Not always. Article 50 of the EU AI Act puts two different duties on AI images. The tool that made the image has to mark it in a machine-readable way. You, as the website that publishes it, have to add a visible label only when the image is a deep fake: realistic content that resembles real people, objects, places or events and could falsely appear authentic.
Key point: An AI illustration that obviously isn't real usually needs no visible label. A photorealistic AI image of a person, product, place or event that a visitor could take as genuine does, from the moment they first see it.
Two duties, two different people
| Duty | Who it applies to | What it requires |
|---|---|---|
| Article 50(2): marking | The provider of the AI tool (the company behind the image generator) | Outputs marked "in a machine-readable format and detectable as artificially generated or manipulated" |
| Article 50(4): disclosure | The deployer: you, if you publish the image | Disclose that a deep fake has been "artificially generated or manipulated" |
The machine-readable marker (for example C2PA metadata or a watermark from the generator) is the provider's job. The European Commission's Article 50 FAQ is explicit that this metadata does not satisfy your own disclosure duty: a deep fake needs a label people can actually perceive.
When is an image a deep fake?
The AI Act defines a deep fake as AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful (Article 3(60)). In practice, ask three questions:
- Does it show something that exists, or could exist? A dragon, a flying person or a talking animal can't pass as real. A person in an office, a product on a table or a street scene can.
- Is it realistic? Photorealistic renders are in scope. Flat illustrations, obvious 3D cartoons and clearly stylised art generally aren't.
- Could a visitor take it as genuine? If someone could reasonably believe it is a real photograph, treat it as a deep fake.
Common website cases:
- AI "team" or customer photos of people who don't exist: realistic and resembling persons. Label them.
- Photorealistic AI product shots: the Commission's draft guidance listed consumer goods under "objects", so a realistic AI image of a product can fall within scope. Label them.
- AI hero illustrations, icons, abstract backgrounds: usually not deep fakes. No visible label needed, though some sites add one anyway.
- Advertising: the lighter regime for evidently artistic, creative, satirical or fictional work doesn't help where the purpose is mainly commercial.
What a label can look like
Article 50(5) says the information must be clear and distinguishable, given at the latest at first exposure, and meet applicable accessibility requirements. For images that means:
- On or next to the image, visible without hovering, clicking or opening a menu. A line in your terms or an image's alt text alone is not enough.
- Plain words: "AI-generated image" or "Image created with AI". The AI image and video disclosure template has captions, badges and alt text to adapt.
- Readable: enough contrast and size to read on mobile.
- Accessible: include the same words in the alt text, so screen-reader users get the disclosure too.
<figure>
<img src="team.jpg" alt="AI-generated image of a support team at work">
<figcaption>AI-generated image</figcaption>
</figure>
What about the 2 December 2026 date?
Article 50 has applied since 2 August 2026. Generators already on the market before that date have until 2 December 2026 to add machine-readable marking (Article 50(2)). That grace period is for the providers of the tools. It does not delay your duty to label deep fakes you publish. Content created before 2 August 2026 doesn't have to be labelled retroactively, although the Commission encourages it. For what else changed with that date, see the Article 50 deadline of 2 December 2026.
How to check your own site
- List every page that uses AI-generated or AI-edited images.
- For each image, run the three deep fake questions above.
- Add a visible caption and matching alt text to every image that could pass as real.
- Re-check after redesigns and new campaigns, which is when unlabelled images usually creep back in.
What our scan does and does not check
The Article50.io scan loads your live pages in a real browser and reads up to 8 images per page. It flags images that carry an AI marker (such as C2PA or generator metadata) but show no visible label, and checks whether disclosures you do have are visible and readable. Image labels are currently recognised in English only, so a German or French caption may show up as missing. It cannot tell whether an image without any marker was made with AI, and it cannot decide for you whether an image is realistic enough to be a deep fake: those findings are flagged for your judgement. For what still needs a manual review, see Article50.io vs manual compliance checklists. Run a free scan.
Frequently asked questions
Do I need to label every AI-generated image on my website?
No. Article 50(4) requires a visible disclosure for deep fakes: realistic images that resemble real people, objects, places or events and could pass as authentic. Clearly non-realistic illustrations usually don't need one.
Is the C2PA or watermark metadata from my image generator enough?
No. Machine-readable marking is the provider's duty under Article 50(2). The Commission's FAQ says metadata alone does not satisfy the deployer's duty to disclose a deep fake, which must be perceivable without special tools.
Do realistic AI product photos count as deep fakes?
They can. A photorealistic image of a product that a shopper could take as a real photograph meets the realism and resemblance tests. Labelling it is the safer course.
Where should the label go?
On or directly next to the image, visible at first view without any click or hover, with the same words in the alt text.
Sources: Article 50, AI Act Service Desk; European Commission, Transparency obligations under Article 50 (FAQ); Guidelines on transparency obligations, 20 July 2026. This article is general information, not legal advice.