EU AI Act: Marking AI-Generated Images

AI now produces a large share of the product imagery in fashion e-commerce, and with that shift comes a fair question from shoppers, marketplaces, and regulators alike: can you tell what came from a camera and what came from a model? The European Union has answered it with a rule. From 2 August 2026, any system that generates or substantially edits synthetic images has to mark its output so that software can tell it was made by AI.

For a fashion brand, this is not an abstract compliance headache. It is a question your legal or procurement team will ask before they sign off on any tool that touches your catalog: "If we publish these images, are they labelled the way the law expects?" We think the honest answer a vendor should be able to give is a simple one. Yes, and you did not have to do anything to make it happen.

This post covers what the EU AI Act's transparency rules actually ask for, what On-Model now does automatically on every image, and where your responsibility as a brand begins.

What the EU AI Act actually asks for

The relevant part of the law is Article 50. It sets out two duties that matter here.

The first is a provider duty. Whoever builds the AI system that generates or substantially alters an image must mark the output in a machine-readable format, so that it is detectable as artificially generated. On-Model is that provider.

The second is a deployer duty. Whoever publishes an image that qualifies as a deep fake must disclose that it is artificial. When you publish images downstream, that role is yours.

The phrase to hold onto is machine-readable. It does not mean a visible logo stamped across the picture. It means information that travels with the image file, in a place people rarely look but software always can, so that a platform, a browser, a marketplace, or a verification tool can read it and know the image was AI-generated. In June 2026 the European Commission published a Code of Practice on marking and labelling AI-generated content that spells out what "good" looks like: standardized metadata first, with more durable layers recommended on top. The Code is voluntary, but it is the blueprint regulators will measure against.

There is a narrow exception for "assistive" edits that do not substantially change an image. Swapping the model in a photo, or generating a person onto a flat-lay garment, is not that. We treat our outputs as in scope, which is the safe and honest reading.

How On-Model marks AI-generated content

Whichever tool you use, whether Model Swap, Flat-to-Model, Create Packshot, or any of the others, the finished image leaves the platform carrying a standardized marker that identifies it as artificially generated. The marker uses a value defined by a recognised industry standard, not a private flag only we can read, so any tool that knows how to look for AI-provenance information can find it. That is what makes it interoperable, which is exactly what Article 50 asks for.

It covers the whole surface, not just the first attempt. First generation, re-runs, regenerated variants, and post-processed versions all carry the marker. Free plans are included. That last point was a deliberate decision: the law does not have a free tier, so neither does our marking. An image made on the free plan leaves marked in the same way an enterprise image does.

Note what we chose not to do. We did not make this a paid add-on or an enterprise-only toggle buried in settings. Compliance that someone has to remember to enable is compliance that eventually fails in production. Making the marker unconditional is the only version of this that actually holds up at catalog scale.

Why marking at the source matters

On-Model is less a creative app you open to make one image, and more a piece of infrastructure your catalog flows through. The useful comparison is Stripe rather than Photoshop: product catalog in, processed imagery out, on into your storefront, your marketplaces, and your ads. The AI is one step inside a larger system.

Because the marking happens inside that pipe, at the exact point the image is made, every brand publishing through On-Model inherits it. You do not add a compliance step to your workflow. You do not train your team to remember it. You do not audit each image by hand. The label is already applied before the file ever reaches you. Compliant by default.

That is the difference between a rule you have to operationalize and a rule that is simply handled for you. In the same way you do not reimplement card-network compliance every time you take a payment, you should not have to reimplement content-transparency marking every time you generate an image.

Where your responsibility begins

Marking the image is our job as the provider. It is not the whole of the law, and it would be dishonest to imply otherwise. If you publish an image that a viewer could reasonably mistake for a real photograph of a real person, in a context the law treats as a deep fake, you as the deployer may still need to disclose that it is AI-generated. The marker we embed makes that easy to prove and easy to automate, but the disclosure decision sits with you, because only you know where and how the image will run.

What we ask in return, and what our Terms now set out, is straightforward: do not strip the marker. Removing the embedded provenance information defeats the purpose and can push the downstream publisher onto the wrong side of the rule. Our Terms describe this shared-responsibility split in full, including your obligations as the deployer of the content.

For most catalog use, a model wearing your garment on a clean studio background that is clearly a product image, the machine-readable marker is enough and no extra visible disclosure is expected. For anything that leans hard into looking like a candid photograph of a real individual, check the context before you publish.

If you decide a visible label is warranted, you do not have to add it by hand. On paid plans, the options step of every tool includes an optional "AI-generated watermark" switch that bakes a visible disclosure mark into the corner of each output. It is off by default and permanent once applied, so it is there for exactly the cases where you want the disclosure on the image itself. The automatic machine-readable marker is always present underneath, whether or not you turn the visible one on.

What comes next

Machine-readable marking is the foundation, not the finish line. Metadata can be stripped by someone who edits or re-exports a file in other software, which is precisely why the Code of Practice recommends layering rather than relying on a single signal. We are building additional provenance and robustness on top as the standards settle, so the marker survives more of the journey from our pipeline to a shopper's screen. The direction is the one the regulation points toward: make it progressively harder to lose track of what was made by AI.

Common questions

Does On-Model comply with the EU AI Act?

On-Model marks every image it generates in a machine-readable format that identifies it as AI-generated, which is what Article 50(2) asks of providers. This is live today, on every plan. Your own legal team should confirm how the rules apply to your specific use, but the marking obligation on our side of the line is met.

Do I need to switch anything on?

No. The machine-readable marking that satisfies the law is automatic and applies to every image from every tool, on every plan. There is no setting to enable and no way to accidentally disable it. A separate, optional visible label is available if you want one, covered just below.

Does this apply to free plans?

Yes. The marker is on every output regardless of plan. Transparency the law requires should not depend on what you pay, so with us it does not.

Can I add a visible "AI-generated" label too?

Yes. Separately from the automatic machine-readable marker, paid plans include an optional switch in each tool's options step that bakes a visible "AI-generated" label into the corner of the image. It is off by default and permanent once applied, so it is there for when you decide a visible disclosure is warranted, for example on content that could be mistaken for a real photograph. The invisible marker is applied either way.

Can the marking be removed?

The embedded information can be stripped by editing or re-exporting the file in other software, the same way any metadata can. That is why we treat marking as the first layer, are adding more durable ones, and ask in our Terms that you do not remove it. Within On-Model, every delivered image carries it.

Is On-Model an EU company?

Yes. On-Model is built by PiktID, a company based in Austria. We operate inside the same EU regulatory framework the AI Act belongs to, alongside GDPR and Austrian data-protection law. We are complying with EU rules as an EU company, not retrofitting a product from elsewhere to fit them.

Further reading

• European Commission — Code of Practice on marking and labelling AI-generated content
• EU AI Act — Article 50 (full text)
• On-Model — AI-Generated Content & Transparency (our Terms)
• The AI infrastructure for fashion imagery

For the broader market context, our fashion e-commerce statistics hub gathers sourced figures on AI adoption across the industry.

Want to see it in practice? Sign up and run a product through On-Model. Every image you get back is already marked the way the law expects, without a single extra step on your side.