In short
- Regulation (EU) 2026/1744, in force since 27 July 2026, inserts points (ba) and (bb) into Article 5(1), covering AI generated non-consensual intimate imagery and child sexual abuse material.
- Those two points carry their own compliance date of 2 December 2026, a year before the deferred high-risk dates of 2 December 2027 and 2 August 2028.
- It applies regardless of high-risk status. An organisation with 0 Annex III systems is still in scope if it operates image or video generation.
- Article 99(3) attaches the top penalty tier to Article 5 as a whole, reaching EUR 35 million or 7% of worldwide annual turnover, so the new points inherit it.
- The practical control is not a policy line. It is input and output filtering, plus a written account of what the filter is designed to stop.
Regulation (EU) 2026/1744 is remembered for what it delayed. It also added something, and the addition arrives first.
The prohibition it inserts into Article 5(1) of the EU AI Act as points (ba) and (bb), covering AI generated intimate imagery and child sexual abuse material, takes effect on 2 December 2026. That is twelve months before the high-risk obligations it shares a statute with.
01Why does a prohibition behave differently?
Because it is not scoped by classification.
Most of the EU AI Act works through a risk tier. A system is high risk or it is not, and the obligations follow. Article 5 does not work that way. It bans practices, and the ban reaches anyone performing the practice regardless of what tier their systems sit in.
That is why an organisation that concluded, correctly, that it operates no Annex III systems cannot conclude anything from that about Article 5.
02What are the dates that actually matter?
| Provision | Applies from | Scoped by risk tier? |
|---|---|---|
| Original Article 5 prohibitions | 2 Feb 2025 | No |
| New prohibition, Article 5(1)(ba) and (bb) | 2 Dec 2026 | No |
| High-risk, Annex III | 2 Dec 2027 | Yes |
| High-risk, Annex I | 2 Aug 2028 | Yes |
Two of those four rows are already unconditional and a third becomes unconditional in December. The deferral that dominated the coverage of Regulation (EU) 2026/1744 applies to the bottom two, and to nothing above them.
03Who is realistically in scope?
Anyone operating a generation capability, including through a supplier.
That includes consumer and business products offering image or video generation, platforms hosting user generated models, and organisations that expose a third-party generation model through their own interface. The last category is the one that gets missed, because the capability is bought rather than built and the risk register records a vendor rather than a function.
The drafting closes the obvious escape route. Points (ba) and (bb) reach material that is a reasonably foreseeable and reproducible outcome of the system, not only material it was built to produce. A general purpose image model that was never intended for this, but that reliably produces it when asked in a particular way, is not outside the text because of what its documentation says it is for. This is the same gap between stated purpose and actual behaviour that the labelling duties in Article 50 run into from the other direction.
The question worth asking is therefore not whether you built a generator. It is whether a person can produce such an image through something you operate.
04What does the penalty tier imply?
That this is not a risk to be accepted and monitored.
| Breach | Ceiling | Turnover basis |
|---|---|---|
| Prohibited practice, Article 5, including the new points | EUR 35 million | 7% |
| Most other obligations | EUR 15 million | 3% |
| Incorrect information to authorities | EUR 7.5 million | 1% |
Article 99(3) is drafted against Article 5 as a whole rather than against a list of points, so the new prohibition inherits the top tier without anything further being said about it. For an SME, Article 99(6) reverses the comparison and the cap is the lower of the two figures rather than the higher.
A ceiling is not a forecast. What the tier communicates is category: prohibitions are treated as conduct that should not occur rather than as risk to be managed to an acceptable level, and a risk register entry accepting residual exposure here reads badly.
05What does a defensible position look like in December?
A control, a test, and a written rationale.
The control is filtering at input and output, aimed specifically at the categories in points (ba) and (bb) rather than at general safety. The test is evidence that it was exercised against representative attempts, with a date. The rationale is a short document stating what the filtering is designed to prevent, what it is known not to catch, and what happens when a report arrives, which is the shape of evidence an audit accepts.
That third artefact is the one organisations skip and the one that distinguishes a system with a safety filter from an organisation that can explain its position.
06Frequently asked questions
What new prohibition was added to the EU AI Act in 2026?
A ban covering AI systems that generate or manipulate non-consensual intimate imagery and child sexual abuse material, inserted into Article 5(1) as points (ba) and (bb) by Regulation (EU) 2026/1744, with a compliance date of 2 December 2026.
Does it apply if none of our systems are high risk?
Yes. Prohibitions in Article 5 are not scoped by risk classification. They apply to the practice itself, so an organisation with no high-risk systems is still in scope if it operates a capability the prohibition covers.
What are the penalties for a prohibited practice?
The top tier. Article 99(3) applies to non-compliance with Article 5 as a whole, so the new points are covered: administrative fines of up to EUR 35 million or 7% of total worldwide annual turnover, whichever is higher. For an SME, Article 99(6) makes it the lower of the two figures rather than the higher.
What should an image generation product do before December?
Three things: implement input and output filtering aimed at the prohibited categories, write down what the filtering is designed to prevent and how it was tested, and define the escalation path for reports. The written account is what turns a technical control into evidence.
07References and method
- Regulation (EU) 2026/1744 of 8 July 2026, amending Regulation (EU) 2024/1689 as regards the simplification of the harmonised rules on artificial intelligence. Published in the Official Journal on 24 July 2026, in force from 27 July 2026. This is the instrument that inserts points (ba) and (bb) into Article 5(1) and sets their compliance date of 2 December 2026.
- Article 5 of Regulation (EU) 2024/1689 as amended, for the wording of points (ba) and (bb), including the extension to output that is a reasonably foreseeable and reproducible outcome of the system. Checked against the consolidated numbering on 17 September 2026.
- Regulation (EU) 2024/1689, Article 99, for the penalty tiers: EUR 35 million or 7% of worldwide annual turnover for Article 5, EUR 15 million or 3% for most other obligations, EUR 7.5 million or 1% for incorrect information, and the reduced basis for SMEs in Article 99(6).
- Regulation (EU) 2024/1689, Article 113 as replaced by Regulation (EU) 2026/1744, for the deferred high-risk dates of 2 December 2027 under Article 6(2) and Annex III, and 2 August 2028 under Article 6(1) and Annex I.