Updated September 2026. Reading time: about 6 minutes. General information, not legal advice.
Most of what we have published in this series delivers uncomfortable news: wide definitions, duties with no transition period, exemptions that do not apply. This piece is the other kind. The EU AI Act contains a genuine, usable exemption for everyday image work, and knowing exactly where it ends is worth real money, because it is the difference between marking everything you touch and marking only what the law actually covers.
The core idea: the marking duty is decided per operation, not per tool. Nothing becomes reportable because AI was somewhere in the software. What matters is what was done to the image.
The principle: assistive editing is exempt
The Commission Guidelines carve out AI that performs an assistive function for standard editing: work that does not substantially alter the content or what it communicates. The same photo, cleaned up, is still the same photo in the eyes of the law. The duty begins where the image starts saying something it did not say before: a scene that was never there, a garment on a body that never wore it, a person who never stood in front of a camera.
This is why asking "is our tool AI?" is the wrong question, and "which operations did we run?" is the right one. A single platform session can contain five exempt operations and one that triggers the duty.
The triage, operation by operation
| Operation | Marking needed? | Why |
|---|---|---|
| Colour correction, cropping, compression, dust or red-eye removal | No | Listed as standard editing in the Commission Guidelines. |
| Removing or hiding a background; blurring faces | No | Listed as not triggering the duty. |
| Upscaling that adds no new detail | Probably not | Content is preserved. Keep a record of the assessment. |
| Generating a new background or scene | Yes | New content that changes what the image communicates. |
| Garment swap, virtual try-on, inserting or removing objects or people | Yes | Listed as substantial alteration. |
| A fully synthetic model wearing the product | Yes | Generated content, clearly in scope. |
| Short product text (under about 200 tokens) | Minimal | Free text cannot carry metadata, and short text is exempt from watermarking under the Code of Practice. |
The line, in one before-and-after
Take one product photo and run it through two workflows. Workflow one: colour-correct it, crop it, remove the background, place it on white. Every operation preserved what the image says: this garment, photographed. No marking duty was triggered.
Workflow two: same photo, but the background is not removed, it is replaced with a generated Mediterranean terrace, and the garment is swapped onto a model who never wore it. The image now communicates things that never happened. Machine-readable marks are required in the file, and if the result reads as authentic photography of a person, the visible-label question from our deepfake classification opens on top.
Same source photo, same afternoon, opposite legal outcomes. The operation is the unit of analysis.
The grey zone worth documenting: upscaling
Upscaling sits at the boundary because good upscalers technically synthesise pixels. The Guidelines' logic points to exemption when no new content is added: the image says nothing it did not say before, only more sharply. The defensible practice is to treat it as exempt and write down the assessment: which upscaler, what it does, why you concluded content is preserved. A one-paragraph record turns a judgment call into a documented position, and documented positions are what answer letters.
The exemption people want, and cannot have
Somewhere in every compliance discussion, someone finds the business-to-business exemption and hope flickers: we sell to businesses, not consumers, so perhaps none of this applies. The exemption is real, but it has three cumulative conditions, and the first is that the outputs are not intended to reach the public. E-commerce photos, campaign visuals and lookbooks exist for exactly one purpose: publication. The exemption fails for fashion by design, not by accident. Any vendor or adviser who waves it at you has misread the conditions, and that is worth knowing about the vendor.
Map your pipeline in an afternoon
- 1. List the operations your imagery actually goes through, from raw capture or generation to final export. Most fashion pipelines contain eight to twelve, and the list is stable.
- 2. Sort each against the table above: exempt, marked, or marked plus visible label. The result is a one-page map of exactly where your duties live.
- 3. Record the borderline calls, upscaling above all. The map plus the recorded calls is most of the one-page answer we keep coming back to in this series.
How the marks themselves work is covered in our C2PA explainer, and the two-disclosure split that sits above all of this is here in our machine-readable marks vs visible labels breakdown.
