Who Really Owns Your AI-Generated Content?
Three major legal systems have looked at the same kind of picture — a Stable Diffusion output, built from a hand-tuned prompt — and reached three incompatible answers about who owns it. That contradiction isn’t a bug regulators will patch. It’s the actual state of the law in 2026, and it means the platform’s terms of service and the copyright office’s opinion are answering two different questions.
I used to give people a clean, wrong answer to this question. Somewhere around 2023 I told a client, flatly, that AI-generated images were automatically public domain because “a machine can’t hold copyright.” That’s true — and it’s also a half-truth that gets people in trouble, because it skips the actual mechanism: it’s not that AI output is declared public domain, it’s that in the U.S. it usually never qualifies for copyright in the first place unless a human’s own creative fingerprints are on it. Those are different legal categories with very different consequences, and the gap between them is where most of the real risk sits.
- Two separate questions get conflated constantly: “Do I own this?” (a contract question, answered by a platform’s terms of service) and “Is this protected by copyright?” (a law question, answered by courts and copyright offices). A platform can hand you “ownership” of something the law won’t protect.
- The U.S. currently requires meaningful human creative control for AI-assisted work to be copyrightable — prompts alone are, per the Copyright Office, not enough. A typed prompt is treated as an idea, not an expression.
- China’s courts have gone the opposite direction in some cases, crediting the human prompter as “author” — but even inside China, one court has already ruled the other way on similar facts.
- The UK and EU haven’t actually answered the ownership question yet. Getty v. Stability AI (2025) was about training data, not about who owns an AI image once it exists.
- For working publishers, the practical fix isn’t a legal opinion — it’s a documentation habit: track and preserve your own edits, selections, and arrangement, because that’s what the law is actually asking about.
Exhibit ATwo Different Questions Wearing One Costume
Here’s the confusion at the center of almost every “who owns AI content” argument I’ve seen play out in comment sections and Slack channels: people ask it as if there’s one authority to consult. There isn’t. There are two independent systems layered on top of each other, and a platform can win on one while losing on the other.
Layer one is contract. When you sign up for ChatGPT, Midjourney, or Adobe Firefly, you agree to terms that say who, as between you and the company, gets to call the output “theirs.” OpenAI’s consumer terms are explicit: “you (a) retain your ownership rights in Input and (b) own the Output,” with OpenAI assigning “all our right, title, and interest, if any” in that output to you.
Layer two is copyright law, which doesn’t care what a Terms of Service page says. Copyright is a government-granted monopoly on copying, and it only attaches to “works of authorship” — a category the U.S. Copyright Office and courts have repeatedly said requires a human author. If your output doesn’t clear that bar, OpenAI can assign you all the “right, title, and interest” in the world; there’s no enforceable exclusive right underneath it to assign. You can still use the image, publish it, sell prints of it — but you generally can’t stop someone else who finds the same output (which, per OpenAI’s own terms, “may not be unique”) from doing the same thing, because copyright infringement requires a copyrighted work to infringe.
My old “it’s public domain” framing missed the hybrid-work problem entirely. Most commercial AI content isn’t pure, un-edited output — it’s a raw generation that a human then selects, arranges, crops, rewrites, or combines with other material. The Copyright Office’s own January 2025 report is explicit that this kind of human “selection, coordination, or arrangement,” and any creative modification, absolutely can be protected — the AI-generated raw material just can’t be, on its own. Treating every AI-touched asset as automatically unprotectable is as wrong as treating it as automatically yours.
Exhibit BAn Original Framework: The Two-Layer Ownership Model
To stop conflating contract rights with copyright rights, I’ve started mapping every AI-content ownership question onto two axes instead of one. Call it the Two-Layer Ownership Model: the contract layer tells you who the platform says can use the output; the authorship layer tells you whether there’s anything legally exclusive to use in the first place. Only when both layers are satisfied do you have something an outside party is actually barred from copying.
This model resolves the question people actually mean to ask. “Can I stop a competitor from taking my AI-generated hero image and reusing it?” isn’t a Layer 1 question — OpenAI granting you the output doesn’t bind that competitor at all. It’s a Layer 2 question: did a human (you, or your editor) do enough to that image that a court would call it an original work of authorship? If yes, ordinary copyright infringement law protects you against the competitor. If no, your only real protection is separate — trademark, unfair competition, or plain old contract terms with people you directly license to.
Exhibit CHow We Got Here: The Case Law, in Order
Four rulings did most of the work of defining where the lines currently sit. None of them is final in the sense of settling the whole field — but together they’re the closest thing to a map that exists right now.
Thaler v. Perlmutter — United States
DeniedComputer scientist Stephen Thaler listed his own AI system, the “Creativity Machine,” as the sole author of an image and sought to register it, with himself as owner. The Copyright Office refused; the D.C. Circuit unanimously affirmed on March 18, 2025, holding that the Copyright Act “requires all eligible work to be authored in the first instance by a human being.” The Supreme Court declined to hear a further appeal on March 2, 2026, leaving the human-authorship requirement intact as binding precedent.
