December 2, 2026. That's the Date Nobody Is Talking About.

The May 7 Agreement Had a Surprise. Most Businesses Missed It.

When the European Parliament and Council reached a provisional agreement on the AI Act's Digital Omnibus in the early hours of May 7, 2026 — after a nine-hour trilogue under the Cypriot Council Presidency — the headlines were predictable: high-risk AI deadlines pushed to December 2027, relief for SMEs, another reprieve from the regulatory burden that has been building since 2024.^[1]

Buried in that agreement, in the same paragraph that set the new December 2027 timeline, were two other dates: December 2, 2026. Those are the dates that most businesses aren't talking about. They should be.

This article is for informational purposes only and does not constitute legal advice.

What December 2, 2026 Actually Means

Two distinct obligations land on December 2, 2026, and neither is part of the high-risk AI regime that everyone has been focused on delaying. They are both transparency and prohibition obligations — and they were not part of the discussion about whether to push deadlines back.

The first is Article 50 transparency — watermarking. The obligation to mark AI-generated content in a machine-readable format — so that detection tools can identify artificially generated audio, images, video, and text — was originally due to apply from August 2, 2026. The Omnibus agreement gives providers four more months. The new date is December 2, 2026.^[3]

That sounds like a reprieve. It is — a small one. The Commission originally proposed a six-month postponement. Parliament pushed for three. Four was the compromise. The practical effect: providers placing generative AI systems on the EU market now have until December 2, 2026 to implement the full Article 50(2) marking infrastructure — machine-readable labels embedded in outputs, metadata, content fingerprinting, and the detection tooling to support it.^[1]

The second December 2 obligation is new — and it is not a postponement. It is a ban.^[2]

The New Prohibition Nobody Had Before

The Omnibus introduces a ban on AI systems used to generate non-consensual intimate imagery and child sexual abuse material. This provision was Parliament's non-negotiable in trilogue — the thing that could have broken the deal. It made it through.^[1]

The drafting is broader than it sounds. The prohibition does not only catch purpose-built "nudifier" apps. It catches any AI system whose functionality renders the generation of non-consensual intimate imagery reasonably foreseeable. That is a foreseeability standard, not an intent standard. It does not matter whether the system's labelled purpose was something else entirely. If a generative image or video platform can, with reasonable effort, be used to create non-consensual intimate content, the provider and deployer must demonstrate that they took design-stage and deployment-stage measures to prevent it.

For businesses building or deploying generative image and video systems — marketing content tools, product visualisation platforms, synthetic media services — this is a compliance obligation that arrives on December 2, 2026. The ban applies regardless of the high-risk AI postponement. It is a prohibited practice, not a high-risk requirement.

The Code of Practice Problem Nobody Is Mentioning

Here is the detail that makes the December deadline genuinely tight: the AI Office's Code of Practice on Transparency of AI-Generated Content — the practical rulebook that tells providers exactly how to implement Article 50 marking — is not finished yet. A first draft was published in December 2025. A further draft is expected during May 2026. The final text is not anticipated before June 2026.^[1]

Providers will have a matter of weeks between the final Code landing and December 2, 2026 to implement the multi-layered marking infrastructure, metadata embedding, content fingerprinting, and detector tooling the draft Code contemplates. Industry submissions during trilogue had argued for a 12-month postponement precisely because of this — the guidance isn't ready and the implementation timelines are unrealistic. That position did not prevail.

If you are a provider of generative AI systems being placed on the EU market, the four-month extension is real but narrow. Do not assume you have until December to start. Start engaging with the draft Code now and plan to move fast once the final text is published.

Who Should Be Paying Attention Right Now

The December 2, 2026 date is the first binding new deadline for organisations producing or distributing AI-generated content — in marketing, media, customer service, or product surfaces.^[3] If that is your business, December 2, 2026 is your operational compliance priority for 2026, not December 2027.

Specifically:

Generative AI providers: Your Article 50(2) marking infrastructure needs to be operational by December 2, 2026. The four-month grace period is the only buffer you have. Start now.
Businesses deploying generative image or video systems: Check whether your tools' functionality makes non-consensual intimate imagery generation reasonably foreseeable. If yes, document what safeguards your vendor has built in and what you have implemented on your end.
Marketing and content teams using AI-generated visuals or video: The visible disclosure requirements under Article 50(3) — the human-readable labels your audience sees — are not part of the postponement. Those remain August 2, 2026.

The Honest Bottom Line

The May 7 agreement is genuinely good news for businesses with high-risk AI obligations. The 16-month extension to December 2027 gives real runway for conformity assessments, technical documentation, and the other heavy compliance work.

But two things in December 2026 were not part of that relief. The transparency watermarking deadline moved four months — not 16 — and the nudifier prohibition is brand new. Both arrived in the same agreement that gave everyone else more time. The businesses that plan for December 2027 while missing December 2026 are going to have an uncomfortable first week of that month.