AI Act Article 50 fines: up to EUR 15M or 3%, and who enforces it

The European Commission's Transparency obligations under Article 50 of the AI Act page says fines can reach up to 15 million euros or 3% of total worldwide turnover for the preceding financial year. Enforcement is mainly by national market surveillance authorities, not by one EU body. This post reports what that page says; it is not legal advice.

Article 50 has applied since 2 August 2026, so a team that generates or publishes AI video for the EU market is already inside the period these figures describe.

Who enforces Article 50?

The FAQ says compliance "will mainly be enforced by national competent market surveillance authorities." The AI Office has a limited role. The European Data Protection Supervisor covers AI systems used by EU institutions, bodies and agencies.

The practical effect is that an answer to a regulator may come from a different country's authority than the one you expect, which is one reason the Commission's separate page on the Code of Practice stresses "across all Member States" certainty for signatories.

How large can the fines be?

The page states a ceiling of up to 15 million euros or 3% of total worldwide turnover for the preceding financial year, and adds that proportionality can be taken into account for small and medium-sized enterprises and small mid-cap companies. It does not say what any single case would cost, and it does not list how an authority sets the amount, so this post does not estimate one.

Does the AI Office supervise every video generator?

No. The FAQ limits the AI Office to the three cases in the table. A video tool that is a separate system built on someone else's model, and not part of a designated very large platform, falls to the national authorities by that description.

What can a video team keep to answer a question later?

The FAQ does not list records an authority will ask for, so nothing here is a required checklist. What Sume's docs do give you is a durable job: the Jobs and results page says generation endpoints create durable jobs, tells you to store the job id from the submit response, and shows completed jobs returning artifacts with a media.sume.com URL.

Keeping the job id, the artifact id and the URL you published next to each file is a cheap way to show which output was generated, when, and which version you delivered. The post on what to log per generated file goes through the fields.