Your shop can strip the AI marking from your photo and never tell you

You ordered photographs carrying an AI marking and uploaded them to your shop. Between your hard drive and the buyer's screen, though, stands a channel that will process that image, and some channels strip the marking on the way through without telling anybody.

We checked nine of them, file by file, and the result is less obvious than it looks. Dates sit beside every measurement, because channels change their behaviour and a result from last month is not a promise for next quarter.

This piece says what those measurements mean for your shop. How exactly we took them, and where the method runs out, is covered in a separate technical post.

These are two different questions, not one

An image's marking is two layers inside one file. The first is a signed provenance record: an entry added by the tool that produced the image, protected by a cryptographic signature. The second is an ordinary metadata entry stating outright that the image came from a generative model. The buyer sees neither with the naked eye.

So every channel has to be asked two questions, not one.

Does the channel label the publication. Does a label appear beside the image or the post, one the buyer will actually read.

Does the channel carry provenance onward. Does the file the channel serves your buyer still hold those layers inside it.

The market conflates these two questions, and the answers are often opposite. There is a service that puts a label on the post and at the same time serves a file stripped to zero. There is another that labels only conditionally, but carries provenance through intact. The first question is about what the buyer sees. The second is about what you can later demonstrate.

What your shop does

The shop is the first link after your hard drive, and it usually decides everything that follows. We measured four routes: we uploaded the same file and read every size the shop then serves.

The sharpest single result came from PrestaShop. The file the admin panel calls the "original" comes back with exactly the same dimensions as the one uploaded, weighs a fifth of what was uploaded, and holds neither of the two layers. The shop recompresses even an image it does not resize, and there is no setting that turns this off. Version 9 behaves exactly like version 8.

WooCommerce has a way through, but a narrow one. Above 2560 px WordPress creates a variant and substitutes it for the "full size", and the signed record is lost. Below that threshold the file passes byte for byte. The metadata entry does better: it reaches every derivative, including the one the product page renders.

One qualification on that row, because without it the row would be a promise. We measured metadata propagation on 22.08 on WordPress 7.1 with Imagick available. An older installation, or one without Imagick, was not measured and this result does not cover it.

IdoSell behaves differently from both of those platforms. The metadata entry passes at every size, and the signed record only when you untick scaling by the default profile on the individual photograph. The price is not small: the shop then generates no derivatives at all, the product page loses its social media preview, and the buyer is served a file of nearly six megabytes on the page. This is a choice worth weighing, not a recommendation.

This row rests on weaker evidence than the others, and we would rather not blur that. We measured the metadata entry passing through on 20.08 on a single product in a single shop, on somebody else's file that carried a packet written by an entirely different tool. We checked the signed record on 5.09 with our own upload, but a file built the way our exports are built has still never passed through IdoSell. The behaviour looks like a property of the platform rather than a setting of that one shop, and that is how we read it, but that is a reading rather than a second confirmed attempt.

What the platforms do

The same two questions, asked of the four services on which sellers most often publish material.

The two "not measured" cells are in that table deliberately. The material we uploaded to Instagram and Facebook had already lost its marking before the upload, in a video editor. There was nothing left to preserve, and about that axis on those two platforms we know nothing. That is not the same as "checked, and the answer is zero".

Meta reads the record inside the file, not the watermark in the pixels. Measured 6.08 with two posts carrying the same image and differing only in the metadata layer. The file with the record received the label "AI info" in about two minutes. The file carrying only a pixel watermark received no label, neither then nor at the follow-up check a day later on 7.08.

TikTok labels conditionally. Of two pieces of material of the same nature, uploaded the same way at the same time, only the one whose provenance was attested by a recognised external issuer received a label, rather than by our own signature alone. We do not know at which point in that day the label appeared: after a quarter of an hour it was not there, and after a day it was.

On the second axis, though, TikTok is the best of the four. When it transcodes material it rebuilds the provenance chain and nests it under its own signature instead of losing it.

YouTube does exactly the opposite. The publication receives an "AI" label after a few minutes, and the file the service then serves carries not one provenance marker. The buyer sees the label at the source, but anyone who downloads that material onward will reconstruct nothing from it.

Google Merchant deletes nothing, it requires

Merchant Center is sometimes named as another link that strips markings. It is not one. Measured 6.08: Google rehosts thumbnails only and clears them entirely, but clicking a sponsored result leads straight to the seller's shop, because Google does not host the full image at all.

Merchant is a recipient that requires the metadata entry, and reads it when it fetches the image named in your feed. The link that strips is therefore your shop, not Google.

The consequence is practical and sharp. A shop on PrestaShop cannot satisfy that requirement by any route, because it strips metadata from every product image and the feed points at an image that has already been processed. WooCommerce and IdoSell pass that entry through at every size, so there the requirement can be met with no effort on your part - although for IdoSell that rests on the single measurement described above, not on two independent ones.

