Updated September 2026. Reading time: about 12 minutes.
This guide is general information for professionals working with AI-generated imagery in fashion and e-commerce. It is not legal advice. Rules are applied case by case, and you should confirm your specific situation with counsel.
Why this guide exists
Since 2 August 2026, European Union law requires AI-generated and AI-manipulated content to be marked and, in many cases, visibly labelled. The rules sit in Article 50 of the EU AI Act. They apply to the platforms that generate images and to the brands that publish them, and the penalties reach 15 million euros or 3 percent of worldwide turnover, whichever is higher.
Fashion is more exposed than most industries. It produces images at enormous scale, it has adopted AI imagery faster than almost anyone, and a large share of AI fashion imagery falls under the strictest part of the rules: the deepfake provisions. At the same time, most of the guidance available is written for lawyers or for engineers. This guide is written for the people who actually commission, produce and publish fashion imagery: brand and marketing teams, e-commerce managers, studios and photographers, and the compliance officers who now share the problem.
Everything here is based on the four official documents that govern the topic, listed below with their exact legal force. Where something is still uncertain, we say so.
The rules at a glance: four dates that matter
| Date | What happens | Who it affects |
|---|---|---|
| 2 August 2026 | Article 50 became applicable. Marking and labelling duties are live and enforceable. | Everyone in scope, now. |
| 2 December 2026 | End of the transition window for generative AI systems that were already on the market before 2 August 2026. The window covers only the machine-readable marking duty of Article 50(2). | AI system providers. Brand duties were never covered by this window. |
| 2 February 2027 | Interoperability deadline under the Code of Practice: marks must point verifiers to a detection method. | Code signatories and, in practice, anyone measured against the Code. |
| Ongoing | Enforcement is national. Any person can file a complaint, including competitors and NGOs. | Everyone in scope. |
One point that surprises most teams: content generated before 2 August 2026 never needs to be marked retroactively. The duties apply to what you generate and publish from that date forward.
The four documents, and how seriously to take each one
Most confusion about Article 50 comes from treating four very different documents as one thing. They are not equal.
| Document | Status | Legal force |
|---|---|---|
| AI Act, Article 50 (Regulation 2024/1689) | Applicable since 2 August 2026 | Binding law. Fines up to 15 million euros or 3 percent of worldwide turnover. |
| Digital Omnibus on AI (Regulation 2026/1744) | In force since 27 July 2026 | Binding. Created the narrow transition window to 2 December 2026 described above. |
| Commission Guidelines on Article 50 | Final version, 20 July 2026 | Not binding, but authorities are expected to follow it. This is where scope, exemptions and the deepfake definition are explained. |
| Code of Practice on Transparency of AI-Generated Content | Final 10 June 2026, assessed as adequate July 2026 | Voluntary to sign, but recognised as the way to demonstrate compliance. Non-signatories must prove their methods are equally adequate, measured against the same text. |
The practical conclusion from that last row deserves emphasis: the Code is the exam whether you sign it or not. Companies that do not sign still have to demonstrate compliance through equally adequate means, and the Commission has said non-signatories may face gap analyses against the Code and more requests for information.
Who carries which duty: providers and deployers
Article 50 splits responsibility between two roles, and the split is decided by facts, not by contracts.
- A provider is the company that develops an AI system and offers it under its own name. Providers carry the machine-readable marking duty of Article 50(2): every AI-generated output must carry marks that software can detect.
- A deployer is the organisation that uses an AI system under its own authority. Brands, retailers and agencies publishing AI imagery are deployers. Deployers carry the visible labelling duty of Article 50(4) for content that qualifies as a deepfake, and their duties have applied since 2 August 2026 with no transition window.
Two consequences follow. First, a brand cannot outsource its labelling duty to its image platform. The Commission has stated directly that deployers cannot rely on the provider's machine-readable marking to satisfy their own obligation. Second, a platform publishing AI images in its own marketing is a deployer of its own system, with the same visible-label duties as anyone else.
