Don’t forget the new prohibited practices — the fine band here is the 7% one

From 2 December 2026 two more practices become outright prohibited under the EU AI Act — and prohibited is the top fine band: up to €35 million or 7% of worldwide annual turnover, whichever is higher, not the 3% band that governs most other breaches.

Who this is for

• Providers placing image, video or audio generation models on the EU market.
• Deployers who let staff, customers or end users run those models under their own control.
• AI governance and compliance leads who own the Article 5 prohibited-use list.

What’s happening

The Digital Omnibus amendments added two new entries to Article 5, the AI Act’s list of prohibited practices. From 2 December 2026 it is banned to place on the market, put into service or use an AI system that generates or manipulates non-consensual intimate imagery (NCII) of identifiable people, or one that generates or manipulates child sexual abuse material (CSAM) as defined in Directive 2011/93/EU. The rest of Article 5 has applied since 2 February 2025; these two carry a transitional period that closes on 2 December 2026. I wrote up the scope and the drafting in the December 2026 Article 5 addition if you want the detail.

Why it bites

Two reasons. First, the money. Breaching Article 5 sits in the highest penalty band under Article 99 — up to €35 million or 7% of total worldwide annual turnover, whichever is larger. That is the same band that has always covered prohibited AI, and it is roughly double the 3% ceiling that applies to most other AI Act obligations. This is not a documentation-gap fine.

Second, the trigger is wider than “we built it to do this”. For providers, the ban bites where generating such material is the intended purpose or where it is a reasonably foreseeable and reproducible outcome without significant technical modification and you have not put adequate safeguards in place. So a general-purpose image model with weak guardrails is squarely in scope. For deployers, it is narrower — you have to be using the system intentionally for that purpose, including by circumventing a provider’s safeguards — but “one of our people jailbroke it” is exactly the scenario the drafting anticipates.

The one thing to do before 2 December

Add both practices to your Article 5 register, then for every generative model you provide or deploy, write down the answer to one question: what stops it producing this output, and who checked? If the answer is “the vendor’s filter”, get that in writing and confirm it covers both NCII and CSAM. If the answer is silence, that is your gap — document the foreseeable-misuse assessment and the technical safeguard before the date, not after the first complaint. Keep it on the same risk register you use for the rest of your AI Act obligations; this is one line item, not a separate programme. If you are already mapping the AI Act against DORA and NIS2, fold it into that one convergence view.

My advisory work here is scoped and fixed-fee — you know the number before I start, and it is a fraction of the band we are talking about.

If you want a second pair of eyes on where your generative models sit against the new Article 5 entries before 2 December 2026, that is what I do — see how I work.