Who Owns AI-Generated Work? Your Terms vs Copyright Law

Open OpenAI's Terms of Use and find the sentence about who owns what the model writes:

"you (a) retain your ownership rights in Input and (b) own the Output. We hereby assign to you all our right, title, and interest, if any, in and to Output."

Now open Anthropic's Consumer Terms:

"we assign to you all our right, title, and interest (if any) in Outputs."

And Midjourney's:

"You own all Assets You create with the Services to the fullest extent possible under applicable law."

Three companies, three lawyers, the same escape hatch. They are assigning you whatever rights they happen to have — and quietly declining to promise there are any.

That hedge is the entire subject of this article, because it sits exactly on top of the thing that decides whether you can sell what you just made.

The two questions everyone runs together

When a solopreneur asks "do I own this?", they're actually asking two separate questions with two different answers.

1. Does the vendor permit me to use it commercially? Yes. All of them. OpenAI, Anthropic and Midjourney all assign you their interest in the output, and none of them claim a licence back for your commercial use. On the contract, you are fine.

2. Do I own a copyright in it? Often no. And copyright is what lets you stop someone else from using it, and what you're promising a client when you sign an IP assignment.

A vendor's terms cannot grant what the law does not recognise. That's what "if any" is doing in those sentences. As one practitioner analysis puts it: if the output isn't copyrightable, there are no rights to assign.

Almost every article on this topic answers question one and presents it as the answer to question two.

What the law actually says, as of today

Two things settled recently, and one thing is still open.

Prompts alone are not authorship

The US Copyright Office published "Copyright and Artificial Intelligence, Part 2: Copyrightability" on 29 January 2025. Its central conclusion, in the Office's own words:

"prompts alone do not provide sufficient human control to make users of an AI system the authors of the output."

The reasoning is mechanical rather than philosophical. AI systems don't reliably follow instructions; they fill in the gaps a prompt leaves, and they generate different outputs from identical prompts. If the same instruction produces different expression each time, the instruction isn't controlling the expression.

Iterative prompting doesn't rescue it either. The Office treats "I prompted it two hundred times" as a version of the sweat of the brow argument copyright rejected long ago. Effort isn't authorship — which is worth holding in mind next to the prompt libraries everyone keeps, ours included. A better prompt gets you a better draft. It does not get you an author credit.

The Office's other conclusions matter as much: existing law is sufficient without new legislation; using AI to assist human creativity doesn't affect protection; human expression inside an AI-generated work is protected; and there should be no special new rights for machine output.

Human authorship is now settled law

In Thaler v. Perlmutter (D.C. Circuit, 18 March 2025), the court held that human authorship is "a bedrock requirement to register a copyright" and that an AI system "cannot be deemed the author of a work for copyright purposes" — required, the court said, "as a matter of statutory law."

The reasoning is textual and rather elegant: the Copyright Act only functions if authors are human. Duration is measured by an author's lifespan. Copyrights are inherited. Transfers require a signature. None of that works for a machine.

The Supreme Court declined to review it on 2 March 2026. That question is closed.

What is not settled: how much human is enough

Thaler answered "can a machine be the author?" It expressly did not answer "how much AI assistance is too much?" The court called those separate line-drawing issues and left them alone.

That line is being drawn right now, in Allen v. Perlmutter (District of Colorado, filed September 2024, still pending with cross-motions for summary judgment as of January 2026). Jason Allen's Théâtre D'opéra Spatial won an art competition. He then tried to register it, disclosing that he had used at least 624 text prompts, plus Midjourney's variation and upscaling tools, plus Photoshop cleanup, plus a second upscale.

The Copyright Office refused registration. Allen argues that Feist requires only a "minimal level of creativity" and that photographers got copyright despite using mechanical devices. The Office's Review Board countered that when AI produces a complex work "solely from a human prompt", the traditional elements of authorship are executed by the technology, not the user.

624 prompts and Photoshop was not enough for the Copyright Office. Whatever you did this morning was probably less.

This is the same shape of question as whether AI can replace an assistant: the interesting line is never whether the tool produces something usable, it is how much of the result is genuinely yours.

The trap that actually costs solopreneurs money

Here's where this stops being academic.

