AI labeling requirement from Aug 2: Myths and facts for marketing teams

We hear this question in almost every client conversation now. Do we have to label our entire website because texts were revised with AI?

The short answer is no. Promotional website copy does not inform the public about a matter of public interest. And even if it did, the editorial exception applies once a named person has reviewed the content and taken responsibility for it.

That is the most important point. Everything else is clarification — and that’s what’s missing right now.

No, not every use of AI triggers a labeling requirement

For weeks the topic has been traded as an anxiety asset. Countdown graphics, penalty numbers, ready-made compliance packages. The uncertainty is bigger than the obligation, and some of it is manufactured.

I’m not the first to notice. Attorney Dr. Carsten Ulbricht wrote in late July that “contrary to numerous misleading LinkedIn posts” not every AI use triggers a labeling obligation. If a specialist lawyer publicly calls out the panic, we don’t need another argument.

Role matters before content does

Article 50 of the AI Act separates two roles. Providers develop AI systems — model houses and tool makers. Operators use an AI system in their own professional responsibility. If you merely use AI, you are an operator.

So you don’t have to develop anything to fall under the rules. It’s enough to use a third‑party tool professionally. One exception is often overlooked: if you integrate or commission a chatbot on your own website, you can move into the provider role. That’s the only situation where marketing goes beyond the operator role.

For operators there are exactly two cases that matter. First: image, audio, or video that is a deepfake. Second: text published to inform the public about a matter of public interest. Everything else does not trigger an obligation.

Case one: Deepfakes — the three check questions

Under the regulation, a deepfake is AI‑generated or AI‑manipulated content that resembles real people, objects, places, institutions, or events and could be mistaken for genuine material. The decisive factor is not intent to deceive, but the potential to deceive. That creates three practical questions:

1. Was the content generated or manipulated by AI?
2. Does it resemble real people, objects, places, institutions, or events?
3. Could a viewer perceive it as a real recording?

If the answer is yes to all three, label it. One no usually means no obligation.

In practice, more content is excluded than you might think. You must label a photorealistic lifestyle photo of a fictional family on a real beach, a KI avatar giving a product review, and a convincingly real voiceover laid over authentic footage. You do not need to label a real product photo with color corrections and removed shadows. Clearly artificial graphics are also out, because no one mistakes a comic illustration for a photo.

A useful rule of thumb: if something is only optimized, you don’t need to label it. If it implies a recording situation that never happened, you do.

One question remains open: whether a realistic but fictional face alone is enough. Courts will have to clarify that. Until then, favor caution.

• 3 Prüffragen – Grundlage zur Beurteilung von Deepfakes
• 3 Mal Ja – Kennzeichnungspflicht
• 1 Mal Nein – in der Regel keine Pflicht

Case two: AI‑generated text rarely affects marketing

This is the biggest relief. Three conditions must be met together. The text was generated by AI. It is published — accessible to an indeterminate number of unrelated readers. And it informs the public about a matter of public interest, like politics or health.

That excludes product descriptions, promotional landing pages, social posts, SEO guides, and newsletters in most cases. The purpose of the rule is to protect democratic discourse, not to police marketing copy.

Even if a text addresses a public‑interest topic, the obligation disappears when two things coincide. A human has reviewed the content, and a named person accepts editorial responsibility. A clicked approval box is not enough. The review must be sufficient to identify errors, misleading claims, and weak sources.

That is the point of the whole regulation, and almost nobody talks about it. The label alone doesn’t save you. The named person with real editorial judgment does. Which is precisely what you need anyway to guard against AI hallucinations.

This article is itself an example. It covers a public‑interest topic, it was produced with AI support, and we reviewed and stand behind it. At our organization, thought leadership is tied to content with a named author. Editorial responsibility therefore always rests with the author, whether it’s a blog post, a social post, or a presentation.

The chatbot is the only immediate technical hotspot

Users must clearly recognize at the first interaction that they are talking to an AI system. There is no grandfathering for running systems and no mandated format. A notice at the start of the conversation is sufficient, a visible badge works as well, and for voice assistants an audible notice is acceptable. That is the only point with real technical work required before Monday.

How to label correctly

The label must be clear, unambiguous, accessible, and visible at first perception. The imprint, terms and conditions, profile description, alt text, or tiny grey text at the image edge are not enough. Metadata alone also fails under the EU guidelines. The label must sit on the content.

For images, a high‑contrast in‑image note or a clearly visible caption meets the requirement. For video, place the notice at the start and in the description; repeat it for recorded content. For audio, include it in the title and episode description. For text, place it up top, near the headline — not at the end. And the label should travel with the content when it is redistributed.

1. Image
   A high‑contrast note directly in the image or a clearly visible caption meets the requirement. The label sits on the content and must be perceivable without detours. It should travel with the asset when shared.
2. Video
   The notice belongs at the start of the video and in the description. For recordings, repeat it so context remains clear even if viewers skip sections.
3. Audio
   For audio content, place the label in the title and episode description. That keeps the information accessible outside the player.
4. Text
   Texts should be labeled at the top, near the headline — not at the end. The label must be clear, unambiguous, and accessible, and it should accompany the content when reused.

