The Digital Omnibus postponed the high-risk requirements and, in the same Regulation, introduced two prohibitions and a marking deadline that take effect on 2 December 2026.
If your company read this summer that “the AI Act has been delayed” and put the matter aside, this is the article you should read.
Regulation (EU) 2026/1744 of 8 July 2026 — the so-called Digital Omnibus — which entered into force on 27 July, did two things at once.
It postponed obligations for high-risk AI systems until 2 December 2027 for those covered by Annex III and 2 August 2028 for those linked to Annex I, subject to the relevant transitional provisions. And it established two new requirements for 2 December 2026.
They should not be confused, because they belong to different legal regimes and must be addressed separately:
- Two new prohibitions under Article 5, concerning the generation of intimate images without consent and child sexual abuse material. They affect those who develop or deploy systems capable of generating or manipulating images, video or audio.
- Machine-readable marking under Article 50(2), applied to generative AI systems already placed on the market before 2 August 2026. In other words, it is precisely the older product that faces the deadline now.
1. The Two New Prohibitions
Article 5(1)(ba) prohibits the placing on the market, putting into service or use of AI systems that generate or manipulate realistic images, video or audio depicting the intimate parts of an identifiable person, or that person engaging in sexually explicit activities, without their freely given, specific, informed, unambiguous and explicit consent.
Article 5(1)(bb) prohibits AI systems that produce child sexual abuse material within the meaning of Directive 2011/93/EU. Recital 11 of the amending Regulation expressly mentions “nudification” applications as an example of the risk it addresses.
So far, this is what one might expect. What deserves particular attention is the following paragraph.
2. The Hidden Risk: Foreseeable Outcomes
Article 5(1a) defines when the prohibition under point (bb) applies to the provider and covers two situations, not just one: where producing such material is the intended purpose of the system, or where it constitutes a reasonably foreseeable and reproducible outcome of the system without adequate technical safeguards.
That is where the real risk lies.
A general-purpose image generator may fall within a prohibition even if it was not designed to produce such content, where the outcome is foreseeable and reproducible and the safeguards are inadequate.
It is not enough for the provider to declare that the system is not intended for that purpose: its actual operation, the possibilities of misuse and the effectiveness of the technical measures adopted must be examined.
For the professional user deploying the system, the rule is narrower: the prohibition applies when the system is used for that purpose.
And the consequences matter, because infringements of the prohibitions under Article 5 fall within the highest tier of penalties under the Regulation, without prejudice to the applicable rules on determining penalties.
This is not a documentation defect: it is a prohibited AI practice.
If your company offers image generation or editing involving people — even as an ancillary feature within another product — the question it must be able to answer by 2 December is not merely whether it has prohibited those uses in its general terms and conditions.
It is what technical safeguards it has in place, how it has verified that they work and which foreseeable risks it has identified.
3. Marking Under Article 50(2), and What Is Excluded
The second obligation concerns transparency.
Article 50(2) requires providers of generative AI systems to ensure that synthetic content is marked in a machine-readable format and is detectable as artificially generated or manipulated, taking into account the characteristics of each type of content, implementation costs and the generally acknowledged state of the art.
The Digital Omnibus introduced the transitional regime under Article 111(4): systems placed on the market before 2 August 2026 have until 2 December 2026 to comply.
This deserves emphasis because it is a frequent source of confusion: the postponement of high-risk requirements does not allow transparency obligations to be postponed.
The Commission’s Guidelines on Article 50 — adopted on 20 July 2026 and published in the Official Journal on 12 November 2026 — clarify the scope of these obligations. This is where businesses can avoid unnecessary compliance work.
Standard Assisted Editing
The broadest exclusion is the one established by Article 50(2) itself: the function of standard editing assistance.
Examples of excluded activities include grammatical and spelling correction, formatting and format conversion, compression, noise reduction, minor colour adjustments and the transcription of conversations.
Outputs that do not substantially alter the input data provided by the user or their semantics are also excluded — subject to a case-by-case assessment — as are systems authorised under Union or national law for the detection, prevention, investigation or prosecution of criminal offences, subject to the relevant safeguards.
Very Short Texts
There is also a technical limitation: very short text — below approximately 200 tokens — cannot be marked using the current state of the art.
Consequently, the Code of Practice does not require watermarking in such cases, although Article 50(2) does not expressly exclude them.
Business-to-Business Applications
Regarding business-to-business applications, it is important to say what the Guidelines actually say, and no more.
There is no three-condition test. What applies is an assessment of proportionality.