You finish a project. Your contract says the client owns all intellectual property in the deliverables — standard language, in nearly every freelance and agency agreement ever signed. You used AI for part of it.

You cannot assign rights that don't exist.

That's the phrase a practitioner writing on AI contract clauses uses, and it's the clearest statement of the problem we found anywhere. If parts of your deliverable aren't copyrightable by anyone, you've promised your client something you can't hand over. Not because you did anything wrong — because the promise was impossible when you signed it.

Practitioners advising freelancers on this suggest the fix is in the drafting rather than in the work: assign what's assignable and licence what isn't, disclose AI use rather than leaving the contract silent, qualify the originality warranty, and include fallback licensing so the client can still use the deliverable even if outright ownership fails.

Whether that's right for your contract is a question for a lawyer, not for us. What we can tell you is that the risk is real, it's specific, and silence doesn't make it go away — as one of those practitioners puts it, "when your contract is silent, every assumption rushes in to fill the gap, and assumptions are where disputes are born."

The demands are already arriving from the other direction too. Freelancers report clients asking for disclosure of which tools were used and how; outright AI bans, or AI restricted to brainstorming; full IP rights "regardless of how they were created"; originality warranties despite nobody knowing what's in the training data; and indemnity clauses that put the freelancer personally on the hook if a claim lands.

The Authors Guild has gone as far as publishing model contract language. Its clause reads:

"Author shall not be required to use generative AI or to work from AI-generated text. Author shall disclose to Publisher if any AI-generated text is included in the submitted manuscript, and may not include more than [a de minimis/5%] AI-generated text."

That's a professional body writing a numeric cap into a standard contract. It tells you where this is heading.

The indemnity that doesn't cover the plan you're on

If AI output does infringe someone's copyright, who pays?

Several vendors offer indemnification. Read which tier it applies to, because the pattern is consistent and it isn't in your favour.

OpenAI's announcement was explicit about scope: "we can defend our customers and pay the costs incurred if you face legal claims around copyright infringement, and this applies both to ChatGPT Enterprise and the API."

The $20/month plan is the uncovered one. That is the plan most solopreneurs are on, and it's the one every "is AI content safe to use commercially" article forgets to mention when it tells you Copyright Shield exists.

Microsoft's commitment is worth reading closely too, because its conditions are strict. You must not have "disabled, evaded, disrupted, or interfered with the content filters, restrictions in Metaprompts, or other safety systems." You must have sufficient rights in your input. And there's a carve-out that lands exactly where a small business would trip:

Trademark claims are excluded. Microsoft's own worked example is a Copilot-generated logo that inadvertently resembles an existing trademarked one. That's not a hypothetical for a solopreneur — that's Tuesday.

We looked at Adobe Firefly's indemnification too and can't tell you what it covers. Sources disagree on whether it's an enterprise-only entitlement or included with all paid Creative Cloud plans, and Adobe's own FAQ wouldn't load. Rather than pick the flattering version, we're telling you it's unresolved — check your own agreement.

What does earn you a copyright

Not everything is lost. The Copyright Office has been clear about what works, and it's consistent across the report and the registrations it has actually granted.

Creative selection, coordination and arrangement of AI outputs. Zarya of the Dawn — a comic with Midjourney images — was registered for the human-authored text, selection and arrangement, and not for the images.

Human-authored material perceptible in the output. In Rose Enigma, a human drawing was used as the input. The registration covers the "perceptible human expression"; the applicant had to disclaim the AI-generated elements such as the realistic rendering and lighting.

Creative modification of AI output. Substantially reworking what the model produced, rather than picking the best of four.

And what doesn't work, explicitly: "selection of a single output is not itself a creative act." Generating four options and choosing one is not authorship.

The Office is emphatic that this is case-by-case and advises applicants to file a detailed statement of their human contribution. Legal analysts reading the report converge on the same practical advice: disclose and disclaim — identify which parts are yours and say plainly which parts the machine made.

That has a workflow implication worth taking seriously. If protection depends on demonstrating human contribution, the record of that contribution is worth keeping — prompts, raw outputs, and what you changed. We already argue for documenting process in building a content calendar with ChatGPT; this is a second, sharper reason to do it.

