EU AI Act
AI-generated Your chatbot is answering a customer question right now. Last week’s newsletter came from an AI. So did the image in the last post.
Three cases. Which one has to be labeled?
Most companies guess wrong. Either they label everything out of fear, or they ignore the question entirely. Both are wrong. The answer is narrower and clearer than you think.
The question isn’t whether you use AI, but what you have to disclose
The EU AI Act governs AI transparency in Article 50. It applies from 2 August 2026, and unlike the big high-risk deadlines, the Digital Omnibus did not postpone this date. Article 50 affects almost every company that uses AI for content and communication. So practically everyone.
But it doesn’t require branding every AI word. It clearly distinguishes who labels what.
What the provider does, not you
The technical labeling of AI-generated content, the machine-readable watermark in image, text, audio and video, is the obligation of the provider of the AI system. So OpenAI, Google or the maker of your image tool. Not you.
That’s the relief up front: you don’t have to embed watermarks in files. That happens on the provider side, increasingly automatically.
What actually is your obligation
As a deployer, a different, manageable part applies to you.
Chatbots have to identify themselves. Anyone who writes or speaks with your AI must know that no human is answering. Unless it’s obvious anyway.
Deepfakes you have to label. If you create or alter image, audio or video so that it could appear real, it must be clear: this is AI. An exception applies only to recognisably artistic, satirical or fictional works; there a brief note that doesn’t disrupt the work is enough.
AI texts on matters of public interest you have to disclose. If you publish AI-generated pieces on socially relevant topics, a labeling obligation applies.
And if you use emotion recognition or biometric categorisation, you have to inform the people affected.
The exception almost everyone overlooks
For AI texts on public topics there’s an important relief. If a human has reviewed the content and someone bears editorial responsibility, the labeling obligation falls away. What matters is not that a name appears above it, that’s almost always the case with published texts anyway, but that a human review actually took place.
An important caveat: this exception applies only to text. For AI-generated images, audio or video that count as deepfakes, it doesn’t help.
A practical example
A press release, drafted with AI. As soon as a human reviews and approves it, the exception applies, no labeling needed. And because that’s exactly the norm for press texts, an approved piece practically never has to be marked as AI-generated. It’s different for fully automatically generated content published without review, such as an automated news feed or mass-generated web pages. There, editorial control is missing, and the AI origin must be disclosed. The difference isn’t the name above the text, but whether a human is responsible for it.
If the reviewed article contains an AI-generated deepfake image, that image still has to be labeled. The text review doesn’t save the image.
And the images?
Here it gets practical, and the answer is stricter than many hope. According to the draft of the Commission’s guidelines on Article 50, even a photorealistic image of an invented person counts as a deepfake, because such a person could exist in reality. So it doesn’t matter whether a real person is depicted, but whether the image could appear real. Only what is obviously impossible stays out, such as a dragon at the wheel or a sphinx above the Eiffel Tower.
For the typical photorealistic marketing and blog images, that means: they have to be labeled. For recognisably artistic or fictional works, a discreet note is enough, but a note it must be. A simple “Image: AI-generated” fulfils this, makes the difference from a licensed photo immediately visible and protects against the accusation of deception. (These guidelines are still in draft, but the direction is clear.)
A second point lies beyond the AI Act: personality rights. A certain resemblance to real people can never be entirely ruled out with AI images, the models are trained on real persons. It only becomes critical, though, when a specific, identifiable person becomes recognisable. So don’t use images that show real people recognisably, and use tools with a clear commercial usage license.
Not whether, but how
If you have to label, the real question isn’t the whether, but the how. And here the AI Act is clear: the notice must be clear and distinguishable, appear at the latest on first perception, and be accessible without barriers (Article 50(5)). Hidden in the terms and conditions or behind a link does not fulfil the obligation.
Is a plain “AI” enough then? Probably not. The EU guidelines and the Code of Practice on labeling, both still soft law, suggest clearer wording, such as “AI-generated” or “AI-edited”, complemented by a uniform symbol. That isn’t binding yet, but through the AI Office’s expanded powers after the Digital Omnibus, these guidelines gain practical weight.
And it’s not just the AI Act. Anyone who uses AI content in a misleading way can also come into conflict with competition law, and in multilingual markets the respective language requirements come on top. Those who label clearly and understandably are on the safe side against all of that.
Transparency isn’t a risk, it’s a signal
Those who draw a simple line now, what we label and what we don’t, are prepared before the guidelines become more binding.
And it pays off. Saying openly where AI is involved doesn’t look weak, it looks confident. Customers trust a company that discloses its tools more than one that hides them.
The prerequisite is that you even know where AI generates content in your house: in marketing, in service, in sales. That’s exactly what the NADOVO platform makes visible, including the tools nobody officially registered. What gets captured is not just the tool but the AI process, because the same text generator can trigger a different transparency obligation depending on its use. Which of them need labeling and how to set up a lean transparency rule is something we clarify in our AI compliance consulting.
One question to close. If a customer asks tomorrow whether they’re currently writing with an AI or reading an AI text: would you have a clear answer? If not, your transparency starts right there. Where you stand is shown by our quick check.
The fundamentals of the EU AI Act are covered in more depth in our article on what companies need to know now.
About the author
Jochen Stier is a co-founder of NADOVO with over 20 years of experience in process management and IT service management. He helps German SMEs implement the requirements of the EU AI Act systematically and pragmatically. His 5-phase NADOVO framework combines regulatory requirements with practical feasibility, without enterprise budgets or complex tools.
Further reading:
- Deployer obligations in the EU AI Act (Contoro)
- Code of Practice on labeling AI-generated content (European Commission)
- EU AI Act, Article 50 full text (EUR-Lex)