AI Compliance · 2026-09-08 · 8 min read

EU AI Act: What Actually Changes on 2 December 2026

Two deadlines land on 2 December 2026 — the machine-readable marking grace period ends, and two new prohibited practices start at the €35M/7% tier. Here is what Regulation (EU) 2026/1744 actually says, including the penalty date almost every write-up gets wrong.

Two things happen to the EU AI Act on 2 December 2026. The grace period for machine-readable marking of AI-generated content ends, so systems that were already on the market before 2 August 2026 lose the only extension Article 50 was ever given. And two new prohibited practices — non-consensual intimate imagery and child sexual abuse material — start applying at the Act's highest penalty tier, up to €35 million or 7% of worldwide turnover.

That is roughly twelve weeks away. Neither change is well covered, because most reporting on the Digital Omnibus stopped at "the EU delayed the AI Act" — which is half true and, for anyone shipping an AI feature into Europe, dangerously imprecise. This is what the adopted text actually says, and what to do before December.

What the Digital Omnibus actually changed

The Digital Omnibus on AI is now law: Regulation (EU) 2026/1744, voted by Parliament on 16 June 2026, adopted by Council on 29 June, published in the Official Journal on 24 July, and in force since 27 July 2026.

It moved the high-risk regime and very little else. The dates it moved are now fixed, which is itself news: the Commission had proposed tying them to a standards-readiness trigger, and that mechanism was dropped. Annex III standalone high-risk systems apply from 2 December 2027; Annex I embedded systems from 2 August 2028.

What it did not touch is the part that binds most companies today. Article 50 came through unchanged. Article 5 prohibitions were not deferred — they were expanded. Article 6(3), the self-assessment route out of high-risk classification, survived the Commission's attempt to remove it. Regulatory sandboxes slipped to 2 August 2027.

If you build or deploy a user-facing AI feature and you are not in a high-risk category, the Omnibus gave you almost nothing.

Correcting a date almost everyone gets wrong

You will read, in a great deal of coverage — including, until this week, our own — that "the EU AI Act penalty regime applies from 2 August 2026." That is not right, and the distinction matters if you are reasoning about exposure.

Article 99, the general penalty regime, has applied since 2 August 2025, under Article 113(b). It did not switch on last month. What began on 2 August 2026 was the substantive Article 50 transparency duty and the Article 101 fines specific to general-purpose AI model providers.

The practical difference: your exposure under Article 50 is new, but the enforcement machinery it plugs into has been live for over a year, and national market surveillance authorities have had that time to organise. Do not plan around a grace period that does not exist.

The tiers themselves are unchanged: €35M or 7% for prohibited practices, €15M or 3% for most other breaches including Article 50, €7.5M or 1% for supplying incorrect information. SMEs pay the lower of the fixed sum and the percentage rather than the higher. The Omnibus added Article 99(6a) for small mid-caps.

Deadline one: machine-readable marking, 2 December 2026

Article 50(2) requires providers of generative AI systems to mark synthetic audio, image, video or text in a machine-readable format detectable as artificially generated. Not a visible label — an embedded, detectable signal. Watermarking, metadata, cryptographic provenance, or a comparable technique.

Article 111(4) gave exactly one concession: systems placed on the market before 2 August 2026 have until 2 December 2026 to comply with it. Anything launched after 2 August 2026 has owed machine-readable marking from day one.

This catches more teams than expected, because the obligation attaches to the generative system, not to whether you consider yourself an AI company. If your product shipped before August and generates images, writes copy, synthesises voice, or produces video, this deadline is yours, and the extension you have been relying on expires in twelve weeks.

Three things worth knowing before you scope the work:

  • Using a foundation-model API does not move the duty upstream. If you place a generative system on the EU market under your own name, you are the provider of that system. Your model vendor's watermarking may help you satisfy it; it does not discharge it.
  • Text is the hard case. Robust machine-readable marking of generated text is genuinely unsolved at the state of the art. The Act asks for techniques that are effective, interoperable, robust and reliable "as far as technically feasible," which is a real qualifier — but it is an engineering-judgment defence you have to be able to evidence, not an exemption you can assume.
  • Marking and disclosure are separate duties. Telling a user "you are chatting with an AI" satisfies Article 50(1). It does nothing for 50(2).

