The common understanding is that the EU AI Act was pushed back to 2027 and there is nothing to do until then. Half of that is right. The high-risk obligations did move. Two prohibitions did not, and they begin on 2 December 2026.
Regulation (EU) 2026/1744, the digital omnibus on AI, was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026. It delayed the high-risk requirements from their original date of 2 August 2026: Annex III systems, which cover employment, credit, education and biometrics, now apply from 2 December 2027, and Annex I regulated products from 2 August 2028. That is where the sense of a general pause comes from.
The same regulation added to the prohibition list. Article 5(1) gained points (ba) and (bb), with new paragraphs 1a and 1b, and those apply from 2 December 2026.
Point (ba) prohibits AI systems that generate or manipulate realistic image, video or audio content showing the intimate body parts of an identifiable person, or an identifiable person engaged in explicitly sexual acts, without that person's freely given, specific, informed, unambiguous and explicit consent. Point (bb) prohibits systems that generate or manipulate child sexual abuse material as defined in Directive 2011/93/EU.
The new paragraph 1a is the provision that should concern general-purpose product teams, because it sets a different test for providers than for deployers. For a provider, the prohibition applies where producing that output is either the intended purpose of the system or a reasonably foreseeable and reproducible outcome in the absence of adequate safeguards. A deployer is caught where it intentionally uses the system for that purpose.
The provider test is not about what you meant to build; it asks what your system reliably does when someone pushes it, and whether your safeguards hold. A Spanish marketing-technology company offering generative image editing to retail clients in Spain and the UK does not intend to produce non-consensual intimate imagery. If a tester can reproduce that output through its interface, intent is not the question that will be asked.
The original eight prohibitions have applied since 2 February 2025 and are unchanged: subliminal manipulation causing significant harm, exploitation of vulnerability, social scoring, untargeted scraping for facial recognition databases, emotion inference at work and in education, biometric categorisation inferring protected characteristics, predictive policing based solely on profiling, and real-time remote biometric identification in public spaces for law enforcement within narrow exceptions. The enforcement weight sits here: breach of a prohibited practice carries penalties up to EUR 35 million or 7 per cent of worldwide annual turnover, whichever is higher.
For an SME that has wrapped a third-party image or video model into a product, three things are worth doing before December. Establish whether you are a provider, a deployer, or both, because the test differs. Write down the safeguards that stop the prohibited outputs, with the date each was put in place and the evidence that it works, since "adequate safeguards" is the phrase the regulation turns on. Then test adversarially and keep the results, because reproducibility is the standard and your own records are the only account of it you will have.
Questions on the 2 December AI Act prohibitions
Was the EU AI Act delayed to 2027?
Half of that is right. The high-risk obligations did move. Two prohibitions did not, and they begin on 2 December 2026.
Which regulation moved the dates?
Regulation (EU) 2026/1744, the digital omnibus on AI, was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026.
When do the high-risk obligations apply now?
Annex III systems, which cover employment, credit, education and biometrics, now apply from 2 December 2027, and Annex I regulated products from 2 August 2028. Both moved from the original date of 2 August 2026.
Does the delay mean nothing applies until 2027?
No. The same regulation added to the prohibition list, and those additions apply from 2 December 2026.
What applies on 2 December 2026?
Article 5(1) gained points (ba) and (bb), with new paragraphs 1a and 1b, and those apply from 2 December 2026.
What does point (ba) prohibit?
AI systems that generate or manipulate realistic image, video or audio content showing the intimate body parts of an identifiable person, or an identifiable person engaged in explicitly sexual acts, without that person's freely given, specific, informed, unambiguous and explicit consent.
What does point (bb) prohibit?
Systems that generate or manipulate child sexual abuse material as defined in Directive 2011/93/EU.
Does the prohibition depend on what we intended to build?
For a provider it does not. The provider test is not about what you meant to build; it asks what your system reliably does when someone pushes it, and whether your safeguards hold.
What is the test for a provider?
Under the new paragraph 1a, the prohibition applies to a provider where producing that output is either the intended purpose of the system or a reasonably foreseeable and reproducible outcome in the absence of adequate safeguards.
What is the test for a deployer?
A deployer is caught where it intentionally uses the system for that purpose.
We wrapped a third-party image model into our product. Which are we?
Establish whether you are a provider, a deployer, or both, because the test differs.
Our tool is for retail image editing, not intimate imagery. Are we safe?
A Spanish marketing-technology company offering generative image editing to retail clients in Spain and the UK does not intend to produce non-consensual intimate imagery. If a tester can reproduce that output through its interface, intent is not the question that will be asked.
What are the original prohibited practices?
Subliminal manipulation causing significant harm, exploitation of vulnerability, social scoring, untargeted scraping for facial recognition databases, emotion inference at work and in education, biometric categorisation inferring protected characteristics, predictive policing based solely on profiling, and real-time remote biometric identification in public spaces for law enforcement within narrow exceptions.
When did the original prohibitions start?
They have applied since 2 February 2025 and are unchanged.
What is the penalty for a prohibited practice?
Breach of a prohibited practice carries penalties up to EUR 35 million or 7 per cent of worldwide annual turnover, whichever is higher.
What records should we keep on safeguards?
Write down the safeguards that stop the prohibited outputs, with the date each was put in place and the evidence that it works, since adequate safeguards is the phrase the regulation turns on.
Do we need to test the system adversarially?
Yes. Test adversarially and keep the results, because reproducibility is the standard and your own records are the only account of it you will have.
Where does ComplianceSME hold these obligations?
More from ComplianceSME
- A machine fired the drivers. The bill is 824,990,000 euro.
- Europe Blamed the AI Act. The Delays Were Data Protection Law.
- The Two Rulebooks, the One Person, and the Answer That Is Already Out of Date
- The 24 hour clock that starts on 11 September
- The leaders who called compliance too complicated, and what the work costs
- ComplianceSME Completes the Free Business Compliance Libraries of All Twenty-Eight Nations
- The Digital Omnibus deferral: what moved, what did not, and what applies now
- Italy's NIS2 basic measures, and what has to be working by 31 October
- Platform work's deadline is 2 December, and the rules reach past the couriers
- Software becomes a product on 9 December, and the liability follows it
- Poland's cybersecurity register closed on 3 October, and the real work starts now
- Every verified deadline between now and the end of December
- Pay gap reporting arrives before the software does
- Six questions Irish companies keep asking, answered with the section that settles each