Can You Copyright What AI Makes? Mostly Not — Here's Where the Line Is

Here is a question that sounds trivial and isn’t. You open an image generator, type a careful prompt, tweak it twenty times, and get a picture you’re rather proud of. Do you own it? Most people assume yes — you did the work, you had the idea, you pressed the button. In the United States, the answer is a flat no: you own the copyright to precisely none of the AI-generated image, and that has real consequences the first time someone else decides to use it.

Answer first, and then the nuance, because the nuance is where the useful part lives. Under current US law, anything generated by an AI on its own is not copyrightable, because copyright requires a human author. A prompt, however elaborate, doesn’t make you that author. But the human contributions in and around an AI-assisted work — your own words, your editing, your selection and arrangement — can still be protected. The UK, almost alone, does something different and stranger, and is about to stop. Knowing exactly where that line falls is the difference between having something you can defend and having something anyone can legally take.

This isn’t legal advice, and the edges are genuinely unsettled. But the core rule is now clear enough, backed by a detailed government report and an appeals-court ruling, that anyone making things with AI should understand it before they build anything they care about on top of it.

The bedrock rule: no human, no copyright

US copyright protects “original works of authorship.” For more than a century, “authorship” has meant a human author — the doctrine that sank the famous “monkey selfie” case applies just as neatly to a diffusion model. The US Copyright Office has now said so repeatedly and in detail, most fully in the second part of its Copyright and Artificial Intelligence report, published in January 2025, which concluded that existing law already answers the question and no new legislation is needed: works generated wholly by AI are outside copyright, full stop.

The courts agree. In Thaler v. Perlmutter, the computer scientist Stephen Thaler tried to register an image called “A Recent Entrance to Paradise,” naming his AI system — the “Creativity Machine” — as the sole author. The Copyright Office refused, the courts backed the refusal, the D.C. Circuit affirmed in 2025 that a copyrightable work must be authored by a human being, and in 2026 the Supreme Court declined to hear the appeal, leaving the human-authorship rule firmly in place. There is now no serious ambiguity in the US about the extreme case: a work with no human author gets no copyright.

The principle isn’t peculiar to pictures. The same Stephen Thaler ran a parallel campaign in patents, naming his AI as the inventor, and lost the same way on both sides of the Atlantic: the US and UK patent systems, like copyright, are built around a human creator. Two different areas of law, asked essentially the same question, returned the same answer — rights attach to people, and a machine is a tool, not an author or an inventor.

Why a great prompt still isn’t authorship

The obvious objection is the interesting one. “But I did author it — I wrote a 300-word prompt, I iterated for an hour, the output reflects my choices.” The Copyright Office considered exactly this and rejected it, and the reasoning matters more than the conclusion. When you prompt a model, you are describing what you want; you are not controlling the specific expressive output — the exact composition, colours, phrasing or notes — that the model then produces from that description. Copyright protects the specific expression, not the idea or the instruction behind it, and the gap between “what I asked for” and “what the model chose to make” is precisely the gap where authorship would have to live.

The concrete example is instructive. Jason Allen’s image “Théâtre D’opéra Spatial” won a fine-art prize at the 2022 Colorado State Fair, and Allen sought to register it, reportedly having run more than 600 prompt revisions in Midjourney to get there. The Copyright Office refused: the work contained more than a trivial amount of AI-generated material that Allen would have to disclaim, and he declined to disclaim it, so it couldn’t be registered as submitted. Six hundred prompts is a lot of effort and skill. It still wasn’t authorship of the specific image, because the model, not the prompter, fixed the actual expression on the canvas.

What you can still protect

This is where the “you own nothing” headline oversells it, and where the practical value is. The rule bites on the AI-generated material specifically — not on everything you touched. Three kinds of human contribution survive:

• Your own expression. Text you actually wrote, a melody you actually composed, a drawing you actually drew — ordinary copyright, unaffected by the fact that AI sits elsewhere in the project.
• Selection and arrangement. If you choose, order and arrange AI-generated elements in an original way, that compilation can be protected — though only the arrangement, not the underlying AI pieces.
• Substantial human editing. If you take AI output and meaningfully transform it with your own creative choices, the human-authored changes can qualify. Using AI as one tool in a human-driven process doesn’t poison the whole work; it just means the machine-made parts have to be carved out.

The registration of the comic “Zarya of the Dawn” is the canonical illustration. Its creator, Kris Kashtanova, made the images with Midjourney and wrote the story herself. When the Copyright Office learned how the images were made, it cancelled protection for the individual pictures — “not the product of human authorship” — but kept protection for Kashtanova’s written text and for the original selection and arrangement of text and images into a comic. Same work, two answers, and the line runs exactly along the human/AI seam. If you make things with AI, that seam is the thing to understand.