The case is narrow — Thaler deliberately claimed zero human involvement, which is not how most people actually use these tools — but it cemented the baseline: no human in the loop means no copyright, full stop.
USCO Copyright and AI Report, Part 2 — United States
Prompts insufficientReleased January 29, 2025, this is the document that actually governs day-to-day registration decisions, and it’s more nuanced than Thaler. The Office’s core finding: typing a prompt, even an elaborate one, doesn’t give the human enough “control over expressive elements” to count as authorship, because the same prompt fed into the same model can still produce meaningfully different outputs. But the report also draws a clear, useful line — a human’s own expressive input that is perceptible in the output, or their creative selection, arrangement, or modification of AI-generated material, is protectable. The Office illustrated this with the real example of the graphic novel Zarya of the Dawn, where the human-written text and overall arrangement were registrable even though the individual Midjourney-generated panels were not.
Li v. Liu — Beijing Internet Court, China
GrantedIn November 2023, the Beijing Internet Court found that an image a plaintiff generated with Stable Diffusion — after iterating on prompts and parameters and selecting a final version — reflected enough of his own “intellectual investment” to qualify as a copyrightable work of fine art, with the plaintiff as author. The court awarded 500 yuan against a defendant who reused the image without credit.
The ruling was celebrated internationally as evidence China would take a more permissive stance than the U.S. Real, but incomplete: a second Beijing-area court reached the same conclusion on similar facts in 2024, while the Zhangjiagang People’s Court denied protection to an AI image in early 2025 — meaning the “China grants it” headline oversimplifies a jurisdiction that is, itself, still working the question out case by case.
Getty Images v. Stability AI — UK High Court
Doesn’t answer output ownershipHanded down November 4, 2025, this was the UK’s first major AI copyright trial — but it addressed whether training Stable Diffusion on scraped Getty images infringed UK law, not who owns an image the tool later produces. Getty abandoned its primary copyright claims mid-trial and lost its secondary infringement claim outright; the court held that a trained model’s statistical weights aren’t a “copy” of the training images under UK law. Getty won a narrow trademark claim over outputs that reproduced its watermark. The UK and EU still have no binding ruling on AI-output authorship itself.
Layered underneath all of this is the still-unresolved training-data fight in the U.S. — The New York Times’ copyright suit against OpenAI and Microsoft survived a motion to dismiss in March 2025 and is proceeding toward trial, and the Copyright Office’s own May 2025 Part 3 report concluded that some AI training uses are unlikely to qualify as fair use, released, awkwardly, the same week its Register, Shira Perlmutter, was removed from her post. None of that determines who owns a given output — but it determines whether the models producing your content were built legally in the first place, which is a liability question every publisher using these tools inherits by proxy.
Exhibit DWhat the Platforms Actually Promise You
Because Layer 1 (contract) is the layer most people actually interact with, here’s what the major platforms’ current terms say — read directly from their published policies, not from secondhand summaries.
| Platform | Who owns the output (contract) | Commercial use | Key limitation |
|---|---|---|---|
| OpenAI (ChatGPT / DALL·E, consumer) | User owns output; OpenAI assigns its rights, if any | Permitted, subject to usage policies | Output “may not be unique” — assignment doesn’t cover identical output others receive |
| OpenAI API / Enterprise | Customer owns output; OpenAI assigns its rights | Permitted | Same underlying authorship gap as consumer tier |
| Midjourney | User owns output on paid plans | Free/Basic: personal only. Standard+: commercial. Firms over $1M revenue must be on Pro/Mega. | Midjourney retains rights to use prompts/images for model training |
| Adobe Firefly | User owns output on paid plans | Full commercial rights on paid tiers | Trained only on licensed Adobe Stock + public domain data, specifically to reduce infringement exposure; offers enterprise indemnification |
Adobe’s approach is worth pausing on, because it’s the clearest sign of where the market is heading regardless of how the courts eventually rule: license the training data cleanly, and the output-ownership question becomes much less fraught, because there’s no upstream infringement claim hanging over the model itself. Expect more platforms serving commercial publishers to follow that path, not because the case law forces them to, but because indemnification is a sales feature.
Exhibit EThe Control-to-Copyrightability Spectrum
Reading the U.S. Copyright Office’s actual examples, a spectrum emerges that’s more useful in practice than a binary yes/no. This is my own synthesis of the Office’s stated reasoning, not a quote from the report — but it maps cleanly onto every example the Office itself gave.
The unpopular take
Most of the “AI content ownership panic” in publishing circles is aimed at the wrong risk. Losing copyright exclusivity over a blog hero image or a social graphic is, for the overwhelming majority of commercial content, close to irrelevant — you were never going to sue a competitor over a stock-style AI illustration anyway, and you don’t need copyright to publish, license through a contract, or build a brand around something. The risk that actually costs people money is upstream: using a tool whose training data is the subject of active litigation (Disney and Universal’s suit against Midjourney is still open as of March 2026), or publishing AI text close enough to a specific source that it raises its own infringement exposure regardless of who “owns” your version. Chase documentation of your own editorial contribution for the pieces that matter competitively — not for everything you touch with a generator.
