Launch weeks have a convention: you talk about what the product does, and the limits go in footnotes nobody reads. We are breaking the convention on purpose. This week we announced certification on every image and opened the file to show what travels inside it. Today we publish the list that usually stays internal: what none of it can do.
The reason is not modesty. In compliance, overclaiming is the product defect. A brand that believes its certification does more than it does will make decisions based on protection it does not have, and discover the gap at the worst possible moment. The vendor who tells you plainly what their product cannot do is the vendor you can believe about what it can. So, in writing, versioned, and staying published: the five limits.
1. Certification is not legal immunity
No marking technology makes an image "compliant" by itself, ours included. Signed provenance and watermarking satisfy one specific duty: the machine-readable marking that European law requires of AI platforms. Compliance as a whole is a posture, not a stamp: correct marking, visible labels where the law demands them, and records you can produce when asked. Our layers are one load-bearing wall of that posture. They are not the building.
2. Our marks do not discharge your labelling duty
If you publish AI imagery that qualifies as a deepfake, and much commercial fashion imagery does, the visible label is your obligation as the publisher, and the European Commission has said explicitly that a provider's machine-readable marks cannot stand in for it. We built the guidance and tooling for that half, but the duty itself cannot be transferred by technology or by contract. Anyone who tells you otherwise is selling you exposure.
3. A determined actor can strip the marks
Our layers are engineered to survive ordinary distribution: re-encoding, resizing, format conversion, the everyday violence platforms inflict on files. They are not engineered to survive deliberate sabotage, because nothing is. Sufficiently aggressive transformation, heavy regeneration by another AI system, or a screenshot of a screenshot of a print can degrade any watermark in existence. What survives sabotage is the export record: our authoritative log of what was created, when, with which marks. Evidence does not have to live in the file to exist.
4. We do not detect unmarked AI content
Our verifier answers one question with authority: does this file carry provenance, and does it verify? It does not answer "was this image, carrying no marks at all, made by an AI somewhere?" Forensic detection of unmarked content is a genuinely unsolved problem; the EU's own Code of Practice declines to require it, noting the technology is not mature. Vendors advertising reliable AI detection of arbitrary unmarked images are ahead of the science. We would rather tell you that than join them.
5. Provenance does not make an image truthful
Our certification proves origin: AI-generated, vouched for by Modelia, at a fixed moment. It does not prove the image is a fair representation of your product. A certified render that flatters a garment beyond what the buyer receives is still a deepfake under the Guidelines and still misleading advertising under consumer law. Render fidelity is an editorial decision that stays in your hands, and no cryptography can make it for you.
What to do with the gaps
Each limit has a corresponding practice, and together they are short: run visible labels in your publishing workflow for deepfake-class content; keep your export records where you can produce them; write down your render-fidelity rule; and treat any vendor claim of universal AI detection or transferable legal duties with the scepticism it has earned.
This page stays published, versioned, and linked from our compliance page. If our capabilities change, this list changes with a visible date. Hold us to it.
Every image exported from Modelia carries signed provenance and an invisible watermark by default, and anyone can check any image, free, no account, in our public C2PA checker. The full picture, rules and industry data, is in The State of AI-Generated Content in Fashion 2026.
