Short answer: The EU AI Act transparency duties have applied since August 2026. Chatbots must identify themselves as AI, and synthetic output must carry machine-readable marking. In parallel, purely machine-generated text and images enjoy no copyright protection in Germany. Together, that calls for clear internal rules.
Most marketing teams have used generative tools for two years without ever deciding what that means at the point of publication. While no rules existed, that had no consequences. Since 2 August 2026 the transparency obligations of the EU AI Act apply, and a second, frequently overlooked question follows immediately: who actually owns the image a model produced? This article answers both for day-to-day business.
What the AI Act requires on transparency
Article 50 of the AI Regulation governs transparency. The European Commission summarises the duties in its FAQ on transparency obligations under Article 50. The core fits into four sentences:
- Anyone operating an AI system that interacts directly with people must ensure those people know they are talking to an AI, unless that is obvious from the context.
- Providers of systems generating synthetic text, images, audio or video must mark the outputs in a machine-readable format so they are detectable as artificially generated.
- Anyone publishing deepfakes must disclose that the content is artificially generated or manipulated.
- Text published to inform the public on matters of public interest must also be disclosed as artificially generated, unless a human has taken editorial responsibility and reviewed it.
That last point matters most to companies, because it contains the exception most of them fall under: a blog post drafted by a model and then edited and owned by a human is not the same as a text published automatically with no human control.
| Content type | Labelling required? | Who is responsible | Practical implementation |
|---|---|---|---|
| Chatbot or voicebot on your website | yes, unless obvious | you as operator | one sentence at the start of the dialogue |
| Blog post, editorially reviewed and owned | usually no | you as the editorial team | named responsibility, documented sign-off |
| Automatically published text with no human review | yes | you as operator | visible notice on the post |
| AI-generated image in advertising or social media | machine-readable marking by the provider; disclosure for realistic depictions | provider and you jointly | preserve metadata, add a caption |
| Deepfake of a real person | yes, unambiguously | you as operator | clear notice regardless of context |
| Internal documents, not published | no | – | internal marking still advisable |
The second question: who owns AI output?
This is where an expensive misunderstanding lives. German copyright law is unambiguous: under section 2 (2) of the German Copyright Act, works within the meaning of the Act are “only personal intellectual creations”. Personal means: by a human. Where no human created anything, no copyright arises.
Three practical consequences:
- You can use a purely machine-generated image – but you generally cannot stop anyone else from using the same image. A competitor copying your AI logo may not be infringing copyright at all. Trade mark and unfair competition law may help; copyright often will not.
- The more you shape, the more likely protection becomes. An image you substantially edit, compose and integrate into a larger work of your own can be protected – what is protected is your human contribution.
- The provider’s terms still bind you. Whether you may use outputs commercially is governed first by your contract with the provider, independently of copyright. Read those clauses before deploying a tool in marketing.
And a warning we have to give regularly: an AI image can absolutely infringe someone else’s rights, for instance when it reproduces a protected trade mark, a recognisable real person or a distinctive design. The fact that you generated it yourself is no defence.
A worked example: what a labelling process costs
A marketing team produces 60 pieces of content a month: 20 social posts, 12 blog sections, 18 images, 10 email sequences. The internal rate is EUR 55 per hour.
- Built in from the start: 3 minutes per item for review, metadata and a note. That is 180 minutes, or 3 hours a month – EUR 165 a month, EUR 1,980 a year.
- Cleaned up afterwards: wait two years and you have around 1,400 published items to review. At 8 minutes each – find, decide, correct, document – that is 186.7 hours, or roughly EUR 10,270 as a one-off, on top of the ongoing process you then need anyway.
The number matters less than what it says: the ongoing process is cheap, the backlog is expensive. And the backlog grows every month nobody writes a rule down.
Seven rules that fit on one page
What companies need is not a twenty-page policy but a single page everyone in marketing has actually read:
- A named human owns every published item. This is the most important rule, because it defuses almost every transparency question.
- Do not strip metadata. Providers mark outputs in machine-readable form. Running images through converters that discard metadata destroys that marking.
- Realistic-looking images get labelled. An obviously illustrative motif needs no notice; a photorealistic image of a situation that never happened does.
- Do not generate recognisable real people. Not celebrities, not employees, not as a joke.
- Chatbots introduce themselves. One sentence at the start is enough.
- Sources and figures get checked before they go out. Models invent studies that do not exist. That is routine, not an edge case.
- Maintain a register of the tools in use. Who uses which tool with which data belongs in one place. Our article on introducing GDPR-compliant AI tools shows how to set one up.
That page belongs in onboarding and in a short annual refresher. The AI Act expects operators to ensure a sufficient level of AI literacy among the people using the systems anyway.
One page of rules your team will actually read
We work with your marketing team on labelling and sign-off rules, review your existing content for gaps and anchor both in the process you already run.
What a label should look like
Do not overdo it. A banner reading “This text was created with AI” above every post looks like a warning and undermines credibility without being legally necessary. Three forms work well:
The responsibility note. One sentence at the end: the name of the person who owns the piece and the date of the last review. It satisfies the purpose of transparency and builds trust at the same time.
The caption. For generated visuals, a short addition in the caption, such as “Illustration, AI-generated”. Discreet, unambiguous, and it does not interrupt the text.
The opening line in a dialogue. For a chatbot, state up front that an AI assistant is answering and how to reach a human. The second half matters more – nothing irritates customers like a bot with no exit.
How to weave these elements into a campaign without losing its effect is covered in AI in marketing: more qualified enquiries.
Frequently asked questions
Do I have to label every blog post as AI-generated?
Under the structure of Article 50, the disclosure duty bites for text published to inform the public on matters of public interest where nobody has taken editorial responsibility. A reviewed company article with named responsibility typically falls outside it. Do not rely on that classification alone – have your case checked.
Can I use AI images in advertising?
In principle yes, provided the vendor’s terms allow it and the image infringes no third-party rights. Be careful with photorealistic depictions of products, spaces or results that do not exist – that risks unfair competition problems entirely separate from the AI Act.
Who owns a text written by ChatGPT?
No copyright arises in a purely machine-generated output, because section 2 (2) of the German Copyright Act requires a personal intellectual creation. Whether and how you may use the text is governed by your contract with the provider. In practice: you may usually use it, but you can hardly stop others from using an identical text.
What happens if I forget to label something?
The Regulation provides for member state penalties for breaches of the transparency duties, and enforcement practice in Germany will become clearer over the coming months. The bigger immediate risk is usually not the fine but the loss of trust when customers notice for themselves.
How do I sustain this day to day?
By attaching it to existing steps rather than building a new process. The review belongs in the sign-off, the note in the template, the tool register in your existing record of processing activities. Anything that needs its own meeting stops happening after eight weeks.
Next step
We help marketing teams write the one page that actually gets read, and review your existing content for gaps. Start with our AI check, look at our marketing services or simply get in touch. For a clean rollout of generative tools in the first place, see our ChatGPT rollout guide.
This article is not legal advice. For binding guidance, consult a lawyer specialising in IT and media law.