Deadline two: two new prohibited practices, same date

The Omnibus added two prohibitions to Article 5, and they apply from 2 December 2026:

  • Article 5(1)(ba) — non-consensual intimate imagery
  • Article 5(1)(bb) — child sexual abuse material

These sit in the €35M / 7% tier — the same band as social scoring and manipulative techniques, and more than double the Article 50 tier.

The part to read carefully is the new Article 5(1a), which sets how the prohibitions bite: an intended-purpose and foreseeable-outcome test. A general-purpose image or video generator is not prohibited because misuse is conceivable. The question is what the system is intended for and what outcomes are reasonably foreseeable from the way it is placed on the market.

That is a design-and-safeguards question, not a policy-document question. If you ship image, video or voice generation to the public, the twelve weeks are for evidencing what you actually do about it — input and output filtering, abuse reporting, red-team results against exactly these two categories, and the reasoning behind where you set the thresholds.

Article 4 was quietly rewritten

The Omnibus reworded the AI literacy obligation. It now reads as a duty to "take measures to support the development of AI literacy" — an obligation of effort, without a guaranteed level of competence for any individual.

That is a meaningfully lower bar than the original phrasing, and it is still binding on both providers and deployers, with no headcount threshold. It has been in force since 2 February 2025.

Read it as the standard shifting from "prove every employee is trained to a level" to "prove you run a programme and can show what you did." Records still matter; a defensible programme is still cheaper than arguing about the wording.

What to do in the next twelve weeks

  1. List every generative feature you shipped before 2 August 2026. That list is your Article 111(4) exposure, and it is the only one with a December date attached.
  2. Determine your marking approach per feature. Provenance metadata for images and video, vendor watermarking where it exists, documented technical-feasibility reasoning for text. Write down the reasoning at the same time — the evidence is the deliverable, not the decision.
  3. If you generate images, video or voice, run the two new prohibitions as an explicit review. Intended purpose, foreseeable outcomes, safeguards, and what your abuse reporting actually does.
  4. Confirm your Article 50(1) disclosures are live and correct. They have been owed since 2 August 2026. If any are missing, that is current exposure, not a December problem.
  5. Get your AI literacy programme on paper. Effort-based, but evidenced.
  6. Re-check your high-risk assumption. The dates moved to December 2027 and August 2028, and they are now fixed. If you assumed the trigger mechanism would slip them again, it will not.

The strategic read

The Omnibus is being sold as relief, and for anyone deep in Annex III conformity work it genuinely is — sixteen months of it. For everyone else it narrowed things. The transparency duties that apply to ordinary product teams were left alone, the prohibitions grew, and the high-risk dates stopped being negotiable.

The enterprise buyers asking AI-governance questions in security reviews have not slowed down either. That is the forcing function most teams actually feel first — long before a market surveillance authority calls.

FAQ

What happens on 2 December 2026 under the EU AI Act?

Two things. The Article 111(4) grace period ends, so generative AI systems placed on the EU market before 2 August 2026 must comply with the Article 50(2) machine-readable marking requirement. And two new Article 5 prohibitions — non-consensual intimate imagery and child sexual abuse material — begin to apply, at up to €35 million or 7% of worldwide turnover.

Did the Digital Omnibus delay the EU AI Act?

Only the high-risk regime. Regulation (EU) 2026/1744 moved Annex III standalone systems to 2 December 2027 and Annex I embedded systems to 2 August 2028, and made those dates fixed by dropping the proposed standards-readiness trigger. Article 4, Article 5 and Article 50 were not deferred, and Article 5 was expanded.

When did the EU AI Act penalty regime start?

The general penalty regime in Article 99 has applied since 2 August 2025, under Article 113(b). What started on 2 August 2026 was the substantive Article 50 transparency obligation and the Article 101 fines for general-purpose AI model providers. Coverage that dates the whole penalty regime to August 2026 is incorrect.

Does using OpenAI or Anthropic move the marking duty to them?

No. If you place a generative AI system on the EU market under your own name or trademark, you are the provider of that system and Article 50(2) attaches to you. Your model vendor's watermarking may help you meet it, but the obligation does not transfer.

What are the EU AI Act penalty tiers?

Up to €35 million or 7% of worldwide annual turnover for prohibited practices under Article 5; up to €15 million or 3% for most other breaches including Article 50; up to €7.5 million or 1% for supplying incorrect information. SMEs pay the lower of the fixed amount and the percentage. The Omnibus added Article 99(6a) covering small mid-caps.