Labeling AI-generated marketing content: a guide
AI-generated images and copy in marketing: how to meet AI Act art. 50 and the machine-readable marking rules live from August 2026. A practical guide.
The AI Act does not ban using AI-generated images or copy in marketing. It just asks you to say so, in a way the user understands and the machine recognises.
AI Act art. 50 requires anyone publishing synthetic content (images, audio or video generated or manipulated by AI that could look authentic) to disclose it clearly to the user, and requires the providers of generative tools to mark it in a machine-readable format. In marketing, the obligation applies from 2 August 2026.
From August 2026 this becomes an operational obligation, not a best practice. Here is how to label AI-generated marketing content without slowing down your publishing flow.
Key takeaways:
- AI Act art. 50 requires transparency on synthetic content (images, audio, video) that could look real: whoever publishes it discloses it to the user, the tool that generates it marks it in a machine-readable way.
- The obligations apply from 2 August 2026; for systems already on the market before that date, the providers’ marking duty starts on 2 December 2026.
- Not every AI-assisted text needs a label: the test is the risk of it being mistaken for authentic.
- Machine-readable marking (metadata, watermark) is distinct from the visible label: the tool applies the first, you decide the second, and you need both.
- A clear labeling process covers AI Act, GDPR and consumer law together.
What art. 50 actually asks
Art. 50 of the AI Act is about transparency. It splits into two obligations that, in marketing, often overlap:
- Disclosure to the user: when you publish content generated or manipulated by AI that could appear authentic (a person, place or event that looks real), you must say so clearly and understandably.
- Machine-readable marking: the content must carry a marker readable by machines, typically metadata or a digital watermark, certifying its artificial origin even when re-shared beyond your channel. It is a duty of the generative tool’s provider (para. 2): your part is not losing it along the way.
The difference matters. The visible label serves the human looking at the post. The machine-readable marking serves platforms, engines and other systems that process the content downstream.
What does NOT need labeling
This is where most consultancies over-reach. Not everything the AI touches becomes “synthetic content” to mark.
- Text written with ChatGPT’s help but reviewed and signed off by a person needs no explicit label.
- A product image edited with standard tools (exposure, crop) is not synthetic content.
- An internal AI-generated report that is never published falls outside the obligation.
The operational test is one: could the audience mistake this for authentic when it is actually AI-generated or manipulated? If yes, label it. If no, don’t.
The 4 steps to fix your process
1. Map where you use synthetic generation
Do the honest inventory: images generated for social and ads, video with avatars or synthetic voices, “invented” product shots, fictional testimonials. That is where the obligation bites, not the assisted copy.
2. Define a standard label
A short, consistent wording (e.g. “Image generated with AI”) applied uniformly. You don’t need a long legal disclaimer, you need clarity.
3. Verify the machine-readable marking
Many serious generative tools already embed provenance metadata (standards such as C2PA). Don’t assume: check it survives the export and the publish step on your platform. If your flow strips the metadata, that is the point to fix.
4. Write one line in your AI policy
One line in the internal policy stating: anyone producing commercial communication with synthetic content applies label + marking. Light governance, but traceable.
When you don’t need a heavy intervention
If your synthetic content production is occasional (a few images for one-off campaigns), the fix is process, not software: a publishing checklist and a standard label are enough. You don’t need a dedicated compliance platform.
It becomes a real topic when you generate synthetic content at volume, across several people and channels. There the marking must be automated in the multi-channel publishing flow, otherwise someone forgets the label and consistency breaks.
One last point that often gets overlooked: cleanly marked content is also more trustworthy for AI engines. If you are working on visibility on ChatGPT and Perplexity, content provenance and transparency pull in the same direction as compliance.
For the full picture of AI Act obligations on your company, start from the AI Act guide for SMEs.
Want to work out where your communication falls and what to do? Let’s talk in 20 minutes, or take the check-up in 3 minutes.
Frequently asked questions
What people usually ask us.
Do I need a label on every AI-assisted post?
What does machine-readable marking mean?
When do the obligations kick in?
Do I risk a fine if I don't label?
Keep reading
AI systems inventory for SMBs: your AI Act register
How to build the AI systems register the EU AI Act calls for: a template, risk classification and the first operational step for every COO and DPO.
AI Act & recruitment: what to do before December 2027
The AI Act classifies AI used for staff selection as high-risk. The 5 concrete obligations (human oversight, candidate notice, logs) to avoid fines.
Next step
Where are you on the AI journey?
The check-up gives you an AI readiness score (0-100) + 3 concrete next steps. 3 minutes, no email.