AI images that prove what they are

Shoppers have started asking a question that did not exist five years ago: "is that photo real?"

It is a fair question. AI-generated product imagery is everywhere now, and most of it carries no answer at all. The picture looks like a photograph, sits where a photograph would sit, and says nothing about where it came from. Every time a customer spots the difference on their own, a little trust leaks out of the page.

We think the answer should live inside the image itself. Not in a caption someone might delete, not in a policy page nobody reads. In the file.

The rule that arrives in August

From 2 August 2026, the EU AI Act's transparency rules apply to AI-generated content. In plain terms: if an image was made or meaningfully altered by AI, that fact must be marked in a machine-readable way, and realistic images of people get the strictest treatment. If you sell into the EU, this applies to your product pages whether or not your business is based there. We wrote a fuller guide to the Act for shop owners in our EU AI Act guide.

Most companies will treat this as a compliance chore and bolt something on in July. We took it as a design brief.

The certificate inside the picture

The standard we adopted is called Content Credentials, built on C2PA, the open provenance standard developed by Adobe, Google, Microsoft, and the camera manufacturers. It works like a wax seal on an envelope, pressed into the image file itself.

The seal records what made the image, when, and that it is AI-generated. It is invisible in the picture and adds around 13KB to the file. Crucially, it is cryptographically bound to the exact pixels it was issued for. Edit the image and present it with the old seal, and the seal visibly breaks when checked.

We did not take that on faith. When we built our signing pipeline, we tested it by cheating: we signed an image, changed a single byte, and read the file back. The verification flagged the tamper immediately. And because a seal only vouches for the exact bytes it was issued for, we designed the pipeline to sign the final stored file, after our image conversion has already run, so the certificate matches the file exactly as it will be served.

Anyone can check a credentialed file using free public tools such as the Content Credentials verifier. No account, no Omniops involvement. That is the point: the proof belongs to the file, not to us.

What happens when someone screenshots it?

Days after this post went live, a reader asked exactly that. It deserves a straight answer: no, the seal does not survive a screenshot.

A screenshot is not a copy of the file. Your phone takes a fresh picture of whatever is on screen and saves a brand-new image, with none of the original's metadata inside. The seal is not broken, it is simply absent. Check a screengrab with a verifier and it will not say "tampered", it will say "no credentials found". The same happens when an image passes through apps that strip metadata on upload, which most social and messaging apps do.

That is a limit of the standard itself, not of any one implementation. A Content Credential proves what a file is. It cannot speak for a copy it was never sealed into.

The industry's answer is a second layer: an invisible watermark woven into the pixels themselves when the image is generated. Google's version, SynthID, is the best known. Because it lives in the picture rather than in the file's metadata, Google says it is added when the content is created and designed to stand up to common modifications such as cropping, filters and lossy compression; detection is handled by Google's tools. That layer belongs to the model provider rather than to us, and it arrives with the model: the image engines behind our coming model photography feature include Google's, whose output carries SynthID in the pixels from the moment it is generated. Those images will carry both marks, the provider's watermark in the pixels and our seal in the file.

Underneath both sits our provenance record. Every image Studio generates is logged with the tool, the model and the timestamp, so even when a copy has been stripped bare, we can vouch for the original we served.

One more reassurance for shop owners: the August rules put the marking duty at the point of generation and publication. You label the image you publish. A shopper screenshotting it afterwards does not undo your compliance. The habit that matters is keeping the sealed file as your master copy, so every image you publish carries its own answer.

The part we are most proud of

Here is the design decision that matters more than the technology.

Omniops Studio is growing from product and scene imagery into full creative work, including AI-generated fashion models for shoots. Before building that feature, we wired in a hard rule: it refuses to run unless the certification layer is switched on.

That is not a policy document or a setting someone could forget. It is code, already merged and tested ahead of the feature it protects, and it fails closed. If the signing machinery is missing or misconfigured, the generation simply does not start.

We built the lock before the door.

What is true today

We hold ourselves to saying only what we can show, so here is the current state, plainly:

• Live now: every image generated in Omniops Studio carries a visible "AI-generated" label, and we store a provenance record alongside it recording the tool, the model, and when it was made.
• Built and proven in testing: the cryptographic Content Credentials pipeline, including tamper detection on read-back, in a standalone test rig. It is not yet signing production images, because the feature that needs it has not shipped.
• Coming: AI model photography for fashion brands. When it ships, cryptographic signing is on from the first image, because the feature cannot run without it.

Why this is good for your shop, not just for regulators

It would be easy to read all this as legal housekeeping. We think it is the opposite.

Your customers are getting better at spotting AI imagery, and their default reaction to unlabelled AI is suspicion. A shop that can answer "is that photo real?" with "it's AI-generated, and the image itself will tell you so" is not admitting a weakness. It is demonstrating that it does not need to hide anything, which is the same signal an honest returns policy sends.

Trust compounds. The shops that treat AI transparency as part of their brand, rather than a box to tick, will be the ones whose product pages still feel trustworthy in five years, when nobody can tell real from generated by eye alone.

And the August deadline? If your imagery runs through a system that seals the label in automatically, July stops being a scramble. The box is ticked before you knew there was a box.

You can read exactly what we sign and how to verify it in our documentation, and our wider commitments on the trust page.