The AI Act Delay Did Not Move These Two December Dates
Dr. Abeer Alshammari · Published 11/4/2026
The EU's Digital Omnibus on AI bought high-risk system providers a lot of time. It did not buy everyone time. Two obligations land on 2 December 2026, less than a month from now, and both are easy to miss precisely because the headline story was about delay.
What moved, and what did not
Under the AI Omnibus, the high-risk deadlines moved from 2 August 2026 to 2 December 2027 for stand-alone (Annex III) systems and to 2 August 2028 for AI embedded in regulated products (Annex I). Regulatory sandbox deadlines also slipped, to 2 August 2027. Those are the dates most compliance calendars were rebuilt around.
Two items kept an earlier date. First, a new Article 5 prohibition covers AI systems that generate child sexual abuse material or non-consensual intimate imagery, including so-called nudifier tools, and it applies from 2 December 2026. Second, the grace period for marking synthetic content as machine-generated ends on the same date, but only for generative systems that were already on the market before 2 August 2026. A system placed on the market on or after that date gets no grace period.
Why these two are governance problems, not legal footnotes
The prohibition is drafted around foreseeable outputs. As reported by law-firm summaries of the agreement, a safe harbour exists for systems with effective technical safeguards that reliably prevent such content. That is an evidence question: a provider will need to show what safeguards exist, how they were tested, and how failures are detected. A policy statement will not carry that burden.
The marking obligation is an engineering question. Whether you provide an image, audio, video or text generator, you need to know which of your systems fall inside the grace period, which do not, and what mechanism makes output identifiable. If your product line includes a model launched after 2 August, the clock never gave you any slack.
A four-week readiness check
For organizations that provide or fine-tune generative systems, a short, evidence-based review is proportionate:
- Inventory by launch date. List every generative system and the date it was first placed on the EU market. This one field decides which marking rules apply.
- Test the safeguards you claim. Record red-team results against prohibited-content categories, and keep the logs. The safe-harbour argument depends on documented, reliable prevention.
- Verify the marking mechanism end to end. Check that it survives the formats your customers actually use, such as re-encoding, cropping or copy and paste.
- Check your supply chain. If you build on a third-party model, ask the model provider for its position in writing and map where responsibility sits in your contract.
- Assign an owner. Both items need a named accountable person and a review date, not a shared mailbox.
The wider lesson
Deadline relief creates a quiet risk: teams stop looking at the whole regulation and track a single date. The Omnibus also changed other things, among them how AI literacy is framed and how small mid-cap companies are treated, and anyone relying on a summary should read the adopted text for their own role in the value chain. Governance programmes that run off one headline date are fragile. A better habit is to maintain a dated obligations register in which each requirement has an owner, an evidence source and a status, so that a delay on one line does not erase the others.
Delay is not the same as dismissal. High-risk obligations are still coming, and the December dates show that the Union is willing to move fast where harm is concrete. Use the extra time on the high-risk side to build evidence properly, and spend the next four weeks on the items that did not move.
This article is general information, not legal advice. Check the adopted regulation and your own role under it.
Try it yourself
An interactive CyberAbeer experience for this topic is in development.
Coming soon