It’s not just images: text, music and voices

The rule travels across media, because it’s about authorship, not art form. An article generated wholesale by a chatbot has no more copyright than a prompt-made picture — the human editing, restructuring and original writing you layer on top is what carries protection, not the raw draft. AI-composed music sits in the same bucket: a fully generated track from a text description isn’t yours to own, while human-written lyrics or a genuine human arrangement can be. And the further you get from text, the more other rights start to matter alongside copyright.

Voice is the sharpest example. An AI-generated voice clip isn’t copyrightable on its own — but cloning a real person’s voice can trigger an entirely separate set of protections, from publicity and likeness rights to the specific anti-cloning laws that are part of what AI regulation is starting to protect you from, none of which care about copyright and all of which care about consent. “I generated it, so I can use it” is doubly wrong there: you don’t own the output, and you may be trampling someone else’s rights in the person you copied.

There’s also a procedural catch worth knowing if you ever register something. The US Copyright Office doesn’t weigh human against machine in the abstract; it makes applicants disclaim the AI-generated portions — to state, on the application, which parts the machine produced. Refuse to disclaim, as Jason Allen did, and the work isn’t registered at all. Registration now comes with a compulsory honesty box about how much of the thing a human actually made.

The commercial sting nobody mentions in the demo

Why should you care, if you’re not a lawyer? Because copyright is what lets you stop other people using your work — and if a thing has no copyright, it is, in practice, free for the taking. Generate a logo, a character, a jingle or a stock image purely from a prompt, and you have no legal standing to stop a competitor copying it, selling it, or putting it on their own products. For a hobbyist that’s a curiosity. For anyone trying to build a brand or a business on AI-generated assets, it’s a structural problem the marketing around these tools is conspicuously quiet about.

Concretely: the AI-drawn logo above your shop can be copied by a rival tomorrow; the fully AI-written short story you posted can be reprinted without your say-so; the AI-generated stock image you’re selling is, legally, no more yours than anyone else’s. This isn’t hypothetical pedantry — it’s the difference between having a cease-and-desist that means something and having no standing at all. Add real human authorship — redraw the logo, rewrite and edit the story, composite and retouch the image — and you claw back protection over the parts you actually made. The tool didn’t hand you ownership; your own work did.

It connects to a wider fight we’ve covered from the other side. The same industry that benefits when its outputs are uncopyrightable and freely reusable is the one arguing hardest over the copyright of the inputs — the human work scraped to train the models in the first place, the unresolved question of who owns the words that trained your AI. And it’s a close cousin of a problem developers already hit, where AI-generated code might not be yours to license as you please. Ownership at both ends of the pipeline is murkier than the launch videos suggest, and the murk consistently favours the platform over the person using it.

The UK does something odd — and is about to stop

Almost everywhere, the human-authorship rule holds: the EU, following the same logic, protects only an author’s own “intellectual creation.” The UK is the conspicuous exception. Section 9(3) of the Copyright, Designs and Patents Act 1988 — written decades before generative AI — grants copyright to “computer-generated” works that have no human author, handing it to “the person by whom the arrangements necessary for the creation of the work are undertaken,” for a shorter 50-year term rather than the usual life-plus-70.

Before any UK reader celebrates, the catch is large. The provision is narrow, and legal commentators are openly unsure it even works for modern AI: UK copyright also demands that a work be the author’s own “original” intellectual creation, and it’s genuinely contested whether a purely machine-made output can clear that bar when no human exercised the creativity the test requires. There is no binding ruling settling it. And the direction of travel is towards the exit: the UK government’s 2026 review of copyright and AI leans towards repealing Section 9(3) altogether, keeping protection only for AI-assisted works where a human makes real creative choices — which would bring the UK into line with the US and the EU. So the honest summary for a UK maker is: there may be a thin, short, shaky copyright in purely AI-made work today, and it may not exist much longer. Don’t build on it.

What to actually do with this

None of this is a reason not to use these tools; it’s a reason to be clear-eyed about what you walk away owning. A few practical rules follow from the law as it stands:

• Assume prompt-only output is unprotected. If ownership matters — a logo, a product, anything you’d want to stop others copying — don’t rely on a pure text-to-image or text-to-text generation to be yours.
• Add real human authorship. Write the words yourself, edit and transform the output substantially, or make an original arrangement. The more genuine human creative choice sits in the final expression, the more there is to protect.
• Keep your working. If you ever need to claim the human parts, records of what you wrote, chose and changed — versus what the model produced — are the evidence the Copyright Office asks for.
• Don’t bank on the UK quirk. Section 9(3) is narrow, contested and on the government’s repeal list. Treat it as a curiosity, not a strategy.

The clean way to hold it in your head: AI is a brilliant tool and a poor author. Copyright rewards the author. So the value you can own isn’t in the button you pressed — it’s in the human judgement you brought before and after it. That’s not a loophole to resent; it’s the law quietly reminding you that the part worth protecting was always the part you actually did.