Two different questions get answered as if they were one. Can you sell it is usually yes. Do you own it is usually no. Here is where the line actually falls.
- The tool's permission and copyright are different things. A vendor can let you sell what you make. Only the law can make it yours, and it mostly does not.
- Prompting is not authorship. However long and specific the prompt, the US Copyright Office does not treat writing one as making you the author of the result.
- The Supreme Court left that in place on 2 March 2026 by declining to hear the case.
- One widely reported AI image registration shows the line — it came after repeated hands-on editing and a video of the whole process. The copyright covers the arrangement, not the generated material.
- Read your own tool's terms. Midjourney’s terms say you own what you make “to the fullest extent possible under applicable law”, with some exceptions. The one tied to a plan: companies over $1m in revenue must be on Pro or Mega to own their assets. There is no free plan either — no free trial on its website or in Discord — so in practice the way in is a paid plan.1First-hand: until 23 Sep 2026 this line said “Midjourney grants ownership to paid members only”. The terms in force since 27 May 2026 set no such condition; the only plan-related exception is the $1m revenue rule.
- The practical position: you can almost certainly sell it. You probably cannot stop anyone else using something near-identical.
The two questions people merge
Someone generates an image and asks "is it mine?" Inside that are two separate questions with different answers, decided by different people.
The gap between those two rows is where people get hurt. A vendor saying "you own the output" reads like the second answer. It is only ever the first.
A licence is permission to use. Copyright is the power to stop others. Most AI output comes with the first and not the second.
What the law actually says
Human authorship is the requirement
The US Copyright Office position is that human authorship is an essential part of a valid copyright claim. Material generated by a machine, without a human author, is not registrable.
This was tested and it held. In Thaler v. Perlmutter — an image entered for registration with no human author claimed at all — the Supreme Court declined to take the case on 2 March 2026, leaving the Copyright Office's refusal and the DC Circuit's agreement with it standing.2
A long prompt does not make you the author
The instinct is that a detailed prompt is the creative act. The Office does not agree. Text prompts, however elaborate, do not by themselves make you the author of the output. The reasoning is that you described a result; you did not execute it, and you could not predict precisely what would come back.
This is the single most misunderstood point in the whole subject. Prompt craft is a real skill. It is not, on this reading, authorship.
What does count: the cheese
One widely reported registration in the United States shows where the line sits: "A Single Piece of American Cheese", accepted by the Copyright Office in early 2025.
It was refused first. It was granted on reconsideration after the Office reviewed a video showing the applicant’s whole process — a generated starting image, then repeated inpainting, each pass a deliberate choice about what to change and where. The Office found enough human authorship in how the AI-generated material was selected, arranged and coordinated.
Read that wording carefully. The protection attaches to the selection and arrangement — the human decisions — and not to the generated material underneath. Someone else generating similar raw output is not infringing anything.3
The more you did after the model finished, the more there is to own. A one-shot generation gives you nothing to claim. A composition you assembled, edited, arranged and revised gives you something — and what you own is that layer, not the pixels the model produced.
Which leads to an unglamorous piece of advice: if the work matters, keep the evidence of your process — the practical end of the provenance question. The cheese was registered because its author could show the work, not because he asserted it.
What the tools grant — and what they carefully do not
Every vendor has a clause about ownership. They are not the same, and none of them can override the section above.
Midjourney, as a worked example
Their terms, in the version effective 27 May 2026, say: “You own all Assets You create with the Services to the fullest extent possible under applicable law.”4First-hand: until 23 Sep 2026 this line said the terms gave that ownership to “paying subscribers”. They do not say that.
That last clause is doing the work. Applicable law, per the Copyright Office, gives you very little in purely generated material. The sentence is accurate and it is also much smaller than it sounds.
Two things worth knowing:
The pattern across vendors
Vendors differ in detail but the shape repeats: paid tiers get commercial rights, free tiers often do not, and the grant is always framed as whatever the law allows. This is the same trap as hosting — a free plan that is free for personal use and not for anything that earns, worth reading beside what generation actually costs.
If money is involved anywhere in the chain, assume you need the paid tier, and check. The one thing you cannot do is infer one vendor's terms from another's.
So what can you actually do?
The Office requires you to say which parts are human and disclaim the AI-generated parts. This is not a technicality: the guidance states that failing to disclose AI-generated material risks the registration being cancelled.
If you have already registered something without disclosing the AI involvement, the route back is a supplementary registration correcting the record.
Before you sell anything
Is AI art stealing? Training data and artist consent
The plain point: this is a third question, separate from who owns the output. Image generators learn from huge collections of existing pictures (their training data), often gathered from the web without asking the artists. Whether that is legal is not settled: the US Copyright Office says some training will count as fair use and some will not, and the court cases so far point in different directions. This section describes published positions; it is not legal advice.
