Labelling AI content: what the EU AI Act asks from August 2026

What the law says

Article 50 of the EU AI Act (Regulation (EU) 2024/1689) covers transparency. Paragraph 4 says that whoever uses an AI system to generate or manipulate images, audio or video that amount to a “deep fake” has to disclose that the content was artificially generated or manipulated. Paragraph 5 says how: clearly and distinguishably, at the latest at the moment someone first sees or hears it. These duties apply from 2 August 2026.

A 2026 amendment gives the makers of AI systems until 2 December 2026 to add machine-readable marks to their output. That part is for the tool makers. The visible label is for whoever publishes the content: a business posting on Instagram or TikTok included.

What counts as a deep fake

The European Commission explained the rule in guidelines published on 5 November 2026. Content that shows people, objects, places or events that could plausibly exist counts, and the Commission names depictions of professional and consumer services. An AI video of a barber at work, a yoga class or a café counter is in scope if a viewer could take it for real.

• What the maker intended does not matter. What a viewer could believe does.
• Content made mainly to sell does not get the lighter treatment the law gives to evidently artistic, satirical or fictional work.
• An AI background around a real photo of your real product is fine, as long as the result does not mislead.

Where the label goes

The guidelines rule out two shortcuts. A hidden mark alone, such as metadata or an invisible watermark, is not enough. And where viewers may join a video after it starts, as they do in every feed, a label shown only in the first seconds is not enough either.

The Code of Practice that the Commission judged adequate in July 2026 recommends, for video, an icon at the start, after breaks and at regular intervals, placed in a corner such as the top right. The EU’s own icon is optional. The German advertising association ZAW reads the rules the same way for advertisers: label everything that could be taken as real; “KI” or “AI” is enough; and the label travels with every cut of the content, including stories and short versions.

What it costs to skip it

According to the ZAW, the state media authorities supervise the rule in Germany, fines can reach €15 million or 3 % of worldwide turnover, and competitors can send warning letters under the unfair competition law (§ 3a UWG). For a small business the warning letter is the realistic risk.

The label has a price too. In a study by NYU Stern and Emory, fully AI-made ads drew more clicks than human-made ones, but labelling them as AI cut their click-through by 31.5 %. That is a reason to use AI where it helps the story, not a reason to hide it.

A checklist for your next post

1. Does the picture or clip show a person, place, product scene or event that could pass for real? Then it needs a label.
2. Put “AI” or “KI” visibly in a corner for the whole clip, not only in the first second.
3. Keep the label in every version: the Reel, the story, the cut-down.
4. Switch on the platform’s own AI label as well.
5. Your own photos and footage, only cut and colour-graded, are not AI-made. Once AI changes what is shown, label it.

How Ansamble does it

Every picture or clip Ansamble generates carries a visible label, the caption says the post contains AI-made content, the platform’s AI flag is set when the post goes out, and the file itself carries the IPTC “digital source type” metadata. A post made only from your own photos and footage carries none of that. And nothing is published until you say yes.