Labelling AI content in ads: what applies since August 2026

What applies since when

The transparency obligations in Article 50 of the EU AI Act, Regulation (EU) 2024/1689, have applied across the EU since 2 August 2026 under Article 113. They apply directly; no German implementing act is needed for that.

There is a transition period, but it is narrower than often claimed. It concerns only providers of generative AI systems placed on the market before 2 August 2026, and only the technical marking of outputs under Article 50(2). For those systems the marking duty applies from 2 December 2026. The period comes from the so-called AI Omnibus package, which entered into force in July 2026.

Advertisers get no grace period until December. The duty to label deepfakes visibly has applied since 2 August 2026. In the European Commission's view, content created before then does not have to be labelled retroactively, although it recommends doing so where feasible.

In the first version of this article we described the December date as a transition period for content already published. That was too broad, and we have corrected it.

Provider, deployer and which role you play

Article 50 distinguishes providers from deployers. Providers are those who develop an AI system and supply it under their own name, that is the makers of the tools. Deployers are those who use an AI system under their own authority in a professional capacity. Anyone using an image tool to generate motifs for their own company is a deployer.

If an agency builds the motifs, it is the agency that uses the tool in the first place. Whether the advertising company also counts as a deployer is not clearly answered by the text of the law. In practice that matters less than one might think: platform rules bind the owner of the ad account, and under competition law the advertising company is also liable for breaches by those it engages. So settle in the contract who labels and who checks it.

Providers have to ensure that their systems' outputs are marked in a machine-readable way as AI-generated, for example through watermarks or metadata. Deployers have to ensure that people can see or hear it. One does not replace the other: the Commission makes clear that deployers cannot rely on the provider's technical marking.

What counts as a deepfake under the AI Act

The Regulation defines a deepfake in Article 3(60) as AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful.

Three points are decisive for advertising. First, it is not only about people but expressly also about objects and places, so products and settings too. Second, on the Commission's reading it is enough that what is shown could plausibly exist. A realistic AI person who never existed can therefore be a deepfake. Third, what counts is the impression on the viewer, not the intention of whoever built the motif.

Conversely, not every AI motif is a deepfake. A recognisably drawn illustration, a comic style or a scene that obviously cannot be real creates no false impression of authenticity.

Who must do what: the overview

The table below maps the content types typical in advertising to the duties under Article 50. It is a simplification; in the individual case the actual impression the motif makes decides.

The chatbot deserves a second look. The duty in paragraph 1 falls on the provider. But anyone who integrates a language model and runs the bot on their website under their own name may count as a provider themselves. A clear notice at first contact, such as "You are chatting with an AI assistant", costs nothing and settles the question in practice.

Product image, avatar, voice: the typical advertising cases

The position is clearest for content meant to look convincingly real. An AI avatar that looks like a satisfied customer and appears as a genuine testimonial is exactly the case the rule was made for. There is no way around a clear label here, and an invented customer voice is a competition-law problem in its own right.

The same applies to cloned or synthetic voices that sound like a real speaker, and all the more to any depiction of a real, well-known person. An AI version of a celebrity in an ad is not only subject to labelling, it also raises personality-rights questions that a label does not solve.

The product image is harder. According to summaries of the Commission's guidelines by several law firms, the rough rule is: a real product in front of an AI background is usually unproblematic as long as it does not mislead about the product. An AI image that shows the product differently or better than it is, by contrast, is treated as a deepfake. That is misleading advertising even without the AI Act.

The exception for evidently artistic, creative, satirical or fictional works rarely helps advertising. It does not exempt entirely anyway, it only permits a notice that does not hamper the work. And for content that is both commercial and creative, on the Commission's reading the commercial character generally prevails.

What does not fall under it

Pure advertising copy is not subject to labelling under Article 50. An AI-written subject line, AI-phrased ad copy, a reworked product description: that is advertising, not information to the public on matters of public interest. It is less clear-cut for guide articles on health, finance or the environment that aim to inform and sell on the side. That is worth asking your lawyer.

And even there an exception applies when a human reviews the text for substance and someone holds editorial responsibility. According to the Commission a spell check is not enough for that; a genuine review of the content is required.

What stays internal also stays out of it. An AI-summarised meeting note, a draft, an internal mail trigger no duty, because nobody outside the company is exposed to them. And ordinary image editing such as cropping, colour correction or removing a speck of dust does not turn a genuine photo into a deepfake.

Checklist for advertisers

We use this sequence ourselves before a motif with AI content goes live. It does not replace a legal review, but it sorts the cases.

1. Does the motif contain image, video or audio that was wholly or substantially generated or altered with AI? If not, Article 50 does not apply to the motif.
2. Does it show people, voices, products, places or events that exist or could exist, and realistically so? If it is obviously artificial, there is no duty under Article 50.
3. Would an ordinary viewer take it to be real? If so, it is a deepfake and has to be labelled visibly.
4. Does an AI person appear as a genuine customer, expert or employee? Then a label is not enough, the motif is also risky under competition law. Rework it.
5. Does the motif show the product differently or better than it is? Then do not label it, change it.
6. Put the label in the motif itself, and additionally use the platform's disclosure setting where one exists.
7. Leave provenance data in the file and document which tool the motif was made with and who approved it.

How to label in practice

The Regulation prescribes no wording. It requires the information to be clear and distinguishable, given at the latest at first exposure and accessible. According to the Commission the notice has to be perceivable by people without technical tools. A note in the terms and conditions or only in the metadata is therefore not enough.