One image, two disclosures
This is the single most useful distinction in the whole topic, and the one most often gotten wrong.
- Machine-readable marks live inside the file and are made for software, not people: signed metadata that records what the content is and who vouched for it, and an invisible watermark woven into the pixels. This is the provider's duty under Article 50(2). The most widely adopted standard for the metadata layer is C2PA, also known as Content Credentials.
- Visible labels are made for humans: a visible indication, such as the official EU icons, telling viewers the content is AI-generated or AI-manipulated. This is the deployer's duty under Article 50(4), and it applies to deepfake-class content.
Neither replaces the other. A synthetic fashion image that qualifies as a deepfake legally needs both: the invisible layers from the platform that generated it, and the visible label from the brand that publishes it.
Which image operations need marking, and which do not
The rules do not treat all AI use the same way. There is a genuine exemption for assistive, standard editing. The line runs between operations, not products, so the right question is what was done to the image, not which tool did it.
| Operation | Marking needed? | Why |
|---|---|---|
| Colour correction, cropping, compression, dust or red-eye removal | No | Listed as standard editing in the Commission Guidelines. |
| Removing or hiding a background; blurring faces | No | Listed as not triggering the duty. |
| Upscaling that adds no new detail | Probably not | Content is preserved. Keep a record of the assessment. |
| Generating a new background or scene | Yes | New content that changes what the image communicates. |
| Garment swap, virtual try-on, inserting or removing objects or people | Yes | Listed as substantial alteration. |
| A fully synthetic model wearing the product | Yes | Generated content, clearly in scope. |
| Short product text (under about 200 tokens) | Minimal | Free text cannot carry metadata, and short text is exempt from watermarking under the Code. |
A warning about a tempting shortcut: the business-to-business exemption. It has three conditions that must all be met, and the first is that the output is not intended to reach the public. E-commerce and campaign images exist to be published, so the exemption fails by design for fashion. Do not rely on it, and be sceptical of any vendor who suggests you can.
When a fashion image is legally a deepfake
The word deepfake suggests malicious impersonation. The legal definition is much wider: AI content that resembles a real or plausible person, object or event, and could be mistaken for authentic. The final Guidelines read this broadly. A photorealistic person who does not exist is still covered, because the image resembles someone who could plausibly exist. And the exception for artistic and creative work was narrowed almost to nothing for commercial content: where content mixes information and creativity, the informative character prevails, and every advertising example in the Guidelines requires a label.
| Type of image | Deepfake under Article 50(4)? | What the publisher must do |
|---|---|---|
| Product on an AI-generated background | No | Nothing extra beyond the platform's machine marks. |
| Synthetic photorealistic model wearing a garment | Probably yes | Add a visible label (EU icon or equivalent). |
| Virtual try-on applied to a photo of a real model | Yes | Add a visible label. |
| Photography with retouching only | No | Nothing extra. |
| A render that makes the product look better than reality | Yes, explicitly | Label it, and treat it as a consumer-protection question too. |
The last row matters beyond the AI Act. A render that flatters the product beyond what the buyer will receive is also misleading advertising under consumer law. Limits on render fidelity are becoming a legal feature of image production, not a style preference.
The Code of Practice: what it actually requires
The Code, published in final form on 10 June 2026 and assessed as adequate the following month, translates Article 50(2) into concrete measures. Its core is a layered approach, built on a blunt observation: no single marking technique survives the real world on its own.
- Layer one, signed provenance metadata: a signed, timestamped record embedded in the file (in practice, a C2PA manifest). It states what the content is, who vouched for it and when, and it detects tampering. This is the strong layer for proof.
- Layer two, an invisible watermark: an imperceptible signal in the pixels themselves. It is the layer that survives when platforms re-encode files and strip metadata.
- Optional layers: fingerprinting and logging are optional additions. The Code states fingerprinting alone is not sufficient, and forensic detection of content whose marks were removed is explicitly described as not mature enough to require.
