Most conversations about disclosing generated model imagery are conversations about wording. What should the label say, where should it appear, how will customers read it.
Those questions are answerable. The question that blocks them is earlier and duller: which images are we talking about, and does anyone know?
The term covers too much to act on
“AI-generated” behaves as though it names a category, and in a fashion image pipeline it names a range. At one end is a model who never existed, wearing a garment that was never photographed. At the other is a photograph of a person, lit by a real light, with nothing done to it. Between those two ends sits most of the actual work.
A partial list of what lives in the middle:
• A photographed model wearing a garment that was substituted after the shoot.
• A photographed model and garment, placed against a background that was generated.
• A photograph in which something was removed and the resulting gap filled in by a model rather than by a retoucher.
• A neckline, a lining, or a brand label that was inferred rather than captured.
• A second colorway derived from the image of the first, so that only one of the two garments was ever in front of a camera.
• A set of frames in which one face was carried across several bodies or several passes.
Nobody has drawn a line through that list, and the line cannot be inherited from a definition of the term, because the term does not contain one. It has to be drawn by the business, written down, and applied consistently — and the difference in risk between the entries above is real rather than a matter of degree, which is the subject of two edits at two risk levels.
You can only disclose what someone recorded
Disclosure is treated as a communications capability and operates as a records capability. To label an image, you have to be able to say what was done to it, and that statement has to come from somewhere.
In most asset libraries it comes from nowhere. There is no field for it. A retoucher’s output and a generated output arrive in the same folder, under the same naming convention, at the same stage, and become indistinguishable within a week. Six months later the question “was this model photographed” has no holder — the person who would know has moved to another season, and the file carries no trace of the answer.
There is a reason the field does not exist, and it is not negligence. Asset systems were built when the questions asked of an image were which version it was and what it was licensed for. How the pixels came to be was not a question anyone needed answered, because there was only one way for them to come to be. Adding a field costs an afternoon. Deciding what values that field can take is the actual work, and it is the same work as drawing the line in the previous section — which is why the two problems have to be solved in that order and usually get attempted in the reverse one.
The timing is the whole of it. Provenance written at the moment of creation, by the person creating it, costs close to nothing: it is one field, filled in by someone who already knows. Reconstructed afterward it is frequently impossible, and what replaces it is inference from the image itself, which is exactly the evidence that cannot be trusted for this purpose.
Where the unrecorded parts usually are
The fully generated model is not the hard case. Everyone knows when they have commissioned one; it is a decision with a budget line and a brief.
The hard cases are small and arrived at sideways. A background swapped because the location fell through. A distracting object taken out, leaving a hole that something had to fill — and what fills it is generated, whatever the tool is called: what replaces a removed object. An interior neckline completed because the original frame never captured it, which in the worst instance produces a brand label on a product page that the brand did not put there: what gets filled in at the neckline.
None of these were flagged, because each one felt like retouching to the person doing it, and retouching has never required a note. The category shifted underneath a workflow that had no reason to notice. This is the practical content of a provenance problem: not a hidden decision, but a series of unremarkable ones that nobody had a place to write down.
Being found out sets the timing and the framing
The reason to do this before it is asked for has little to do with virtue and a lot to do with control.
Discovery, when it happens, is usually social rather than procedural. A customer notices that a hand has the wrong number of fingers, or that the same face appears across two unrelated brands, or that a garment in one image does not match the garment that arrived. A competitor points at something. An employee mentions it somewhere public. None of those events are scheduled by the brand, and none of them come with an opportunity to prepare.
The friendly version of the same event is more common and gets less attention. A buyer asks whether the imagery for a style is photography. A review of an unrelated matter turns up a question about one campaign. A new hire assumes the record exists and asks to see it. Nobody is accusing anyone, and the answer is still that nobody can say. The record earns its cost in those moments too, and they arrive much earlier than the public version.