Wordings that work are short and unambiguous:

• "AI-generated" or "Made with AI" as a discreet but clearly legible note in the image,
• "AI avatar" or "This person is AI-generated" for synthetic people, in video from the first second,
• "Voice AI-generated" as an on-screen note or spoken notice for synthetic voices, spoken at the start for audio-only formats,
• "Scene altered with AI" where genuine footage has been substantially manipulated.

On placement: the notice belongs in the motif, where it is noticed at first glance, not in a corner that the platform layout crops. In video it should be visible from the start, because many users scroll on after two seconds. The EU has published optional icons for this that you may use but do not have to. The notice should be in the language of the ad.

How this works in Meta and Google

Meta labels ads with "AI info" when they were created or significantly edited with Meta's own generative tools, or when Meta detects that a third-party AI tool was involved. The notice sits in the "About this ad" screen behind the three-dot menu, sometimes also next to "Sponsored". Minor changes such as resizing or colour correction do not trigger it.

Since July 2026 Google has expressly allowed text and visual labels directly in image and video ads and is gradually rolling out its own AI label setting in Google Ads, Display & Video 360, Campaign Manager 360, Merchant Center and Ads Editor. Assets from Google's own AI tools may be labelled automatically. Google itself points out that the setting does not guarantee compliance with specific regulations.

There is also a route that gets overlooked: many image tools write provenance data into the file. Platforms can read these signals and apply a label of their own. So anyone assuming a file is inconspicuous because it looks fine is often mistaken. Stripping provenance data to prevent exactly that is, in our view, the wrong approach.

Platform rules and the law: both apply

Platform rules and the AI Act run side by side, and neither replaces the other. The differences in brief:

• The law asks whether content looks deceptively real. The platforms often only ask whether AI was involved, and so label motifs the law does not require to be labelled.
• A platform label behind the three-dot menu is, in our assessment, not a notice that is clearly perceivable at first exposure. For a deepfake you should not rely on it.
• The platform enforces its rules itself, with rejected ads and, for repeat cases, account restrictions. The law is enforced by authorities and, through competition law, by competitors.
• Ads about elections and political issues are subject to separate, stricter disclosure rules on both platforms. According to Meta, it no longer allows such ads in the EU at all.

The first practical risk is therefore rejected ads, restricted campaigns and, in the worst case, a suspended ad account. That hits faster than any regulator.

Does a label cost performance?

That is the question people really ask, and the worry is understandable: a note saying "made with AI" might cost trust.

In our own tests we saw no difference beyond the usual noise between two motifs. That is not a study, it is an observation from a limited number of campaigns, and we would not sell it as a promise.

What does measurably cost performance is a rejected ad. And the likelihood of that rises when a platform works out for itself what you did not declare.

What happens if you skip it

Article 99(4) of the Regulation provides for fines of up to 15 million euros or up to three per cent of worldwide annual turnover for breaches of Article 50, whichever is higher. For small and medium-sized enterprises, paragraph 6 applies the lower of the two. These are ceilings; the amount in a given case depends on gravity, duration and circumstances.

In Germany, market surveillance is in principle the job of the Federal Network Agency (Bundesnetzagentur) under the AI market surveillance and innovation act, which entered into force on 29 July 2026. For certain areas such as media and financial supervision, other bodies remain responsible. How active supervision will be with advertising motifs cannot yet be said in September 2026.

Something else is more realistic anyway: a competitor's lawyer writing to you. An AI testimonial presented as a real customer, or a product image that promises more than the product delivers, can be misleading advertising under section 5 UWG, the German Unfair Competition Act. And withholding material information is unfair under section 5a UWG, towards consumers and other market participants alike. Whether the AI origin of a motif is such material information in a given case has not been conclusively settled.

The same pattern as with cold email, where the problem is not the regulator but the warning notice. How that plays out there we described in our article on section 7 UWG.

What this means for AI creatives

The labelling duty is no reason to give up on AI motifs. It only shifts which motifs are worth it. Stylised scenes, product visualisations without embellished features and avatars clearly recognisable as AI remain unproblematic. The risky ones are motifs whose effect relies precisely on being taken for real.

What AI can and cannot do in ad production we described in the article on AI creatives for ads. How we test motifs for Meta and Google is on the performance creatives page.

What we do in practice

Every motif that was made with AI is recorded as such by us before it enters an ad account. We set the declaration in the account where the platform provides for it, we do not strip provenance data from files, and we do not build avatars that appear as genuine customer voices.

That is not caution on principle. A suspended ad account costs more than any motif can bring in, and in our tests a label has never cost an ad its performance.

The questions for your lawyer

First: which of our ad formats are deepfakes within the meaning of the Regulation and which are not? Second: how and where do we label so the notice is perceivable at first exposure? Third: who is the deployer and who is liable when an agency creates the motifs? Fourth: are our guide articles information of public interest, and is our editorial review enough for the exception?

Four questions, one appointment. The same recommendation as with the UWG, and for the same reason: the assessment belongs to you, and we work to it. If you then need motifs that are properly labelled and still sell, get in touch.

Frequently asked questions

Sources

1. Regulation (EU) 2024/1689 (AI Act), in particular Art. 3(60), Art. 50, Art. 99 and Art. 113 (EUR-Lex)
2. European Commission: Transparency obligations under Article 50 of the AI Act (FAQ)
3. Meta: How AI-generated images in ads are identified and labelled
4. Google Ads: Updates to AI labeling requirements (July 2026)
5. German Federal Ministry for Digital Transformation: Act implementing the AI Regulation
6. Section 5a UWG, misleading by omission (gesetze-im-internet.de)