What 'you own the output' actually means
Vendor terms can assign you an output file. They cannot make it copyrightable. How to separate contractual ownership from protectability before you ship.
Every generator's marketing page says some version of "you own what you create," and every one of them is telling the truth about something narrower than the sentence implies. The clause is real. It just answers a different question from the one most people are asking when they read it.
There are two questions, and collapsing them is where teams get into trouble:
- Between me and the vendor, who holds the rights to this file? That is a contract question, and the answer is usually you.
- Does anyone hold copyright in this file at all? That is a statutory question, and the vendor has no say in it whatsoever.
A licence can settle the first. Nothing in a licence can settle the second. This is not a drafting oversight or a loophole a better vendor could close. It is a structural limit, and knowing where it sits changes what you promise a client.
A provider can only assign what it holds
The first hard limit is mundane contract law. An assignment transfers rights the assignor actually possesses. If the provider holds no copyright in a generated image, the clause that assigns you "all right, title and interest" in that image transfers a set that may be empty.
That is not a reason to distrust the clause. It is a reason to read it as what it is: a promise that the provider will not come after you, plus a waiver of any claim it might otherwise have. Both are genuinely useful. Neither creates a property right where the law says none exists.
The clauses do vary in shape. OpenAI's Terms of Use assign OpenAI's right, title and interest in Output to the user, with commercial use permitted across plans. Midjourney gives paid subscribers ownership of their assets, but subject to revenue-threshold tiering and a broad licence back to Midjourney, and its consumer-plan terms disclaim IP warranties outright. Those are meaningfully different documents that produce the same headline, and the difference only surfaces when a deliverable is already out the door. Read the actual clause for the actual plan you are on, not the summary on the pricing page. Open-weight models are a separate case again, because there the licence travels with the weights rather than with your account.
Where US law draws the line
Human authorship is the requirement, and it is settled at the level that matters for planning.
In Thaler v. Perlmutter, the applicant identified a machine as the author of the work and sought registration on that basis. Registration was refused, the D.C. Circuit upheld the refusal, and the Supreme Court denied certiorari on 2 March 2026 (No. 25-449), leaving that ruling in place. The case is about a work with no claimed human author at all, so it is the outer boundary rather than a ruling on ordinary prompt-based work.
The line for ordinary work comes from the Copyright Office. Its AI guidance — the registration guidance issued 16 March 2023 and the Part 2 report published 29 January 2025 — states the position plainly:
- Prompts alone do not make you an author, however long, detailed or many times iterated. The Office's reasoning is that the prompt does not control the expressive elements the model produces.
- Human-authored selection, arrangement and modification are protectable. If you generate forty clips and cut them into a sequence, the sequence is yours even where the clips are not.
- AI-assisted works are registrable, with disclosure of the AI-generated material and a claim limited to the human contribution.
The practical consequence is uncomfortable if you have been selling raw generations as owned assets: a single-prompt output with no further human authorship is, on the Office's stated position, not registrable by anyone. Nobody holds copyright in it. A competitor who lifts it is not infringing, because there is nothing to infringe.
That is survivable for most marketing work, where the asset's value is in being on-brand and on-schedule rather than in being defensible. It is not survivable for a logo, a recurring character, a packaging illustration or anything you intend to license onward.
The rest of the world has not converged
Do not assume the US answer travels.
| Jurisdiction | Position as of August 2026 |
|---|---|
| US | Human authorship required. Prompts alone insufficient. Selection, arrangement and modification protectable. |
| EU | No harmonised AI-authorship rule. The CJEU's "author's own intellectual creation" standard governs and has not been tested against generative AI. |
| UK | The government abandoned its proposed broad text-and-data-mining exception with rightsholder opt-out in a report published 18 March 2026, keeping the status quo. CDPA s.9(3) on computer-generated works still exists, but its application to generative AI is contested. |
| China | The Beijing Internet Court found an AI image copyrightable in Li v. Liu (November 2023) on the basis of sufficient user input. First-instance decision, not binding nationally. |
For a campaign running in several territories, that means the same file can have a different protection status in each one. It rarely changes what you can do with the asset. It changes what you can promise about stopping someone else from using it.
How to build a protectable asset on purpose
If a deliverable needs to be defensible, the work is to create and record human authorship rather than to find a better clause. Four things do most of the load:
1. Do real editorial work, and keep the evidence of it. Generate wide, then select, sequence, retime, colour, composite and revise. The editor is edit-decision-list based, so the timeline itself is a durable record of human selection and arrangement rather than a flattened export with no history behind it. That record is the thing a registration claim describes.
2. Separate the generated layer from the authored layer. Keep the raw generations, keep the composite, keep them distinguishable. When you disclose AI material on a registration, you need to be able to say which parts are which without reconstructing it from memory. The argument for keeping generated work in editable layers rather than baked frames is made in full in layer separation.
3. Log the generation, not just the file. Model, version, date, prompt, account, plan. This is the same record that answers clearance questions later, and it costs nothing to capture at the moment of generation and a great deal to reconstruct afterwards. The operational version of this sits in AI content governance for brands.
4. Say the true thing in the contract. If you deliver AI-assisted work, the assignment you sign should describe what you can actually convey. Promising a client exclusive worldwide copyright in a raw generation is a promise that may be unperformable. The usage rights framing is the vocabulary to use, and the practical drafting patterns are in legal and licensing basics for AI business content.
None of this is legal advice, and none of it substitutes for counsel on a specific deal. It is the operational shape that keeps counsel from having to do archaeology on a launch you already shipped.
The second limit nobody reads
Ownership between you and the vendor is only half the exposure. A licence to use the output does not immunise you against a third party who says the output reproduces their protected work. The studios suing over generated images are targeting outputs, not only training, and the user-side question that raises is genuinely unresolved. That third party never signed your terms and is not bound by them, which is why the risk register needs a second column entirely. The general shape of it is mapped in the AI copyright and safety guide.
Versely's own terms are explicit about the shape of this: generated content may be used for personal and commercial purposes subject to the applicable third-party AI provider terms, and you remain responsible for ensuring your use complies with applicable law. That is the honest structure of every reseller and aggregator agreement, whether or not it says so as plainly.
FAQ
If nobody owns the copyright in a raw generation, can a client still buy it?
Yes. You can sell the file, the exclusivity of your delivery, the right to use it, and the work you did producing it. What you cannot reliably sell is the ability to stop a stranger from using an identical output. Price and describe the deliverable accordingly, and if exclusivity is the point, build the human authorship that supports it.
Does a longer, more detailed prompt make me the author?
No, and this is the single most common misreading. The Copyright Office's stated position is that prompts do not make the user an author regardless of length or iteration count, because the prompt does not control the model's expressive choices. Authorship comes from what you do with the outputs, not from how you asked for them.
Do I have to disclose AI use when registering?
Under the Office's guidance, yes for the AI-generated material, with the claim limited to the human contribution. This is separate from platform and regulatory disclosure, which is a different obligation with different triggers. The audience-facing side of that is in writing an AI disclosure line nobody scrolls past.
Which assets actually need this treatment?
The ones with a long life or a licensing tail: brand marks, recurring characters, packaging art, anything sublicensed to a partner. Ordinary campaign assets that run for six weeks and retire do not need a registration strategy. Deciding which bucket a deliverable is in, before you make it, is the whole discipline.