What we did not measure: exactly which size the image address in the feed points to. For both layers the answer today comes out the same regardless of size, so the question is housekeeping rather than decisive.

What this does not mean

Something has to be said here that is easy to leave out, because it weakens the drama. Losing metadata in your shop is not, in itself, a breach of the AI Act on your side.

Embedding the marking in the file is the duty of the system provider, under art. 50(2), and not yours. A shop does not have to embed anything in metadata for AI Act purposes. So if the shop strips an entry you never had to put there, no provision is broken by that fact.

It does matter, though, for three other things, and each of them is real.

First, for the label on the platforms. The automatic marking the buyer will see comes from the layer inside the file. Without it, it simply does not happen.

Second, for Merchant Center, which rejects the offer without that entry.

Third, for your ability to demonstrate that the disclosure arrived. That is no longer a question about metadata, but about evidence.

What follows for you

You are a deployer also when you do not generate the photographs yourself but order them from a contractor, as long as you are the one deciding on the use of AI. This follows from art. 3(4) and from the Commission guidelines, points (12) and (14): control here means responsibility for the decision to deploy and for the manner of use, not technical control. Commissioning an agency or a freelancer changes nothing about it.

The duty of visible disclosure arises only where the content is a deepfake within the meaning of art. 3(60). Art. 50(4) and guidelines point (113) require four conditions together. A real product on a generated background falls outside the definition, but only for as long as the product itself is faithfully rendered. An image that shows it more favourably or at better quality than it looks in reality is a deepfake, even with no figure in it at all. A photorealistic synthetic figure presenting a product always falls within the definition.

Where the duty does arise, guidelines point (12) adds in its final sentence the thing the channels make difficult: you are to take proportionate measures so that the disclosure is visible to the end audience also outside your own shop. It names contractual conditions with distributing partners and interface settings explicitly.

The other side of that sentence is uncomfortable. A marketplace, a search engine and an advertising network are not deployers and bear no responsibility under the AI Act (guidelines point 16), and are merely encouraged to preserve markings (point 98). Responsibility for the disclosure arriving stays on your side, even though the channel is under no obligation to carry anything.

There is one more layer that is easy to forget, because it does not come from the AI Act. The disclosure has to be readable by a screen reader, which follows from Directive (EU) 2019/882; guidelines point (144) confirm that the AI Act itself adds no separate requirements here. A label burned only into pixels does not exist for a screen reader. Both layers are needed: the visible one and the machine readable one.

That leaves the question almost every shop asks at this point. Would it not be simpler to drop the marking and hope nobody notices? The worry is understandable and it is about money rather than conscience, so the answer is owed in full.

Here is what was measured: on Meta, material stripped of its metadata layer received no label, including at the follow-up check a day later. An image classifier did not add one, even though a different tool recognised the same image as generated from the pixels alone. We had a control arm only on Instagram and Facebook, though. For YouTube and TikTok there was no such arm, so nothing about them follows from this measurement.

Dropping the marking is not a third option, however, because the duty never concerned metadata. It concerns visible disclosure where the content is a deepfake, and that disclosure reaching the audience. The real choice is therefore this: a marking you control, or a marking stuck on by the platform when and how it decides, plus exposure to the penalty under art. 99(4)(g) of up to EUR 15 million or 3 % of turnover, whichever is higher, and lower for SMEs (art. 99(6)).

What you do not have to do

The list is shorter than you fear.

You do not have to label retroactively photographs generated before 2 August 2026, or audit existing databases. Guidelines point (154) says so outright and rules out disproportionate effort, naming database audits as the example. What decides is the date of generation, not the date of publication.

You do not have to embed anything in metadata for AI Act purposes. That is the provider's duty under art. 50(2). In your hands, metadata is a tool rather than an obligation.

Nor do you have to build your own solution where the platform offers one. When a platform provides a labelling tool, you may rely on it to satisfy art. 50(4) within that platform (guidelines point 126). The words "within that platform" matter, because outside it the matter comes back to you.

What to do now

Four steps, in the order that makes sense.

Check whether the duty applies to you at all. The gate is the definition of a deepfake, not the use of AI as such. The quiz walks through the four conditions and ends with an answer rather than a list of doubts.

See what your own file carries. Upload it to the checker. The file is read in your browser, never reaches any server, is not stored anywhere, and you need no account.

Repeat the same test on a file downloaded from your own product page. That is the measurement that counts, because it measures what the buyer receives rather than what left your supplier. The gap between those two results is exactly the loss this article is about.

If a channel stripped the marking, add a visible one back. The labelling tool puts a label on the image, also in your browser and also without storing the file anywhere.

The whole path, from the decision to the finished file, is covered by the guide How to label AI images. It is written in Polish.

Of the four shop routes we measured, two carry the metadata entry all the way to the buyer and two carry nothing. Finding out which group you are in takes one file and about a minute.

This text is not legal advice. It describes our measurements and our reading of the provisions as at 6 September 2026; your situation may require a lawyer's assessment.