AI labels for videos: how to implement labelling in practice

The legal framework from an expert view is easy: you need to label AI.

The legal framework comes from Article 50 of the AI Act. The regulation applies directly in every member state, so no separate national law is needed.

How exactly (or how loosely) this works in practice, several experts have already taken a close look. Dr. Thomas Schwenke describes the basis like this: the EU Commission "has published a code of conduct and official icons for AI-generated content" . In his detailed guide he makes clear who is affected: every company, every public authority and every freelancer who uses AI professionally. So all of us, with hardly any exceptions.

At its core, the regulation separates two roles:

• Providers of AI tools mark their output in a machine-readable way, for example through metadata or invisible watermarks.
• Companies that use AI label deepfakes visibly for their audience.

For those of us in the second group, things get concrete, as LTO points out: "For deployers, the labelling obligation is immediately visible and therefore directly challengeable" .

Does this now apply to everything that involved AI? No, it depends. The law firm SRD writes that the obligation "is decided not in the abstract, but by the specific case". ABD Partner names a key criterion: "Deepfakes only require labelling insofar as they are used in a professional context".

As you can see, there are rules, and there are opinions. The Commission's guidelines offer orientation, but Heidrich make it clear: "The guidelines remain legally non-binding". For your daily work, this means a lot is still moving, and over the coming years courts will look at many cases in more detail. Ignoring the labelling obligation can lead to fines and cease-and-desist letters. National supervisory authorities handle enforcement. You're call.

So when does a video need an AI label?

A video needs an AI label when three things come together:

1. you created or substantially changed the content with AI,
2. it shows real people, places, objects or events, and
3. it looks so realistic that someone could take it for real.

Two points matter most here. What counts is the effect on your audience, your intention plays no role. And the rule covers professional use, content created privately falls outside it.

Where the debate is still open

The open question is mainly about photorealistic people who do not exist. The Commission sees a possible deception here as well. Lawyer Tobias Voßberg reads the final guidelines more narrowly, since all their examples revolve around real people. His conclusion: "Which labelling practice is actually required, case law will still have to show".

Product images of mass-produced goods are just as open. On the podcast Rechtsbelehrung, Thomas Schwenke sums up the balancing act: "what matters is the degree of similarity, the message, the context of use and the expectations" of the audience. For borderline cases he gives a clear recommendation: "When in doubt, labelling is advisable".

What needs an AI label and what doesn't?

Enough theory, let's look at examples.

The rule of thumb: the more real an AI element looks and the more it relates to real people or things, the more likely it needs a label.

Label required

• Avatar of a real person: for example your CEO in a company announcement, or a member of your service team who guides viewers through a tutorial as an avatar.
• Realistic avatar without a real role model: as mentioned still legally disputed. Recommendation is to label this.
• Cloned voice: if the AI voice sounds like a real person and comes across as genuine, it needs a label. If the cloned voice runs in a video, the label also belongs visibly in the picture.
• Face swaps and altered statements: swapped faces, inserted scenes, or people whose words you change with AI.
• Realistic AI scenes with real places or specific products.

No label required

• Color correction, stabilization, upscaling: classic image optimization remains standard editing.
• Animation and fantasy: clearly animated, cartoon-like or physically impossible scenes.
• A generic AI narrator voice without a real role model should be fine.

What this means for your website and your e-learning platform

To make it even more tangible, here are a few typical situations from everyday work.

Product videos on your website:
Did you work with AI, and does the video show one specific item, such as a particular vehicle? Then it depends on how much you changed. If you only corrected light and color, it remains standard editing. If you generated the vehicle or changed it substantially, the video needs a label. For image videos covering an entire product line the situation is still open, and a label can give you certainty here.

Internal training videos:
The regulation generally makes no difference between internal and external videos. So the obligation applies here too, as soon as an avatar or a cloned voice appears. The notice can be leaner, for example in your learning platform or right before the video.

E-learning and tutorials:
If an avatar guides learners through longer courses, the label belongs at the start and reappears along the way, for example with every new chapter.

Existing videos:
Videos you created before the obligations took effect need no retroactive label. What counts is when you created them.

What should an AI label in a video look like?

We have talked a lot about when a label is needed. That leaves the question of what it should actually look like. Here are the key ground rules:

• Visible from the start: the label appears right when the video starts, easy to spot and free of overlapping elements.
• Repeated or always visible in longer videos: in longer content the label shows up again or is visible the whole time. If only individual sequences are AI-generated, a label for that section is enough. When in doubt, though, we recommend showing the label the entire time. That way you are on the safe side, even when someone joins halfway through.
• Accessible by design: this means strong contrast and good readability. And here comes the catch: screen readers should be able to read the label too. A label you render directly into the video stays invisible to screen readers. If you want to get this right, display the label through the player.

How we handle this: Simple AI labels with Ignite

As a video hosting and streaming platform, we offer you a player that lets you embed videos into your website, app or learning platform with ease. Since we have worked intensively on this topic, AI labelling is built right in, accessibility included.

Here is how it works: when you upload your video, you mark it as AI-generated or edited with AI. The player then shows the label automatically, and you decide where it appears. The label meets the rules described above while staying subtle and discreet. Your product video or e-learning course stays in the spotlight.

We deliberately use a label that stays visible throughout. AI rarely shows up in just one spot in a video, so this way every sequence is covered. For you, it takes a single click, and there is no need to render anything into the video afterwards.

You decide which video needs a label. You set the label, Ignite displays it reliably.

Here is an example video with a label in the top corner:

What else to keep in mind

A label creates transparency, but it does not replace further checks. Thomas Schwenke points this out in his guide as well: an AI label does not automatically make content lawful.

You cannot recreate a person with AI or clone their voice, put an AI label on it and call it a day. You need the person's consent for their face and voice. You also need to check copyright and trademark rights, follow the GDPR and make sure your advertising claims do not mislead.

The AI label covers none of that.

Conclusion:
if you work with AI, label it

The rules are clearer than they first seem. Realistic AI depictions of real people, voices or scenes need a label, standard editing and animation do not. For borderline cases, a simple rule of thumb from legal practice applies: label it.

And one point many people overlook: to also meet accessibility requirements, the label should come through the player rather than being rendered into the video.

P.S.: This guide is not legal advice. There are countless individual cases and distinctions, and our notes give you a rough orientation. For your specific situation, it is best to ask your own lawyer.