Do AI images need to be labeled under the EU AI Act?

The short answer

Yes, but the obligation depends on whether you are a provider or a deployer, and on what the image depicts.

What providers must do

Providers of AI systems that generate synthetic images must ensure outputs are marked in a machine-readable way and detectable as artificially generated. This means the AI system itself should embed metadata, watermarks, or other technical markers that make the image identifiable as AI-generated.

This is a provider obligation — it applies to the company that builds the AI system (e.g., OpenAI, Midjourney, Adobe), not to the company using it.

What deployers must do

If you use an AI system to generate images and publish them, you are a deployer. Your obligations depend on what the image depicts:

Regular AI-generated images

The provider must ensure machine-readable marking. As a deployer, you should also apply a visible label as best practice:

"This image was generated using artificial intelligence."

Deepfakes (images resembling real persons)

If you publish AI-generated or manipulated images that resemble real persons, places, or events, you must disclose that the content is artificially generated or manipulated. This is a deployer obligation under Article 50.

Label:

"This image has been artificially generated and may not depict real events."

When labeling is not required

• Personal, non-professional use — images you generate for yourself and don't publish
• Content obvious as AI-generated — if a reasonably well-informed person can tell it's AI-generated
• Artistic or creative content — disclosure must not hamper the work, but some form of disclosure is still expected

Practical guidance for teams

1. Audit what AI-generated images you publish
2. Apply visible labels at the point of publication
3. For deepfakes, use stronger wording and place the label prominently
4. Log what was labeled, when, and by whom

The full labeling guide

The Labeling Kit (€99) includes a complete labeling guide, wording library, decision tree, and record log for images, audio, video, and text.