Updated September 2026. Reading time: about 4 minutes. General information, not legal advice.
Six weeks ago, on 2 August 2026, a new European rule quietly changed how AI-generated images can be published. It is called Article 50 of the EU AI Act, and if your company generates or publishes AI imagery for the European market, it applies to you today. Not in December. Not next year. Today.
This is the short version. The complete guide, with the full classification tables and an FAQ, lives here: the complete guide to EU AI Act Article 50 for fashion images.
What Article 50 requires, in one paragraph
AI-generated and AI-manipulated content must be identifiable. That breaks into two separate duties. The platforms that generate images must embed machine-readable marks inside every file: signed metadata plus an invisible watermark that software can detect. And the companies that publish certain kinds of AI imagery, the ones the law classifies as deepfakes, must add a visible label that a human can see. Two duties, two different responsible parties, and one common misunderstanding: the platform's invisible marks do not satisfy the publisher's visible-label duty. The European Commission has said so explicitly.
Who is affected
- AI platforms and tools that develop or offer AI image systems carry the marking duty. This includes companies that build products on top of someone else's model: selling under your own name makes you the responsible provider.
- Brands, retailers and agencies that publish AI imagery are deployers under the law. If a published image qualifies as a deepfake, and much commercial fashion imagery does (a photorealistic synthetic model counts, and so does virtual try-on on a real model's photo), the brand must label it visibly. The free official EU icons exist for exactly this.
- Photographers and studios using AI tools professionally sit in the same deployer seat as brands when they publish the results.
The three dates that matter
| Date | What it means |
|---|---|
| 2 August 2026 | The rules became applicable. Publisher duties, including visible labels, are enforceable now, with no transition period. |
| 2 December 2026 | End of a narrow transition window that covers only the machine-marking duty of AI systems that were already on the market before 2 August. It does not delay anything for publishers. |
| 2 February 2027 | Marks must point verifiers to a detection method, so anyone can check content without guessing which tool to use. |
And one date that never arrives: content generated before 2 August 2026 does not need to be marked retroactively. The rules look forward, not back.
The fine, and the more realistic risk
The penalty ceiling is 15 million euros or 3 percent of worldwide turnover, whichever is higher. Enforcement is national: in Spain the authority is AESIA. But the number that should actually change behaviour is a different one: any person can file a complaint. A competitor. A consumer association. An NGO. There is no gatekeeper.
Which makes the realistic short-term scenario not an inspection but a letter: a request to explain how your company complies. The expensive version of that moment is having no answer, scrambling to reconstruct which images were AI-generated, what was done to them, and who was supposed to be responsible. The cheap version is a one-page answer written calmly, in advance.
Three things to do this week
- 1. Ask whoever produces your imagery, internal team or vendor, one question: which of our published images are AI-generated or AI-modified, and what marks do they carry?
- 2. Decide who owns this topic in your organisation. Legal, marketing, or a named person. In many companies today the honest answer is nobody, and that is the gap a regulator's letter finds first.
- 3. Read the classification tables in the complete guide to see which of your images need labels. Most teams are surprised in both directions, by what is exempt and by what is not.
