Two duties, two addressees
Most of the confusion comes from the fact that the relevant article of the AI act prescribes two separate things, for two separate parties.
One is the machine-readable mark, and that is the provider's job: the company that operates the generating system. It has to make sure the output can be recognised by machine as artificially produced. The other is the disclosure to the public, and that one belongs to the publisher — that is, to you. This one is narrower, though: it does not cover every generated piece of content, only the kind capable of deceiving, which the act defines as a deepfake.
The deadlines that have already passed
- The relevant article has applied since 2 August 2026. It is not something to prepare for, it is something to comply with.
- There is a grace period until 2 December 2026 — but only for the provider-side, machine-readable mark, and only for systems placed on the market before 2 August.
- There is no retroactive labelling. Content made earlier does not have to be marked afterwards.
- The grace period does not cover the publisher's disclosure. That one is live already.
What the code of practice says about video
The voluntary code of practice drawn up alongside the regulation was finalised in the summer of 2026 and has somewhere around a hundred and eighty to two hundred signatories. On video it is more concrete than the text of the act itself, and it is worth knowing what it expects.
The label should sit in the top right corner of the frame, at the start of the video and at regular intervals — at minimum after every ad break. There is a stated reason for that: what spreads onward from videos is clips and screenshots, and a single introductory label does not show up on those. The other important point: for a visual deepfake an audio-only disclosure is not enough — if the picture is what deceives, the signal has to be visible too.
The platforms
- The largest of the video sharing platforms is completely clear about it: the label appears under the player, as an overlay on short videos, and since May 2026 the platform detects generated content by itself as well. It accepts a specified version of the provenance standard.
- At that same platform there is a useful exception: cloning your own voice for dubbing does not require a label.
- The short-video platform has also been embedding a hidden mark since November 2025 — but only in content made with its own AI tools, not in every upload. By its own account it has labelled 1.3 billion videos so far.
- A number that shows how far invisible watermarks have spread: one large vendor says its system has marked ten billion pieces of content to date.
What we could not confirm
For this post we wanted to check several platforms' policies from primary sources, and at two large platforms that did not work — one help centre does not return its content to an automated request, the other's returns an error. At a third, the help page served a version from March 2023, which contradicts the company's own announcements from 2025.
So we state nothing here about those platforms. That is deliberate, not sloppy: most of the "platform rule summaries" circulating in this field are built on secondary blogs and are often years out of date. If you are uploading to a given platform, open its own help centre and read it that day.
The penalty, and who sets it
The EU regulation permits fines of up to fifteen million euros or three percent of worldwide annual turnover — whichever is higher. That is the general frame.
Enforcement, however, is a member state matter, so the actual upper limit and the authority that acts differ from country to country. And not everyone is at the same point: in some states the penalty regime is finished, in others the obligation is live but nobody can yet impose a national fine. Open the one that applies to you:
Hungary — the fine band exists, in forint
Implementation went through a separate act and a government decree, and the decree gives the upper limit in forint as well: the band covering the labelling obligation is 5.7 billion forint. Market surveillance sits with the minister responsible for enterprise development — not with the data protection authority.
That figure is obviously not designed for small companies, and under the regulation the ceiling for an SME is the lower of the two values. The obligation itself, though, applies to them in exactly the same way.
Spain — the obligation is live, the national fine is not
The Spanish implementing bill was still in the amendment stage in Congress in September 2026: the council of ministers adopted it and sent it to parliament on 26 May 2026, and the deadline for amendments has been extended four times since. So there is no national penalty regime in force.
That has a very practical consequence: in Spanish law an offence can only be created by a norm of statutory rank, so not a single euro of fine can be imposed in Spain today for unlabelled AI content under the AI Act. What is live: where the content depicts a real person, the data protection authority can act under the GDPR — and it already has in such a case.
Under the bill, breaching the labelling obligation will be a serious infringement carrying up to 7.5 million euros or 1% of worldwide turnover, supervised by the AI agency AESIA. Worth knowing, because the widely circulated claim that "unlabelled AI content costs 35 million euros in Spain" is wrong twice over against the official text: wrong band, and the law has not been passed.
Article 50 itself, on the other hand, has applied directly in Spain too since 2 August 2026, with the grace period for machine-marking systems already on the market running to 2 December 2026. The obligation is there — what is missing is the national sanction behind it.
What does NOT belong here
Two common misunderstandings are worth clearing up, because both cause unnecessary work.
The first: the EU political advertising regulation — which has applied since the autumn of 2025 — does not prescribe AI content labelling. Artificial intelligence appears in that law at the targeting stage, not at the content. Anyone making political ads has other duties, but the AI label is not among them.
The second: the publisher's disclosure does not cover every piece of generated content. A generated background meant as an illustration, a piece of scoring music or an avatar training video does not become label-bound because AI made it — it becomes label-bound if it deceives. In practice, though, platform rules are broader than the law, so they usually make the decision for you.
What to do
- Do not strip the mark out of the file. Most generators write it into the metadata; editors and compressors often throw it away. Check it in the finished file.
- Where the platform asks, tick the box. That takes five seconds and removes most of the risk.
- If the picture is what deceives, the signal has to be visible. An audio statement on its own is not enough.
- Keep a record. Which material was made with AI, which part of it, and when. A spreadsheet is enough, and if a question ever comes it is your answer.
Common misconceptions
- "This becomes mandatory from 2027." It has been in force since August 2026. The December grace period covers only the provider-side, machine-readable mark.
- "Every video made with AI has to be labelled." Legally, the ones capable of deceiving. Platforms, however, ask for more than that.
- "I have to go back and label my old videos." No: there is no retroactive labelling.
- "Political ads need a separate AI label." The EU political advertising regulation prescribes nothing of the sort.
What to take away
The duty is already live, with two addressees: the machine-readable mark belongs to the provider, the disclosure to the public belongs to you — and the latter applies to deceptive content. In practice the platform rule is stricter than the law, so that is where your day-to-day task is decided. On the technical side of labelling — what metadata and an invisible watermark can survive — Watermarks and Provenance is the post to read, on labelling written content AI-Written Text covers the question, and on music licences AI Music and Commercial Use goes through the terms.