General information, not legal advice.
Every fashion brand that publishes AI imagery in the EU now depends on its vendors for half of its legal position. The platforms and studios that generate your images carry the machine-readable marking duty; if they do it badly, the gap surfaces in your published files. And with the last transition window closing on 2 December, this quarter is when procurement teams should be asking the questions.
Here are the ten we would ask, each with what a good answer sounds like and the red flag that should end the meeting early. Send them as they are: the point of a checklist is that it travels.
The ten questions
1. Do your exports carry machine-readable marking by default?
What a good answer sounds like: Yes: every export carries a signed C2PA manifest and an invisible watermark, on every plan, with nothing to switch on.
Red flag: Marking is a premium feature, an API flag, or "on the roadmap". A duty that exists since August cannot live on a roadmap.
2. Can you show one of your files verifying in an independent checker?
What a good answer sounds like: A live demonstration: their export opens in a public C2PA inspector, the signature chains to trusted infrastructure, the timestamp is accepted.
Red flag: A slide about standards instead of a file that verifies. If the certificate is not anchored in the C2PA trust infrastructure, verifiers will reject it, however real the signature.
3. What exactly do your manifests assert, and what do they exclude?
What a good answer sounds like: A precise list: AI-generated statement, system and provider name, trusted timestamp. And explicit exclusions: no prompts, no client identity, no campaign context.
Red flag: They do not know what their own manifests contain, or the manifest leaks workflow data your competitors could read in your published files.
4. What happens to marks after your own pipeline edits?
What a good answer sounds like: They know that compositing and inpainting destroy inherited marks, and they re-mark at final export so every delivered file carries fresh, valid layers.
Red flag: "Our upstream model provider handles that." Inherited marks die in editing; a vendor who has not thought about re-marking is shipping unmarked files and does not know it.
5. Which of your marks survive platform re-encoding, and how do you know?
What a good answer sounds like: Tested, dated results: the metadata layer dies on re-encoding platforms, the watermark survives, and they can show the tests.
Red flag: Survival claims with no test behind them, or the assumption that the manifest alone is enough because "we sign everything".
6. How can we, or a regulator, verify your files independently?
What a good answer sounds like: A free, public verification service anyone can use, which reads both layers and says which one answered, with a clear no-retention policy.
Red flag: Verification only through a support ticket, only for customers, or not at all. If a regulator cannot check a file, your evidence depends on goodwill.
7. What records do you keep of what we generated?
What a good answer sounds like: An export record: file hashes and provenance details for everything delivered, available to you on request, with clear terms on retention and on what happens to your history over time.
Red flag: No records at all, or records they cannot give you. When a stripped copy of your image surfaces somewhere, that log is the authoritative answer to "is this ours, and when was it made?"
8. Do you support our side of the duties: visible labelling?
What a good answer sounds like: Yes: tooling or presets for the official EU icons, placement guidance per channel, and clear documentation that visible labels on deepfake-class content are the publisher's duty, with help doing it.
Red flag: "Our marks make you compliant." They legally cannot: the Commission has said publishers may not rely on the provider's machine marks for the visible-label duty. A vendor claiming otherwise is either uninformed or overselling, and both are your problem later.
9. How do your contracts handle the duty split?
What a good answer sounds like: Terms that prohibit removing marks, and an annex that allocates practical tasks clearly while acknowledging that legal roles are set by facts, not by contract.
Red flag: Contract language that promises to "take on" your legal responsibilities. Roles under the AI Act cannot be reassigned by contract; a promise to do so is either naive or bait.
10. How will you handle change, and will old content stay verifiable?
What a good answer sounds like: A versioning commitment: detector versions maintained, certificates managed, notice of material changes, the February 2027 interoperability deadline on their roadmap by name, and content verified years from now even as systems evolve.
Red flag: No answer about backward compatibility. Content you publish this quarter must still verify in 2029; a vendor without version discipline is accumulating your future problem.
A note on our interest
We are not neutral, and the checklist is better if we say so. Modelia is an AI image platform; we built our answers to these ten questions into the product and published them, including the limits. That is exactly why we know what good answers sound like, and why we are comfortable handing you a list your current vendor might fail. Use it on us too: every claim we make is verifiable in a public checker, which is the standard this checklist holds everyone to.
The scorecard version
For procurement use, the same ten questions are available as a designed pack with a scoring sheet, answer-recording fields and the red-flag summary on one page, ready to attach to an RFP or a vendor review [link to gated pack]. And if the questions surface gaps on your own side of the duties, the visible-label guide and the full legal picture are on our compliance page and in the Article 50 guide.
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.
