Labelling AI Images: What Is Mandatory, What Is Sensible — and How to Do It Properly
Do AI-generated images on websites need to be labelled? Where the obligation applies, why voluntary labelling is a smart move, and what a clean implementation looks like.
Few questions reach us more often: Do I now have to label my AI-generated images? The honest answer has two sides — a legal one and a strategic one. Both ultimately point to the same solution, but for different reasons.
The legal perspective: where the obligation applies
Art. 50 of the EU AI Act requires disclosure primarily where there is a risk of deception: with deepfakes — that is, content that depicts real people, places, or events in a convincingly realistic way — and with AI-generated texts on matters of public interest. An obviously illustrative AI image in an online shop does not automatically fall under this. However, the boundaries are blurred: A photorealistic image that appears to be a genuine photo of your team or your shop is in a different category to a stylised graphic.
There is also the provider side: Generative systems must mark their outputs in a machine-readable way. Tools such as Adobe Firefly already embed provenance information in the file according to the C2PA standard. These markers travel with the image — and reveal when something has been published without labelling.
The strategic perspective: why voluntary labelling is wise
The trend is clear: towards more transparency, not less. Codes of conduct, platform rules, and advertising guidelines are already following suit. Those who only label the bare minimum today will be revisiting the issue in twelve months. Those who label consistently have dealt with the topic — and can even showcase it: transparency is a signal of trust, not a flaw.
Three design rules for effective labelling
- Subtle but present: A small note in the corner of the image is sufficient. No one requires a warning sign over the subject.
- Proportionate to size: On a thumbnail, a dot symbol is enough — a text banner would overwhelm the image.
- Dual approach: visible to people, machine-readable for auditors and platforms. Only both together are future-proof.
The real challenge is the effort — not the rule itself
Manually checking and labelling hundreds of existing images is unrealistic. Automated detection plus a snippet to set the labels can achieve in minutes what would take weeks by hand.
As always: technical assessment, not legal advice. Whether a specific image is subject to the obligation in an individual case is something your lawyer must determine.
Free audit
What is actually on your website?
Enter your domain, wait less than a minute — you will see which content shows traces of AI generation. No registration required.
Read more
Art. 50 EU AI Act: Who Must Label AI Content — and What This Means in Practice
The practical guide to labelling obligations: affected content, the operator’s role, and the most common misconceptions.
How to Detect AI-Generated Content — and Why No Single Detector Is Sufficient
Why metadata provides the strongest evidence, classifiers can be misleading — and why, ultimately, a human should make the final decision.
EU AI Act: The 10-Point Checklist for Your Website
From inventory to monitoring: the ten questions every website operator should now be able to answer.
Deepfakes on Your Own Website: Why the Operator Is Liable — Not the Creator
The law’s most powerful lever does not target AI providers, but you — and deepfakes arise more quickly than most people realise.
C2PA, Metadata, Watermarks: How Machine-Readable AI Labelling Really Works
A look inside the file rather than at the image: the standards that make origin verifiable — clearly explained.
Stock Photos and AI: Why Almost Every Website Contains Unlabelled AI Content
You have never knowingly used AI? Your website probably has — in ways no one expects.