Paragraphs 85 to 87 explain that, in business-to-business applications, the principle of proportionality and limited marginal benefit may justify excluding the obligation, that the costs may be disproportionate to the benefits, and that in narrowly defined circumstances less robust metadata marking may be sufficient.
For example, where the system is integrated into a closed and controlled environment.
Therefore, the fact that a company’s customers are businesses does not, by itself, justify an exception: proportionality must be assessed in the specific circumstances.
Other Important Clarifications
The Guidelines make two further points.
The exception relating to creative works should not be extended to informational or commercial materials.
And there is no obligation to retroactively label deepfakes generated before 2 August 2026.
Finally, two aspects that are often confused must be distinguished:
- The technical marking obligation falls on the provider.
- The disclosure and labelling obligations under Article 50(4) and (5) fall on the deployer.
They are related, but they are not the same.
4. The Code of Practice: What It Offers and What It Requires
The Code of Practice on Marking and Labelling of AI-Generated Content was published on 10 June 2026, and the Commission and the AI Board declared it adequate on 8 and 9 July 2026.
It covers providers’ marking and detection obligations under Article 50(2), as well as deployers’ disclosure obligations under paragraphs 4 and 5.
Adherence is not conclusive proof of compliance and does not prevent market surveillance authorities from investigating. Nor does it replace the need to verify that the measures actually work.
What it offers is regulatory predictability and relevance when assessing the operator’s conduct.
However, it is important to understand the commitments involved, because the Code is voluntary but, once signed, it comes with its own timetable.
Measure 3.4 establishes 2 February 2027 as the deadline for implementing an interoperability solution for detection mechanisms, with several permitted approaches.
Marking alone is not enough: the marking must be detectable. Until public industry standards are available, the Guidelines allow proprietary, third-party or shared solutions as a time-limited arrangement.
Anyone who addresses the December requirements without including this second date in their compliance plan will have to return to the same issue two months later.
5. Checklist for 2 December
Does your product generate or edit images, video or audio of people?
Even if it is a secondary feature, document which technical safeguards you have in place, which uses they prevent and how you have verified their effectiveness. The prohibition may extend to foreseeable and reproducible outcomes, not merely the system’s declared intended purpose.
Did you have a generative AI system on the market before 2 August 2026?
Check whether the transitional regime under Article 111(4) applies to you and, consequently, whether the marking obligation applies from 2 December.
Does your system provide standard assisted editing?
Spelling correction, formatting, compression, transcription and similar functions fall outside the marking obligation. Check whether your product genuinely falls within this category before launching a compliance project.
What does your system produce that reaches third parties?
Identify the outputs subject to marking and examine whether any exclusion applies. Do not assume that everything used internally is excluded.
Are you relying on proportionality in business-to-business applications?
Assess the costs against the marginal benefit and the characteristics of the environment. The fact that your customer is a business is not sufficient.
Are you invoking the creative works exception for commercial or informational content?
Review the legal basis: the exception does not extend to such materials.
Have you adhered to the Code of Practice?
Adherence does not exempt you from liability, but it helps structure your compliance position. And if you adhere, make a note of 2 February 2027.
Conclusion
The postponement of the high-risk requirements was real and gives operators of systems covered by Annex III or Annex I more time. What was not postponed is the set of obligations discussed here, and the paradox is that the same Regulation that granted that additional time created obligations that become applicable earlier.
If your company develops or markets generative AI features, the task between now and 2 December is not simply to review your terms of use. It is to be able to demonstrate what your safeguards prevent and to mark the content produced by your system — unless an exclusion applies, and those exclusions are broader than is commonly believed.
The first issue falls within the regime governing prohibited AI practices; the second concerns transparency obligations. They are two different legal issues and must be addressed separately.
Note: This article is intended for general information purposes and does not constitute legal advice concerning any particular product or system. The classification of an AI system and the applicable obligations depend on its intended purpose, its actual operation and the circumstances of the case.
Sources
- Regulation (EU) 2024/1689 (AI Act), Articles 5, 50 and 111.
- Regulation (EU) 2026/1744 of 8 July 2026, which entered into force on 27 July 2026 and amends the AI Act.
- Directive 2011/93/EU.
- European Commission Communication, Guidelines on the application of the transparency obligations for certain AI systems under Article 50 of Regulation (EU) 2024/1689, C(2026) 5054 final, of 20 July 2026, published in the Official Journal of the European Union, No. 6005, on 12 November 2026, pages 1–43, particularly paragraphs 76 and 85–87.
- Code of Practice on Marking and Labelling of AI-Generated Content, published on 10 June 2026 and declared adequate by the European Commission and the AI Board on 8 and 9 July 2026, particularly Measure 3.4.
