Does your video ad need an AI label? What the AI Act says

No, not every AI video ad needs a label. Since 2 August 2026, you only have to label a deepfake visibly: realistic content that could pass for real, such as AI people, a cloned voice or a product that looks better than it does in reality ^[1][2]. Your real product in an AI setting generally needs no label.

That follows from Article 50 of the AI Act and the guidelines the European Commission published on 20 July 2026. At the end of August, the Raad voor Reclame, the Belgian advertising council, turned them into guidance for Belgian practice, with concrete label texts ^[3][4]. Whether you outsource AI video production or make ads yourself: this is how you decide per ad.

When does your video ad need a label?

Label it if your ad realistically shows something that exists or could have existed, and the audience could wrongly take it for real ^[1]. A realistic AI person who doesn't exist also counts. Anything that defies the laws of nature, such as a talking animal or a dragon, falls outside it ^[2].

The Raad voor Reclame turns this into three questions: how much AI is in the ad, is the result realistic, and does that change how consumers understand the ad or the product ^[4]? AI in preparatory steps, such as the script or storyboard, in principle requires no disclosure ^[3].

Don't count on the exception for creative work. For evidently artistic or fictional work, a more discreet disclosure is enough ^[1], but the Commission interprets "evidently" strictly. It cites a realistic synthetic influencer testing a product as an example that doesn't qualify ^[2], and according to the Raad voor Reclame there is no specific exception for advertising ^[4].

The duty has not been postponed: the Digital Omnibus mainly pushed back the rules for high-risk AI ^[2]. You don't have to label deepfakes from before 2 August 2026 retroactively, although the Commission encourages it ^[2].

How do you label a video ad?

Put the label in the video itself, from the first frame. The law requires a clear disclosure at the latest at the time of first exposure ^[1], and according to the Commission, information hidden in menus or terms and conditions doesn't count ^[2]. Neither the AI Act nor the guidelines impose a fixed text, icon or minimum duration ^[4].

For online video, the Raad voor Reclame recommends an icon or label that is immediately visible and stays on screen for at least five seconds; for TV and connected TV, for the entire spot. The label is neutral, in the language of the ad and with a contrast of at least 4.5:1 ^[4].

Keep the label out of the zones the platform covers: in Reels, buttons and caption overlay the bottom 35% and the top 14% of the frame ^[5]. The Commission's EU icon is allowed, but not required ^[6].

Is the Meta, TikTok or Google label enough?

Only on that platform, and not always. The Commission accepts a platform label if it is clear and recognisable, but if you run the same video elsewhere, you label it there yourself ^[2].

So put the label in the video, and also switch on each platform's AI setting. How TikTok ads work beyond that, from formats to budget, is covered by Jonas in TikTok advertising in Belgium.

Who labels: you or your agency?

Whoever decides that AI is used, and how, does the labelling. An advertising agency that uses AI itself is the deployer; an advertiser that doesn't decide whether and how the agency uses AI is not ^[2]. If you ask for AI avatars or an AI voice yourself, you share in that decision.

The Commission asks the parties in a production chain to make arrangements that ensure the label reaches the audience ^[2]. So set out in your briefing or contract:

• which elements are made with AI: images, voice, people, product;
• who assesses whether it is a deepfake, and who decides on the label;
• the label text per language, its position and its duration;
• the consent for every real person, voice or likeness;
• who switches on the AI setting in Meta, TikTok and Google Ads.

What do you risk, and what does a label not solve?

The AI Act provides for fines of up to €15 million or 3% of worldwide annual turnover; for SMEs, the lower of the two applies ^[2]. In Belgium, the regulator has not yet been designated by law, so for now enforcement can only go through the courts, the FPS Economy told VRT NWS ^[10]. Meanwhile, the JEP, the sector's self-regulatory jury, oversees the advertising sector's AI guidelines ^[11].

A label is not a free pass. Misleading stays misleading, and anyone who imitates the image or voice of a real person needs prior consent ^[4]. A label doesn't change anything about copyright either ^[2]: that is the subject of can you use AI images and AI music commercially? If you work with avatars, read AI UGC and AI avatars in ads first.

Five questions before you publish

1. Does the ad realistically show something that exists or could have existed, which the audience could take for real? Then you label it.
2. Does your product show something it doesn't have or do in reality? Then you cut it.
3. Does a real person, voice or likeness appear in it? Then you need written consent.
4. Is the label on screen from the start, in the language of the ad and clear of the buttons?
5. Have you kept a record per ad of what was made with AI and why you did or didn't label? If you don't sign the Commission's code of practice, you must be able to demonstrate compliance in another way ^[12].

Want to see how that turns out on your own product? Request a free test ad: one video ad on your own brand, at no cost or obligation.

This article is not legal advice. The Commission itself calls its guidelines a first interpretation that it will revise ^[2], and the Belgian regulator has yet to be designated.