At that moment the brand answers a question about specific images on somebody else’s timeline. If the record exists, the answer is short and boring, which is the best available outcome. If it does not, the answer is that nobody can say — and an absence of records reads externally as concealment whether or not anything was concealed. “We don’t know” is the worst answer available and it is very often the true one.
What varies by market and what does not
Requirements around disclosing generated imagery differ between markets and between selling platforms, they are not uniform, and they change. Nothing in this article should be read as establishing what applies to a given business. That determination has to be made market by market and platform by platform, against current requirements, and routed through whoever owns compliance in the organization.
What does not vary is the prerequisite. Every possible version of a requirement — strict, loose, per-market, platform-specific, or none at all — depends on the business being able to say which of its images are which. A brand that has the record can comply with whatever turns out to apply, adjust when it changes, and answer a question from any direction. A brand without the record cannot comply with a requirement it agrees with, which is a worse position than disagreeing with one.
That asymmetry is why the record is the thing to build now, before the policy question is settled. It is useful under every outcome, and it is the only part of this that gets harder with delay.
Where a disclosure sits, and what it says
Placement decides whether a disclosure informs anyone. A note attached to the image, or immediately beside it, reaches the person at the moment the question occurred to them. The same words on a policy page reach people who were not asking. Both may satisfy a requirement; only one of them does the thing disclosure is for.
Wording carries its own exposure, in a direction that is easy to miss. A label claiming more than the business can support — that garments shown are physically accurate, that a depiction represents an actual product state — creates a second statement that has to be true. The safer construction describes what was done to the image rather than characterizing what the image proves, and a claim about the product belongs in copy where it can be checked against facts someone has confirmed.
Consistency does more work than either. A disclosure applied to some images in a set and not to others says something definite about the unlabeled ones, whether or not that was intended, and it will be read that way by anyone who notices the pattern. Partial labeling is therefore not a partial version of disclosing — it is a claim about everything left unlabeled, made silently, by omission. A rule applied to every image in a set is the only version that means what it appears to mean.
Both of those are decisions to make once, write down, and apply. Neither can be made at all until the first question is settled, which returns to the same place: someone has to know which images are which.
Questions teams ask about disclosure
Which of our images actually need to be disclosed? That is a question about the requirements applying to your markets and selling platforms, and it needs to be answered against current sources by whoever owns compliance. What can be settled internally, immediately, is which images fall into which category. Doing the second makes the first answerable instead of theoretical.
Is a retouched photograph the same as a generated image? They sit at different points on a range rather than in two boxes, and the range has no natural break in it. The business has to decide where its own line falls and record that decision in a form other people can apply. A line that lives in one person’s judgment will move without anyone noticing.
When should provenance be recorded? At creation, by whoever made the image, as part of delivering it. That is the only point at which the information exists for free. Every later point requires someone to reconstruct it from an image, which is unreliable in exactly the cases that matter most.
We have already published images we cannot characterize. What now? Stop the accumulation first, because the set of uncharacterized images grows every week the field does not exist. Then treat the existing library as a known gap and say so internally rather than guessing retroactively. A documented gap can be managed; a confident reconstruction that turns out wrong is a new problem on top of the original one.
Does disclosing hurt conversion? Your pages are the only place that can answer that, and the test is available to you — the same label, the same placement, measured against your own traffic. General claims in either direction are not measurements of your customers. Run it on your own pages and use your own numbers.
Who owns this? Two roles, and the usual failure is that neither is named. Someone has to own the policy question, which is a compliance responsibility. Someone else has to own the asset record, which is an operations responsibility. Disclosure fails in the gap between them far more often than it fails on either side.
Where this leaves you
Disclosure reads as a messaging problem and behaves as a bookkeeping one. The wording can be drafted in an afternoon; the ability to say which images the wording applies to takes a field, a habit, and someone responsible for both. Build the record before deciding the policy, because the record is useful under every version of the policy and is the only part that becomes unrecoverable with time. Then take the line the business is prepared to defend, write it down, and let the requirement questions be answered by the people whose job that is.
Written by
