Selection and arrangement copyright for AI edits
Individual generated clips carry no copyright, but the cut you built from them can. How to structure an edit so the human decisions carry the claim.
Twelve generated clips, none of them protectable on their own, assembled into a ninety-second film that probably is. That is not a loophole. Selection and arrangement is the oldest recognised form of authorship in compilation works, and right now it is the only route to a defensible claim over an AI-heavy video that survives contact with a registrar.
The mistake almost everyone makes is trying to make the generations theirs. That door is shut. The one that is open leads through the edit.
This is a working map, not legal advice.
The rule people skip past
Two things settled the US position, and they point in the same direction.
First, human authorship is required. The Supreme Court denied certiorari in Thaler v. Perlmutter on 2 March 2026, leaving the D.C. Circuit's ruling in place. A work with no human author is not registrable, full stop.
Second, and more useful, the Copyright Office's Part 2 report of 29 January 2025 drew the line inside AI-assisted work rather than around it. Prompts alone do not make the user an author, however detailed or iterated the prompting was. But human-authored selection, arrangement and modification are protectable, and AI-assisted works are registrable provided the AI-generated material is disclosed. The registration guidance of 16 March 2023 set up the disclosure mechanics.
Read those together and the operating conclusion is blunt: the clip is not the work. The cut is the work.
One more distinction worth internalising, because it trips up people who have read the terms of service and assumed they were done. Owning the output file under a contract and holding copyright in the output are different things. A provider can only assign what it holds, and terms cannot manufacture copyright where the human-authorship test fails. "You own your outputs" and "your outputs are protectable" are not the same sentence.
The claim is real, and the claim is thin
Compilation-style protection covers the structure you contributed, not the material underneath it. In practice that means:
- A competitor who lifts your clip 4 and drops it into their own video is not infringing your compilation, because the clip was never yours.
- A competitor who reproduces your shot order, your cut points, your pairings and your pacing is much closer to the line, because that is the part you made.
The strategic consequence follows directly. The more of the finished piece's appeal that lives in the arrangement, the more the claim is worth. A montage where the whole effect is "these six shots, in this order, at these lengths, against this track" has a meaningful claim. A single ten-second generation with a logo stinger bolted on the end does not, whatever the invoice says. The same logic applies to still work, which what makes an AI print design legally protectable covers on the image side.
Five decisions that build the record
None of these are ceremonial. Each one is a human creative choice applied to a particular work, which is the shape authorship takes.
1. Select from a surplus. Generate more than you need and reject most of it. Choosing eight shots out of forty is selection in the doctrinal sense; using the eight you happened to make is not much of anything. This is also just how good edits get made, and reroll rates and budgeting credits is the cost side of the same habit.
2. Decide order and juxtaposition. Which shot follows which, and why. Two clips that mean nothing apart can mean something adjacent. That meaning is yours.
3. Set duration and rhythm. Frame-exact in and out points, held beats, hard cuts against the music. On a 25 fps timeline a shot trimmed to 37 frames is a decision, and it is recorded as one. If your source clips arrived at another rate, 24 fps clips on a 25 fps timeline explains what the conform is doing to them.
4. Modify individual shots. Retime, reframe, mask, speed-ramp, grade. Modification is named explicitly in the Part 2 report alongside selection and arrangement, and it is the one people most often skip because regenerating feels easier. Regenerating produces a different machine output. Editing the one in front of you is a judgement about that particular work.
5. Add layers the model did not make. Typography you set, captions you wrote, sound design you built, your own footage, your own stills. Every non-generated layer is unambiguously human-authored and thickens the claim around everything it sits on. The effort-visible edit makes the audience-facing case for the same moves.
Why the timeline is better evidence than the export
A flat MP4 is a bad witness. It renders your contribution and the model's into identical pixels, and nobody examining it later can tell which decisions were yours.
An EDL-based video editor is a much better one, because the timeline is the list of decisions. Every clip reference, in point, out point, order position, layer and effect is a row. The structure exists as data independent of the footage that fills it, which is exactly what a selection-and-arrangement claim asserts you contributed. It also means the structure survives a footage swap: the pattern in reusable editor drafts is the same asset doing double duty as a production template and a record.
There is a practical reason this is affordable to do properly. Rendering the timeline at 480p with preview: true costs no credits, subject to a short per-user cooldown, so cutting the same material six ways to find the right structure does not multiply the bill. Only the final export is charged, once, regardless of how many clips are on the timeline. Structure-hunting is the activity that generates the authorship record, and free 480p previews are why it is not a budget decision.
What to write down at delivery
Keep it to one block per finished edit. The fields that matter:
| Field | What goes in it |
|---|---|
| Pool | Shots generated vs shots used (e.g. 41 generated, 9 used) |
| Structure note | One or two lines on why this order and these lengths |
| Modifications | Per-shot list: retimed, masked, reframed, graded, speed-ramped |
| Human-authored layers | Type, captions, sound design, original footage, stills |
| Generated material | Which elements are model output, for disclosure |
| Timeline reference | The saved EDL revision the export came from |
The last row is doing more work than it looks like. If you ever need to describe your contribution precisely, the timeline revision is the description, and pointing at it beats reconstructing the edit from memory eighteen months later. When work leaves your studio, handing generated clips to a human editor cleanly covers the same handoff from the other side.
FAQ
Does a longer, more detailed prompt improve my claim?
No. The Part 2 report is explicit that prompts alone, however detailed or iterated, do not make the user an author. Detailed prompting improves the footage you have to work with, which is worth doing for entirely separate reasons. It does not move the authorship line.
If the individual clips aren't protectable, can anyone just reuse them?
As a copyright matter, a purely generated clip in the US is not owned by anyone, so reusing one standing alone is not infringement of your work. Other constraints still apply independently: the model provider's terms, any recognisable likeness in frame, and any trademark or protected character depicted. What a copyist cannot freely take is the arrangement, which is the whole reason to put weight there.
Does this reasoning work outside the United States?
It is the clearest in the US because the Copyright Office has said so directly. In the EU the CJEU's "author's own intellectual creation" standard governs and has not been tested against generative AI, and there is no harmonised AI-authorship rule. The UK has a computer-generated works provision on the books whose application to generative AI is contested and untested. Treat non-US positions as open, and treat the human-contribution record as useful in every jurisdiction regardless, because it is the evidence any of them would want.
Do I have to keep every rejected take?
No. A count and a list of which generation IDs made the final cut is enough to evidence selection. Storing forty discarded clips forever costs you storage and buys you very little; storing the fact that there were forty costs nothing.