The market has already made its decision

You don't have to take a position on any of this to notice how the commercial world has priced it.

Getty Images banned AI-generated submissions. Its CEO cited "open questions with respect to the copyright of outputs from these models" and "unaddressed rights issues with respect to the underlying imagery and metadata used to train these models."

Shutterstock doesn't accept AI-generated content from contributors — while selling its own AI image generator. Its stated reason is blunt:

"Since AI content generation models leverage the IP of many artists and their content, AI-generated content ownership cannot be assigned to an individual."

Two companies whose entire business is licensing image rights have concluded they cannot verify ownership of AI-generated images. If they can't, the question of what you're transferring when you invoice a client is not a pedantic one.

The counter-argument, which deserves a hearing

The strongest case that all of this is overblown comes from the Kluwer Copyright Blog, which called Thaler "an easy case making good law" and argued the human authorship requirement "does not impede the protection of works made with artificial intelligence."

Its reasoning: in practice, commercial operators simply name "the person who created, operated, or used artificial intelligence" as the author, and get on with it. On that reading the requirement is largely ceremonial — copyright keeps its humanistic symbolism while accommodating AI production without real friction.

We think that's half right, and the half it gets wrong is the half that matters to you.

It's correct that naming a human author on a registration form is trivial. It is not correct that the requirement is therefore ceremonial — because the Copyright Office refused Jason Allen's registration after 624 prompts, and that refusal is being litigated right now. A requirement that produces a live federal case over a real artist's real work is doing something.

The gap between the two views is roughly the gap between a large company with lawyers and a one-person business signing a standard IP assignment. If you're the second one, ceremonial isn't the word.

What this means if you're invoicing next week

Stated as plainly as we can, and with the limits marked.

What is reasonably clear:

• The vendors' terms let you use output commercially. That part is genuinely settled
• Prompting alone doesn't make you the author of the result, per the Copyright Office
• A machine can't be an author. That's final after March 2026
• Your creative arrangement, modification and perceptible human contribution are protected
• If you're on ChatGPT Plus, no copyright indemnity applies to you

What is genuinely unresolved:

• How much human contribution is enough. Allen may tell us; it hasn't yet
• Whether Adobe's indemnification reaches individual paid plans
• What a court would make of a client suing over an unenforceable IP assignment. We found no case on this and we're not going to invent a prediction

Where you need an actual lawyer, not an article:

• Before signing an IP assignment for work you know is AI-assisted
• If a client asks you to warrant that deliverables are original and non-infringing
• If you receive an infringement claim over anything AI touched
• Any registration where the human/AI boundary is genuinely arguable

We're describing documents, not advising you. This is a good moment to say that we're not lawyers and this isn't legal advice — the difference between the two is exactly the difference between "here's what the Copyright Office published" and "here's what you should do", and we're only doing the first.

If you want the practical companions to this: what AI is genuinely good at in a solo business, what the whole stack costs, whether paid writing tools earn their keep in our Jasper review, and how to use AI on client-facing documents without handing over your judgement in writing business proposals. Two liability parallels are worth reading alongside this one: you own what your automation says to customers, which is AI customer support, and automated sending puts you under a separate set of published rules, which is AI email automation.

---

Sources, all read on 27 August 2026: the US Copyright Office's announcement and reporting of "Copyright and Artificial Intelligence, Part 2: Copyrightability" (29 January 2025); Thaler v. Perlmutter, No. 23-5233 (D.C. Cir., 18 March 2025), certiorari denied 2 March 2026; Allen v. Perlmutter (D. Colo., No. 1:24-cv-2665), pending; Anthropic's Consumer Terms and Microsoft's Customer Copyright Commitment, read directly; OpenAI's and Midjourney's terms as quoted consistently across legal analyses, since both sites refused automated access; the Authors Guild's published model clauses; Shutterstock's contributor policy; and practitioner commentary from law firms and freelance-contract specialists. Vendor terms are revised frequently — every quotation above is dated to the day we read it, and this page is flagged for a terms re-check in February 2027. Where sources conflicted and we could not resolve them, as with Adobe Firefly's indemnification, we have said so rather than picked one. Some links may be affiliate links — read our policy.