In short
- If you publish AI images or AI videos that look real, you have to disclose that they are AI-generated. Article 50(4) of the EU AI Act sets this out.
- The duty lies with whoever publishes the content, not with the tool maker. It also applies to companies outside the EU if the content is used in the EU.
- The EU Code of Practice of 10 June 2026 describes the label: a symbol with “AI”, visible immediately, shown at the start of a video and repeated at intervals.
- Instagram, TikTok and YouTube also require their own disclosure when you upload realistic AI content.
What the AI Act requires
Article 50(4) of the AI Act (Regulation (EU) 2024/1689) covers so-called deepfakes: AI-generated or manipulated images, audio or video that resemble real persons, objects, places or events and would falsely appear authentic (Article 3(60)). Anyone using such content professionally has to disclose that it was artificially generated or manipulated.
The disclosure has to be clear and appear at the latest at the first exposure (Article 50(5)). For evidently artistic, creative, satirical or fictional works, a disclosure that does not hamper the display is enough, for example in the credits or as an unobtrusive symbol.
The duty applies since 2 August 2026. Breaches of the transparency obligations can be fined up to EUR 15 million or 3% of worldwide annual turnover.
What does that mean for advertising?
An ad with a realistic AI person, or a scene that looks like a real shoot, falls under the duty in my reading. An obviously stylised 3D spot probably does not, because nobody takes it for real footage. A product image showing your real product in an AI-generated setting usually looks real, so I label it. When in doubt, I label.
What the label should look like
On 10 June 2026 the European Commission published a Code of Practice on labelling AI-generated content, followed by guidelines on Article 50 on 20 July 2026. The code is voluntary, but following it lets you demonstrate compliance. For deployers it sets out:
- a common EU symbol with the acronym “AI”, or an equivalent label,
- visible immediately, without anyone having to click or expand,
- visible long enough to be noticed, and embedded in the content where possible,
- for video: at the start and repeated at intervals,
- for audio-only content: a spoken notice at the start.
Law firm Lewis Silkin gives a good overview of the code and the guidelines.
What Instagram, TikTok and YouTube require
| Platform | What must be labelled | How |
|---|---|---|
| Instagram and Facebook | Photorealistic video and realistic-sounding audio created or altered with AI | Set the AI disclosure when posting; Meta then shows “AI info”. Meta also labels some content with AI metadata on its own. |
| TikTok | Realistic AI images, video and audio | Turn on the AI-generated content setting before posting (TikTok help). TikTok also reads Content Credentials metadata. |
| YouTube | Realistic AI content, such as scenes that never happened or real people saying things they never said | Answer Yes to the AI use question in YouTube Studio when uploading (YouTube help). |
TikTok and YouTube state that disclosure alone does not limit reach. Whether a platform label already satisfies the AI Act is an open question, so I also put the notice directly into the video.
How I handle it in projects
- In the video: a small “AI” mark from the first second, placed in a corner where it does not cover subtitles.
- When posting: the platform’s AI toggle, plus a sentence in the caption when realistic people appear.
- Metadata: the provenance data from the AI tools stays in the file wherever the export allows it.
- On websites: a small “AI” mark on generated images and an explanation in the footer, as on this site.
- Facts: prices, dimensions and product details never come from a model, they come from the client.
Common mistakes
- The notice is only a hashtag in a collapsed caption. That is not immediately visible.
- The notice only appears in the end card of an ad that is meant to look real.
- The label is only in the profile bio, not on the individual piece of content.
- A video is re-exported so often that its provenance data is gone.
Note: as of 20 October 2026. This is a view from my own work, not legal advice. For your specific case, ask a lawyer.