The three EU icons are voluntary and available from the European Commission. For German‑language audiences a combined label can make sense because not everyone reads the English text: "AI MODIFIED. Dieses Bild wurde mithilfe künstlicher Intelligenz verändert."

One relief few know: content created and published before August 2 does not need retroactive labeling. No one must clean up their archives.

The real risk is not the fine

Formally, fines can reach up to €15 million. For everyday mid‑market operations that’s mostly theoretical.

More realistic is the cease‑and‑desist letter. A labeling breach can count as misleading by omission, and competitors or consumer protection groups can pursue injunctions. The opposite risk is AI washing: claiming AI when people actually did the work creates the same problem from the other side.

So it’s not the regulator that’s most likely to act. It’s a competitor. That’s closer to reality and a better argument for a practical process than any headline fine number.

Why this still fails in day‑to‑day work

So far it’s clear: this is not a blanket rule; these are test steps. Test steps work in individual cases and fail at scale.

A team of five people and three agents produces forty assets in a week. If the rule lives only in heads, they’ll apply it forty different ways and no one can later explain why. A training in August won’t fix it either, because four weeks later no one remembers the line between color correction and reality distortion.

This has stopped being just a compliance issue. It’s an architecture issue.

How we solve this with clients

We ran this with a client a few weeks ago and made the labeling logic a central governance rule embedded in every agent. The result: a few surprising individual cases that no one had anticipated, but relatively little overall that actually required labeling. The rule reduces friction once it’s set up cleanly.

What matters is not the number of measures, but a well‑built governance entry. It can be as short as:

Check every image output for imitation of real people, objects, places, or events and whether it could be perceived as a real recording. If yes, provide a labeling suggestion for the asset including placement. Check every text for public‑interest topics. If yes, indicate the need for a named editorially responsible person.

Where you integrate this rule depends on your setup. In an Operating System it sits as a governance file in the context layer and applies to every agent. If you build your own agents or skills, put the check where content is created. If you work directly in ChatGPT, Copilot, or Claude today, include it in the base system instruction. I tested this in our own assistant and in a medical content workflow. It works.

These approaches vary in strength. An instruction in a personal account applies to one person and disappears when they leave. In the context layer it applies company‑wide and is auditable. But any version beats a rule that lives only in someone’s head.

The most important part isn’t in any file. Responsibility remains with the person who reviewed the content. That mindset must land with teams: we stand behind our content, whether a machine produced it or we typed it ourselves.

Client case: embedding labeling logic in governance

At one client, we added the labeling logic as a central governance rule and made it part of every agent. This produced few labeling cases overall, but a handful of surprising instances no one had foreseen.

Depending on setup, the rule can live in the operating system as a governance file in the context layer, in custom agents or skills where content is created, or as a system instruction in tools like ChatGPT, Copilot, or Claude. We tested the approach in our own assistant and in a medical content workflow — it worked.

What matters is not volume, but a clean governance entry that makes reviews consistent and eases the team’s workload.

What you can do this week

• Check the chatbot. Does it clearly disclose at the first interaction that it’s an AI system?
• Review the last twenty published images — not the entire archive. Which give the impression of a real recording? That sharpens judgment faster than any training.
• Write the three deepfake questions down and place them where content is created. Do it once, not weekly.

What you don’t need: retroactively labeling your archive, blanket labeling everything, or launching a large compliance program.

Frequently asked questions about AI labeling in marketing (FAQ)

Do I have to label website texts if they were edited with AI?

No. Labeling only applies to texts published to inform the public about matters of public interest. Promotional content is generally excluded. The obligation also disappears if a named person conducts an editorial review and accepts responsibility.

What counts as a deepfake and when must I label it?

Three criteria matter: AI generation or manipulation, resemblance to real people or events, and the potential to be mistaken for a real recording. If all three apply, you must label it. Clearly artificial graphics or pure optimizations of real recordings usually do not require labeling.

How and where should I place the label?

The notice must be clear, unambiguous, and accessible directly on the content. For images, a high‑contrast in‑image note or a visible caption works. For video, place it at the start and in the description (repeat for recordings). For audio, include it in the title and episode description. For text, place the notice at the top, near the headline. The notice should travel with the content when it’s shared.

Do older assets need retroactive labeling?

No. Content created and published before August 2 does not need to be labeled retroactively. You do not need to clean up your archive.

How serious is the legal risk in reality?

Formally, fines can be high, but in practice cease‑and‑desist actions are more likely. Violations can be deemed misleading by omission, and “AI washing” carries its own risk. Solid processes and clear responsibilities reduce exposure.

How do I ensure teams apply the rules consistently?

Embed the checks in a central governance rule rather than only in people’s heads or training slides. Depending on your setup, the rule belongs in the context layer, in agents, or in the base system instruction of the tools you use. And keep a named, reviewing person with editorial responsibility.

This is practical guidance, not legal advice. The regulation does not require a sticker. It requires a person who stands behind the content. The only question is whether that person is named at your organization or whether it’s a different person every day.