What the US Copyright Office says
In May 2025 the Office released a pre-publication version of its report on training generative AI. Its view: there is no blanket answer. Non-commercial research is likely to be fair, while copying works from pirate sources to make content that competes with them, where a licence could have been bought, is unlikely to be.8 The Office’s own page says a final version is still to come.9
One point often missed: copying an artist’s style is not, by itself, something copyright protects against. The same report records artists’ concerns about outputs that imitate a creator’s style, “which copyright does not protect”.8
What the courts have said so far
Books, in the US. Two US federal judges ruled in June 2025 that training language models on books was fair use in the cases before them, with limits. In the case against Anthropic, the judge held that training was fair use but that downloading pirated books into a library was not.10 In the case against Meta, the judge ruled for Meta but stressed that the authors had simply made the wrong arguments.11 Neither case was about images.
Images, in the US. The main artists’ case, brought by Sarah Andersen and other artists against Stability AI, Midjourney, DeviantArt and others in January 2023, argues that putting their work in training data infringed their copyright. In August 2024 the judge let those claims go forward.12 It has not been decided: a report on 28 Sep 2026 said the schedule had been extended again.13
Images, in the UK. In November 2025 the High Court ruled against Getty Images on its remaining copyright claim against Stability AI, finding that the trained model does not store copies of Getty’s photos. Getty had dropped its main training claim because there was no evidence the training happened in the UK, so the case that reached judgment no longer included a claim about the training itself.14
Consent and opt-outs
The European Union takes an opt-out approach: AI companies may mine works for training unless the rights holder has reserved them, and the EU AI Act requires model makers to respect those reservations. The US Copyright Office’s report describes the EU rules the same way, and notes “There continues to be controversy” over how they apply to generative AI.8
Critics argue that opting out gets consent the wrong way round. Some comments recorded in the same report say copyright is meant to be opt-in, and that opt-out markers are easily stripped from images.8 AI developers argue the other side. Adobe, for example, told the Office that training uses works for the ideas and styles in them, which copyright does not protect, to teach a model facts about the world.8 That is close to what the UK court found about Stable Diffusion: a model that learned patterns, not one that stores copies.
“Stealing” is a moral word and a legal one, and the two answers differ. Legally, it depends on where you are, how the images were obtained and what the model is used for, and the biggest image cases are not finished. Morally, the argument is about consent: whether artists should be asked first, or only be able to say no afterwards.
Whether training on artists’ work without permission is legal is unsettled: the US Copyright Office says it depends, US book cases went the AI companies’ way with limits, and the main US art case is still open. Copying a style alone is not protected by copyright.
Related on this site: AI Video Platforms Like Google Flow and Higgsfield, Sorted by Job · Art Styles for AI Prompts · Before You Publish AI Content
- 1Midjourney, Terms of Service, version effective 27 May 2026, read at source 23 Sep 2026“You own all Assets You create with the Services to the fullest extent possible under applicable law. There are some exceptions:” ↑
- 2US Copyright Office, “Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence”, Federal Register, effective 16 Mar 2023, read at source 23 Sep 2026copyright can protect “only material that is the product of human creativity” and that “author” in the Copyright Act “excludes non-humans”; Thaler v. Perlmutter, certiorari denied 2 Mar 2026. Checked 28 Aug 2026 ↑
- 3US Copyright Office registration of "A Single Piece of American Cheese", applicant Kent Keirsey of Invoke, as reported by the Harvard Journal of Sports and Entertainment Law, 24 Mar 2025, read at source 17 Sep 2026. It reports the Office acting “After initially denying the copyright claim”, says of the Office’s letter that “the Copyright Office described that it reviewed the video showing in full the graphic interface that Keirsey used, including his selection of component parts and his revised prompts”, and dates the acceptance: “In February, the Copyright Office accepted an application for copyright registration in an artwork titled A Single Piece of American Cheese.” An earlier version gave a registration date of 30 January 2025, a refusal in September 2024 and a quoted phrase from the registration; none of those is in this source, the public record at publicrecords.copyright.gov was not read here, and they were removed. It also called this the one AI image registered in the US, which no source here supports. The draft of this page said "roughly 35 inpainting edits"; the edit count appears in no retrievable source and was removed before publication.↑
- 4Midjourney, Terms of Service, section 4, Content Rights, read at source 23 Sep 2026.↑
- 5Midjourney, Free Trials, read at source 17 Sep 2026 and 23 Sep 2026“A limited trial is available on the niji · journey app, available for iOS and Android devices. No free trial is currently available in Discord or the midjourney.com website.” ↑
- 6Midjourney, Terms of Service, read at source 23 Sep 2026“If you are a company or any employee of a company with more than $1,000,000 USD a year in revenue, you must be subscribed to a “Pro” or “Mega” plan to own Your Assets.” ↑
