Your content agent produced the ad image in thirty seconds, and the copy was ready even sooner. The ad carries the word “Anzeige” (“ad”) in the top left corner, and everything feels settled. Since 2 August 2026, that single word is no longer enough in some cases. Alongside the question “Is this recognizable as advertising?”, there is now a second, separate question: “Is this recognizable as coming from an AI?”

This article is an evergreen topic, not a product launch. It explains why two labeling obligations exist side by side, what each one requires, and where an approval step makes sense. For this, we read the following: the reproduction of Art. 50 and Art. 3 of the AI Act on artificialintelligenceact.eu, the analysis by the law firm Cooley dated 3 August 2026, the Garrigues law firm's post on the publication of the Digital Omnibus, and the wording of Section 5a(4) UWG (German Unfair Competition Act) on dejure.org. We were unable to retrieve the Official Journal text on EUR-Lex. The official legal text therefore remains authoritative, and this article is not a substitute for legal advice.

Where in your marketing is material with an AI component already being created today?

In the free strategy call (30 minutes), we go through together which content your agents should produce and at which point a human gives approval. Free and non-binding.

Book a free strategy call (30 minutes, amazing-agents.com/booking)

Obligation number one: advertising must be recognizable as advertising

Many marketing managers know the older of the two obligations. Section 5a(4) UWG provides that anyone who fails to make the commercial purpose of a commercial practice clear acts unfairly. According to the wording of the law, this applies where the purpose does not follow directly from the circumstances and where the failure to disclose it is likely to cause a person to take a transactional decision they would not otherwise have taken. The provision, in the version in force since 28 May 2022, additionally contains a rule for acting on behalf of third-party businesses: there is no commercial purpose if neither payment nor a similar consideration has been paid or promised for it. However, receipt of a consideration is presumed, and the acting person must credibly demonstrate that there was none.

What matters for our topic: the provision does not ask who created the content. Whether a human, an agency or an agent wrote the text plays no role in whether it is recognizable as advertising. We have already described this separation requirement in a dedicated article on ads in AI chat. The core is enough here: advertising must be recognizable as advertising.

Obligation number two: disclose AI origin (Art. 50 AI Act)

Art. 50 of the AI Act is the newer obligation. It regulates transparency in four areas and addresses two different roles in doing so. Providers of AI systems must ensure that people are told when they are interacting directly with an AI (paragraph 1), unless this is obvious from the context to a reasonably attentive person. Providers of systems that generate synthetic audio, image, video or text content must mark the outputs in a machine-readable format and make them recognizable as artificially generated, as far as technically feasible (paragraph 2). Deployers (in the law: “deployers”) of systems for emotion recognition or biometric categorization must inform the persons concerned (paragraph 3).

For marketing, paragraph 4 is the most important, because it is aimed at deployers, that is, at companies that use AI systems. It requires two things. First, deployers must disclose, in the case of so-called deepfakes, that the content has been artificially generated or manipulated. Second, they must disclose AI-generated or manipulated text that is published in order to inform the public on matters of public interest. There is an exception to this text obligation: it does not apply if the content has been reviewed by a human or is subject to editorial control and a natural or legal person holds editorial responsibility.

Paragraph 5 governs timing and form. The information must be provided at the latest at the time of the first interaction or first exposure to the content, in a clear and distinguishable manner, and must meet the applicable accessibility requirements.

The separation requirement asks: Is it recognizable as advertising? Art. 50 asks: Is it recognizable as AI? These are two different tests with two different legal sources.

What a “deepfake” is under the law, and what it is not

The word conjures images of faked videos of politicians. The definition in Art. 3(60) is broader, and that is exactly what makes it relevant for marketing. There, a deepfake is 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.

Two features carry the definition: the content imitates something that really exists, and it could pass as authentic. A recognizably drawn mascot is unlikely to meet this. A photorealistic image of an “employee” who does not exist, or of a production hall that does not exist in that form, raises the question. Whether a specific marketing image falls under it is a matter of interpretation that we cannot answer for you here. What we can tell you is what it turns on: Does the image imitate something real, and would a viewer take it to be genuine?

For obviously artistic, creative, satirical or fictional works, paragraph 4 provides a reduced obligation: it is sufficient to disclose the existence of such content in an appropriate manner that does not hamper the display or enjoyment of the work.

Availability and deadlines: what applies since when

There is no product availability here, only dates of application. According to Cooley's analysis dated 3 August 2026, the transparency obligations of Art. 50 have applied since 2 August 2026 to all systems in scope, regardless of when they were placed on the market. According to that analysis, content created and published before this date does not have to be labeled retroactively. According to Cooley, the EU Commission adopted guidelines on the obligations on 20 July 2026, and the AI Office has published a voluntary code of practice on the transparency of AI-generated content. For violations, Cooley cites fines of up to 15 million euros or 3 percent of worldwide annual turnover, whichever is higher.

One exception concerns only the technical marking under paragraph 2, that is, the obligation of providers of generative systems. For systems that were already on the market before 2 August 2026, a transitional period applies until 2 December 2026. Cooley, Garrigues and the timeline on artificialintelligenceact.eu consistently report this. Garrigues also describes that the Omnibus Regulation was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026. As mentioned, we have not read the Official Journal text itself; we are reproducing the information from the specialist sources.

