• Existing systems (live before August 2): watermarking deferred to December 2, 2026. Disclosures still due August 2.
• New systems (launched on or after August 2): no grace period on anything. Machine-readable marking applies from day one, alongside all the disclosure obligations.

What to do with this

The practical sequence for an existing generative AI system:

• Now through August 2: ship every disclosure. Chatbot AI notices, deepfake labels, emotion-recognition notices, public-interest text labelling. These are mandatory August 2 and most are quick.
• August through December 2: build and deploy machine-readable marking — metadata plus watermarking — using the Code of Practice as your spec.
• Don't invert that order. The single most common planning mistake is treating watermarking as the August emergency and disclosures as the thing you'll get to. It's the reverse.

The grace period is real and it's useful. But it's a four-month extension on one technical requirement for systems that already exist — not a general reprieve on AI transparency. Plan around the August 2 date, treat December 2 as a bonus on the watermarking piece specifically, and you'll be on the right side of both.