AI Act standard-editing exemption: trim, filter, caption steps
The Commission FAQ exempts AI assisting standard editing from marking. Which plain edit steps Sume exposes (trim, dim, crop, caption), and what to check.
The Commission's Article 50 FAQ says the machine-readable marking obligation "does not apply when the AI system performs an assistive function for standard editing", and that the Guidelines give examples of what is standard editing and what goes beyond. The FAQ page itself does not list the examples, so which edit falls inside is a Guidelines question, not something this post can settle.
What this post can do is name the plain edit steps Sume exposes and what each one does to the file, from the Video trim, Video filter and Video captions docs, read 2026-10-01. Not legal advice.
What does the exemption actually cover?
It is an exemption from the provider's marking duty in Article 50(2) for an AI system doing an assistive editing function. It sits beside other carve-outs the FAQ lists, such as outputs used only in closed-loop industrial or film-production environments unless they are the final output. It does not say that edited content is never a deepfake or never needs a label; those are separate questions under Article 50(4).
Which Sume steps are plain edits?
Three endpoints change an existing clip without generating new people or scenes.
| Step | Endpoint | What the docs say |
|---|---|---|
| Trim | POST /v1/video-trim | Returns a new MP4 holding only [start, end); the source is untouched. |
| Filter (dim, crop) | POST /v1/video-filter | Takes ops[] dim/crop and/or a filters-only filtergraph; returns a new MP4, source untouched. |
| Captions | POST /v1/video-captions | Takes a public HTTPS video URL and returns a job-backed captioned video. |
Does an edit step make a clip AI-free?
No. These steps do not remove whatever made the original. If the source clip was generated, trimming it leaves it a generated clip. The exemption is about the editing function, so the origin of the input still matters when you assess labeling. For the creator-side view on reframing and color work, see reframing, cropping and color correction.
What should I record?
Keep the source URL, the edit request and the output URL together, and note which Guidelines example you relied on. If a step does more than trim, dim, crop or caption, check the Guidelines before assuming it is standard editing.
Sources
Related posts
More in Use cases
- AI Act deepfake disclosure at first exposure: marking is not enough
Per the Commission FAQ, deployers disclose a deepfake at first exposure in a form people can perceive; the provider's machine-readable mark alone is not enough.
- AI Act artistic or satirical deepfakes: a lighter disclosure
For evidently artistic, satirical or fictional works the Commission says disclose in an appropriate manner that does not hamper enjoyment. Credit or banner?
- How long must the EU AI label stay on screen? Cue timing
The Commission gives no seconds: a limited-time disclosure should stay visible long enough to be read by people with cognitive or processing difficulties.
- EU AI icon with a "voices generated with" label on avatar video
The Commission's example pairs the basic EU icon with the text "voices generated with". Here is how to author that line as caption cues on a Sume avatar video.
Written by Sume