Important for your planning: this transitional period concerns the providers of the systems, not your obligation as a deployer under paragraph 4. So you should not wait for your tool to deliver the marking automatically at some point. Whether and how it does so is something to ask your provider.

Two obligations, two legal sources: why “Anzeige” is not enough

That both obligations apply side by side is intended. Paragraph 6 of Art. 50 makes clear that the provision leaves other transparency obligations under Union law or national law unaffected. German advertising law is one example. The consequence can be shown with a simple grid:

The strength of this grid is that it splits the question in two. Anyone who writes only “Anzeige” on it has answered the first question and not asked the second. Anyone who writes only “AI-generated” may not have made the advertising recognizable. The two notices serve different purposes.

A thought experiment: five pieces from one marketing week

The following example is entirely fictitious and only shows how the two questions sit side by side in practice. A mid-sized company has its agents draft five things in one week: a social media ad with a photorealistic image of a production hall, a LinkedIn post with an expert opinion, a short explainer video with an AI voice, a product description for the online shop and a chat assistant for the website.

For the ad, both questions arise: it must be recognizable as an ad, and for the hall image you need to clarify whether it imitates something real and could be taken for genuine. For the LinkedIn post, what matters is whether it is advertising; a text on a topic of public interest would be a case for the question of editorial review and responsibility. For the explainer video, it is open whether the AI voice imitates a real person. The product description is advertising and is therefore more a matter for the separation requirement than for the deepfake rule. For the chat assistant, the issue is the notice that an AI is answering here, and whether your tool already includes it.

The point of the example: five pieces lead to five different combinations of questions. That is exactly why a fixed list helps more than a blanket notice on every piece. Which combination leads to which result in your case, please clarify with legal advice if in doubt.

A checklist with two questions before every publication.

In the free strategy call (30 minutes), we design together the approval step that checks both questions at your company before anything goes live. Free and non-binding.

Book a free strategy call (30 minutes, amazing-agents.com/booking)

Which agent: the research team and the content chain

On amazing-agents.com, we assign this topic to the research team. Nadja is the research lead there, who, according to the website, delivers research briefings with topic and data basis. Daniel is the analytics specialist who collects performance data across all channels and builds reports. The honest limit matters here: both deliver material, research and numbers. A legal assessment is not a task that the website lists for any of our agents, and we do not claim it here. Legal advice belongs with your legal counsel.

Implementation in the material itself lies in the chain described on the website. Lena writes drafts; according to the website, every draft remains a draft until you approve it. Sophia designs visuals, Clara produces videos. Magdalena, as QA specialist, checks brand fit, tone and formal correctness; the website states that no post leaves the team without her review and no visual goes live without a brand check. The brand check is a quality check. That it also covers AI labeling under Art. 50 is not stated on the website. We therefore suggest adding this second question as a separate item in your approval checklist. That is a suggestion for how you proceed and not an existing feature.

On approval, we again reproduce what the website says: every agent runs within a defined framework with approval steps for critical actions, and emails, for example, are created as drafts rather than sent directly.

With and without an approval step: a comparison

Without an approval step, the content agent publishes the material directly. Whether a label is included depends on chance or on the tool's default setting. Nobody has asked specifically whether the image imitates something real, whether the ad is recognizable as an ad, and whether the notice about the AI origin is placed prominently enough. If someone asks later, there is no point at which it can be traced who decided this, and when.

With an approval step, every piece goes live only after a fixed review. A human sees the draft and answers two questions: Is this recognizable as advertising? Where necessary, is it recognizable that it comes from an AI? If an answer is “no”, the piece goes back. An approval step does not automatically make material lawful. It ensures that the question is asked and the answer is recorded.

What you should prepare now

The basic obligations already apply, so this is not a waiting period but a work assignment. Five steps that are possible without a new tool:

  1. Create an inventory: List which channels and formats contain material with an AI component: images, videos, texts, chat assistants on the website.
  2. Ask the reality question: Mark images and videos that imitate real people, places, products or events and could look genuine. Only there does the deepfake question arise.
  3. Define two notices: Define what “Anzeige” and what “AI-generated” each look like and where they appear, so that they are clearly recognizable and do not disappear in the small print.
  4. Extend the approval checklist: Add both questions as fixed items before publication, with the name of the person who approves.
  5. Question your providers: Ask every tool provider whether and how it marks AI-generated outputs in a machine-readable way and whether a notice is included for chat assistants. Get the answer in writing.

If you are unsure about specific cases, please seek legal advice. It is especially worth checking back for images that imitate real people or places, and for texts on topics of public interest.

What we do not know and do not claim

The limits of this article: we did not read the legal text in the Official Journal, but through the reproduction on artificialintelligenceact.eu and through the analyses by Cooley and Garrigues. We report the statements on deadlines and on the Omnibus as these sources report them. We do not assess any individual case, and we do not claim that a particular advertising item must or must not be labeled. We also say nothing about the amount of fines in an individual case; the upper limit cited is the law firm's statement.

What certainly applies is the basic idea: two questions, two legal sources, one process behind them. Anyone who asks them at every publication is closer to a kind of marketing that uses agents and still remains traceable. The difference that ultimately matters is the one between an agent that simply publishes and an agent with limits and an approval step.

An agent with limits and an approval step instead of “it'll be fine”.

In the free strategy call (30 minutes), we determine together which content your agents may draft and which needs your approval before publication. Free and non-binding.

Book a free strategy call (30 minutes, amazing-agents.com/booking)

Sources