Exhibit FWhat This Means If You Publish AI-Assisted Content
- Keep an editorial record, not just a prompt log. The Office’s own guidance points at “creative selection, coordination, or arrangement” and “creative modifications” as the protectable layer — so what matters legally is evidence of your edits, rewrites, cropping, and combination choices, not just that you typed a clever prompt.
- Read the commercial-use tier, not just the ownership clause. Midjourney’s revenue-based tier requirement ($1M+ annual revenue needs Pro or Mega, regardless of actual plan) is the kind of detail that gets missed until a legal team asks about it during due diligence.
- Separate the training-data risk from the output-ownership risk. They’re different lawsuits, different defendants, and different timelines — a platform can be fully “yours to use” contractually while its underlying model is a defendant in active litigation over how it was built.
- Don’t assume “AI-assisted” reads as a single legal category anywhere. The line the U.S. draws (prompt vs. edit) is not the line China draws (intellectual investment) or the line the UK/EU haven’t drawn yet. If you operate across markets, the safest content is content with clearly documented, substantial human authorship everywhere.
- Expect this to keep moving. Part 3 of the Copyright Office’s own report was released the same week its Register was removed from office — this is not a stable regulatory environment, in the U.S. or anywhere else.
Exhibit GFrequently Asked Questions
Does registering an AI-assisted work with the U.S. Copyright Office guarantee protection?
No. Registration creates a legal presumption and is a prerequisite for filing an infringement suit, but the Office can still refuse or later have a registration challenged if it decides the human contribution wasn’t substantial enough. Applicants are now required to disclose AI involvement in the application.
If I heavily edit an AI image in Photoshop, is the final result protected?
The Copyright Office’s own examples support protection for the human-authored modifications and arrangement — the Zarya of the Dawn precedent is the clearest illustration. The underlying raw AI generation itself generally still isn’t independently protectable, but your edits and the combined final work can be.
Can two people legally publish the identical AI output and both claim rights to it?
Under U.S. law, if neither version clears the human-authorship bar, neither has an enforceable exclusive right in the raw output, so yes — this is explicitly why OpenAI’s terms note output “may not be unique.” Whoever added protectable human expression on top has rights to their addition, not to the shared raw layer.
Does using AI content violate a platform’s terms if I don’t disclose it?
That depends entirely on the platform you’re publishing to (not the AI tool) — some require AI-content disclosure, others don’t. Check the publishing platform’s policy separately from the generation tool’s terms.
Is the “prompts are ideas, not expression” rule likely to change as prompting gets more sophisticated?
The Copyright Office itself left the door open, noting a future could arrive where prompts “sufficiently control expressive elements” to count — but said current technology doesn’t offer that level of predictability. That’s a factual, not just legal, question, and it could shift as tools change.
Does the Getty v. Stability AI ruling mean AI training is legal everywhere now?
No — it means one specific secondary-infringement theory failed under UK law, on a narrowed set of claims Getty chose to pursue after abandoning its primary claims. It doesn’t resolve the training-legality question in the US, EU, or even fully in the UK, and Getty signaled it may pursue further claims.
Why did China’s courts reach a different conclusion than the U.S.?
The Beijing Internet Court in Li v. Liu focused on the “intellectual investment” reflected in the plaintiff’s iterative prompting and selection process, effectively treating sufficiently effortful prompting as authorship — a lower bar than the U.S. Copyright Office’s “control over expressive elements” test. But this isn’t settled Chinese doctrine either; other Chinese courts have gone the other way on similar facts.
What happens to content created before these rulings — is it retroactively unprotected?
Copyright status is generally assessed against the facts of how a specific work was created, not against the date a court clarified the standard. Older AI-assisted content that involved substantial human authorship was likely always protectable on that basis; content that was fully machine-generated was likely always unprotectable, even before Thaler made it explicit.
This article is provided for general informational purposes and reflects the state of the law as best understood on the date above. Legal outcomes are fact-specific; nothing in this article constitutes legal advice or creates an attorney-client relationship. Several described cases remain subject to appeal or further proceedings and may change after publication — see “Last updated” above.
Primary sources
- U.S. Court of Appeals for the D.C. Circuit, Thaler v. Perlmutter, No. 23-5233 (Mar. 18, 2025) — full opinion via Justia
- U.S. Copyright Office, Copyright and Artificial Intelligence, Part 2: Copyrightability (Jan. 29, 2025) — full report (PDF)
- Beijing Internet Court, Li v. Liu, (2023) Jing 0491 Min Chu No. 11279, with English translation — Internet Archive
- UK High Court, Getty Images (US) Inc & Ors v Stability AI Ltd [2025] EWHC 2863 (Ch) — summarized via Bird & Bird and Mayer Brown
- OpenAI, Terms of Use and Services Agreement (accessed 2026)
- Mayer Brown, “Supreme Court Denies Review in AI Authorship Case” (Mar. 2026)