Signatories also commit to offering a free public detection service, so that anyone can check whether content carries their marks, and to an interoperability measure by 2 February 2027 so verifiers know which detection service to use.
As of the most recent public signatory list, roughly 95 companies had signed the transparency section covering providers, including the major model developers.
What survives online, and what does not
A signed manifest is only as useful as its survival. Upload an image to Instagram, WhatsApp or X and the platform re-encodes it, stripping the metadata and with it the entire signed record, however well it was made. Some platforms preserve or display provenance information; many do not. The invisible watermark travels in the pixels and survives most everyday transformations, which is exactly why the Code requires both layers.
The practical reading for publishers: your compliance evidence and what the public can verify after distribution are two different questions. We cover the platform-by-platform behaviour in a dedicated analysis on a future blog post.
Enforcement: how trouble actually arrives
Enforcement is national. In Spain the authority is AESIA, the state agency for the supervision of artificial intelligence. Any person can file a complaint: a competitor, a consumer association, an NGO. The realistic short-term scenario for a fashion brand is not a surprise inspection. It is a complaint, followed by a letter asking the company to explain how it complies.
The expensive scenario is having no answer. A documented one, even a modest one, changes the conversation entirely: which images are AI-generated, which operations were performed, what marks they carry, who owns the question internally. The fines exist, up to 15 million euros or 3 percent of worldwide turnover, but the working risk for most companies is the scramble of assembling an answer under deadline that could have been written calmly in advance.
What to do this quarter: a practical checklist
- 1. Inventory your AI imagery: which published images are fully generated, which are AI-modified, and which operations were used.
- 2. Classify against the operations table above: which of your images need machine marks, and which also need visible labels.
- 3. Ask your image platform or vendor what marks their exports carry, and how you or anyone else can verify them.
- 4. Decide who owns this internally. In many companies today, nobody does, and that is the first thing a regulator's letter exposes.
- 5. Set up visible labelling for deepfake-class content: the official EU icons are free, come in three versions, and require no attribution.
- 6. Write the one-page answer to the question: how do we comply? If you cannot write it today, that is the gap to close before December.
Frequently asked questions
Does Article 50 apply to my company if we only publish AI images but do not generate them?
Yes, as a deployer. If the content qualifies as a deepfake, you carry the visible labelling duty, regardless of who generated the image.
Our AI image platform says its images are compliant. Are we covered?
Partially at best. The platform's machine-readable marks satisfy the provider duty. Your visible labelling duty as a publisher is separate, and the Commission has said explicitly that you cannot rely on the provider's marks to discharge it.
Do we need to mark images created before August 2026?
No. The rules are not retroactive. Content generated before 2 August 2026 never needs marking.
Is C2PA legally required?
Not by name. The law requires machine-readable, effective, interoperable, robust and reliable marking. In practice, C2PA is the only metadata standard with a mature ecosystem behind it, and the Code's description of signed, timestamped provenance is hard to satisfy any other way.
What about images we post on Instagram, where metadata is stripped?
Your duty is to publish properly marked content; you are not responsible for what platforms do to files afterwards. This is exactly why the watermark layer and detection services exist. Keep your own records of what you exported and when.
Does a small text disclaimer in the caption count as a visible label?
The label must be clear, distinguishable and visible at first interaction. The official EU icons are the safe option. A buried caption line is a weak position.
We are a B2B business. Are we exempt?
Almost certainly not for imagery. The exemption requires, among other conditions, that outputs are not intended to reach the public. Commercial fashion images exist to be published.
What are the penalties, really?
The ceiling is 15 million euros or 3 percent of worldwide turnover, whichever is higher. In practice, expect proportionality: warnings and orders first for cooperative companies. The reputational cost of being the first named case in your market may exceed the fine.
Where to go deeper
This guide is the hub of a growing series. Detailed pieces cover the machine-versus-visible disclosure split, the C2PA standard itself, what survives on each social platform, the deepfake classification in depth, the operations that trigger marking, and how enforcement works. Each links back here, and this page is updated as the series grows and as the rules evolve.