- 7Midjourney, Terms of Service, section 4 (Content Rights), quoted verbatim from the version effective 27 May 2026, re-read 23 Sep 2026. Free-tier status per Midjourney's own documentation — the trial was withdrawn in 2023. The draft of this page said "free users get no commercial rights at all", which implied a free tier that does not exist; the check before publication caught it. Checked 28 Aug 2026 — terms change, and this one has changed before. Read the current version before relying on it.↑
- 8US Copyright Office, Copyright and Artificial Intelligence, Part 3: Generative AI Training (pre-publication version), 9 May 2025, read at source 5 Oct 2026“The Office expects that some uses of copyrighted works for generative AI training will qualify as fair use, and some will not.” and “the copying of expressive works from pirate sources in order to generate unrestricted content that competes in the marketplace, when licensing is reasonably available, is unlikely to qualify as fair use. Many uses, however, will fall somewhere in between.” ↑a“many commenters raised concerns about AI outputs that imitate a creator’s style, which copyright does not protect”. ↑b“Article 53 obligates AI model providers to establish policies for complying with Union law and to identify and comply with copyright owner opt-outs under the DSM Directive’s Article 4 TDM exception.” and the EU exception “conditions the availability of the exception on lawful access and respecting opt-outs by copyright owners”; “There continues to be controversy, however, over how the TDM exceptions apply to uses involving generative AI and whether and how the opt-out provision will work.” ↑c“Some stressed that copyright is by its nature fundamentally an opt-in system of exclusive rights, or asserted that requiring opt-outs would be burdensome.” and “Others raised concerns about the persistence of opt-outs given the frequency of metadata stripping”. ↑dAdobe’s comments, as quoted in the report: “Inputs are temporarily accessed for the unprotected ideas, concepts, and styles contained in the dataset--say, the number of fingers a human hand has, or what cars look like--to help the AI model learn facts about the world”. ↑e
- 9US Copyright Office, Copyright and Artificial Intelligence, read at source 5 Oct 2026“On May 9, 2025, the Office released a pre-publication version of Part 3 in response to congressional inquiries and expressions of interest from stakeholders. A final version of Part 3 will be published in the future, without any substantive changes expected in the analysis or conclusions.” ↑
- 10The Authors Guild, Mixed decision in Anthropic AI case, on the 23 Jun 2025 order in Bartz v. Anthropic, read at source 5 Oct 2026“The court ruled that both #1 and #2—the acts of training Claude on books, however obtained, and digitizing purchased (i.e., legally acquired) print books for internal storage—constituted fair use.” and “However, it found that #3—for Anthropic to knowingly download and copy pirated books for its library (from which it made subset copies to train its AI)—was not fair use.” ↑
- 11Authors Alliance, Meta wins on fair use for now, but court leaves door open for “market dilution”, 26 Jun 2025, on the 25 Jun 2025 ruling in Kadrey v. Meta, read at source 5 Oct 2026the ruling stands for the proposition that “these plaintiffs made the wrong arguments and failed to develop a record”. ↑
- 12NYU Journal of Intellectual Property & Entertainment Law, Andersen v. Stability AI: the landmark case unpacking the copyright risks of AI image generators, 2 Dec 2024, read at source 5 Oct 2026the artists accused the defendants of “copyright infringement for including the plaintiffs’ work in the dataset to train AI systems”; on 12 Aug 2024 the judge “denied Stability AI and MidJourney’s motion to dismiss the artists’ copyright infringement claims, allowing the case to move towards discovery.” ↑
- 13ChatGPT Is Eating the World, Sarah Andersen’s copyright case schedule gets extended by 3 months, 28 Sep 2026, read at source 5 Oct 2026“Judge Orrick approved the plaintiffs’ motion for an amendment to the case schedule” and “summary judgment motions won’t get resolved until the end of 2027 at the earliest.” ↑
- 14Latham & Watkins, Getty Images v. Stability AI: English High Court rejects secondary copyright claim, on the 4 Nov 2025 judgment, read at source 5 Oct 2026“the Model itself does not store any of those Copyright Works; the model weights are not themselves an infringing copy”; Getty accepted “there was no evidence that the training and development of Stable Diffusion took place in the UK”. ↑
US Copyright Office registration guidance, Federal Register, 16 March 2023 (copyright.gov/ai/ai_policy_guidance.pdf) · Thaler v. Perlmutter, Supreme Court certiorari denied 2 March 2026, leaving the DC Circuit decision standing · Copyright Office registration of "A Single Piece of American Cheese", early 2025, as reported by the Harvard Journal of Sports and Entertainment Law · Midjourney Terms of Service and Free Trials page. Checked 28 August 2026; the registration re-read 11 and 17 September 2026, the Midjourney trial 17 September 2026, and the Midjourney terms and trial page again 23 September 2026.
This is United States law, and it is not settled everywhere else. Other jurisdictions take different views on machine-made work, and this page does not attempt to summarise them. It is also not legal advice — it is a description of published positions, with the dates attached so you can check whether they still hold.
The honest summary: the tools will let you sell it, the law will not let you fence it, and the only part that is really yours is the part you did yourself.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK