Who Owns AI-Generated Content?

In most of the world today, the answer to who owns AI-generated content is short: nobody. A text, image, or video generated entirely by artificial intelligence, without any human creative contribution you can document, isn't protected by copyright. It has no author. Anyone can copy it, republish it, sell it, and you have no legal tool to stop them.

But that short answer hides a world of exceptions and country-by-country differences. China, for instance, has already granted copyright to an image generated with Stable Diffusion. The US Supreme Court just closed the opposite legal battle for good. Italy has a brand-new law trying to draw a line. And the EU is playing a different game altogether, less about who owns the output, more about what happens upstream, in the data used to train the models.

If you write, draw, or create content with AI's help, knowing where that line falls isn't a detail for lawyers to worry about. It's the difference between publishing something that's genuinely yours and publishing something anyone can take.

The common principle: you need a human author

Nearly every legal system in the Western world shares one basic idea, inherited from over a century of copyright law: a work is protected only if it has an author, and an author, by definition, is a person. Not a company, not an algorithm, not a piece of software.

Generative AI has put that principle to the test, and so far it has held up everywhere it's been challenged in court. The most cited case is American.

Thaler v. Perlmutter. Stephen Thaler tried to register copyright on an image created entirely by his AI system, the "Creativity Machine," listing the AI itself as the author. The US Copyright Office refused. Thaler sued, lost at the district court, lost on appeal, and on March 2, 2026, the US Supreme Court denied certiorari, leaving the ruling in place: a work created entirely by AI, without meaningful human involvement, cannot hold copyright under US law (Reed Smith, March 2026). The "humans-only" rule is, at least in the US, settled law now.

Where does Italy stand?

Italy moved first among EU member states with Law No. 132 of September 23, 2025, published in the Official Gazette on September 25, 2025 and in force since October 10, the country's first comprehensive AI legislation (Gazzetta Ufficiale, September 2025). The law also touches copyright, and it does so with a precise test: not every work created with AI's help is protected, only the one where a real creative contribution can be traced back to an actual person (Altalex, May 2026).

In plain terms: if you write a novel using AI to break through a block, rework a sentence, or generate a draft you then rewrite, you keep the rights, as long as your contribution is substantial and, above all, provable. If you just paste a prompt and publish the output as-is, that work risks having no protection at all.

One thing is still open: the implementing decrees, the ones meant to spell out the exact conditions and practical protections, are due by October 2026. Until then, the general principle is clear. The operational details aren't.

Where does the US stand?

Beyond the Thaler case, the US Copyright Office published a report on January 29, 2025 specifically on the copyrightability of works made with generative AI, the second in a planned series of three (U.S. Copyright Office, Copyright and Artificial Intelligence Part 2: Copyrightability, January 2025). The core finding: human authors are entitled to copyright over the part of an AI-assisted work that's recognizably theirs, including creative selection, arrangement, and edits made to the machine's output. But writing a prompt alone, no matter how detailed, isn't enough to make the work protected (Jones Day, February 2025).

One real case tells this story well. Kris Kashtanova published the comic Zarya of the Dawn, registering copyright in 2022 without disclosing that the illustrations were generated with Midjourney. When the Copyright Office found out, it opened a review and, on February 21, 2023, revoked protection on the individual images. In its official letter, the Office compared using Midjourney to using a camera, but noted that someone generating an image with a prompt doesn't have the same degree of control over the final result that a photographer has over their shot. The written text and layout, on the other hand, were creative choices made by the author, not the machine, and stayed protected (U.S. Copyright Office, Zarya of the Dawn letter, February 2023). It's the clearest example of how, in the US, copyright gets applied work by work, element by element, never as a blanket ruling.

The UK: the oldest rule, now on shaky ground

The UK is a case of its own. It has a law on computer-generated works dating back to 1988, decades before generative AI existed. Section 9(3) of the Copyright, Designs and Patents Act states that, for a literary, dramatic, musical, or artistic work generated by a computer, the author is treated as the person who made the necessary arrangements for the work to be created. Protection lasts 50 years, compared to the author's lifetime plus 70 years for human-made works (legislation.gov.uk, Copyright, Designs and Patents Act 1988, section 9).

It's a rule written for a different world, and the UK government knows it. On March 18, 2026, it published a report recommending that section 9(3) be scrapped, while keeping protection for AI-assisted works where a person makes recognizable creative choices. Most respondents to the public consultation agree: works generated entirely by AI shouldn't get copyright (HSF Kramer, March 2026). For now, though, it's only a recommendation. No law has been passed, and no UK court has yet ruled on whether a prompt can meet the originality standard the law requires.

China is the exception: the Li v. Liu case

China is, as of now, the country that has most openly recognized copyright on an AI-generated image. In November 2023, the Beijing Internet Court decided the Li v. Liu case. Li had created an image with Stable Diffusion, carefully choosing the prompts, the order of keywords, and the parameters, then published it online. Liu used it without permission as an illustration for a poem.

The court ruled that the image had originality and deserved protection as a work, because Li's intellectual effort in choosing the composition, selecting and ordering the prompts, and adjusting the parameters was enough to reflect personal, original expression (Kluwer Copyright Blog). In September 2025, the court included this case among eight "typical cases" on artificial intelligence, confirming it as a leading precedent (China Justice Observer).

The difference from the US is subtle but real: where the US Copyright Office tends to treat prompting alone as insufficient, a Chinese court has already treated it as sufficient, when it's documented and elaborate. It's proof that, on the exact same question, different countries can land on opposite answers.

The EU: less about ownership, more about transparency

The EU is tackling the problem from a different angle. Instead of focusing only on who owns the output, the AI Act imposes transparency obligations on model developers: as of 2026, companies must disclose the origin of the data used for training, honor rightsholders' opt-out clauses, and label AI-generated content as such (IAPP). The EU Copyright Directive, in parallel, already lets creators reserve the right to exclude their work from being used to train models.

On the ownership side, the European Parliament's committee vote in January 2026 backed a report proposing to exclude purely AI-generated content from protection and to introduce a labeling requirement. It's a proposal, not yet law, but it points a direction: Europe seems to be converging on the same principle as the US, a recognizable human contribution as the condition for protection.

What if the content is half human, half AI?

This is the question that applies to almost everyone, because in practice content is rarely purely human or purely AI. The answer, in every jurisdiction we've looked at, is the same: it's assessed case by case, and what matters is what you did, not what the tool did.

Elements that, under both US Copyright Office guidance and Italian law, tend to be protected:

• Creative selection of which outputs to keep and which to discard
• Reorganizing and arranging the generated material
• Substantial edits made to the text or image after generation
• The narrative structure, the plot, the chapter organization, if they're genuinely your own work

Elements that, on their own, tend not to be enough:

• A prompt, no matter how detailed or well written
• Simply picking one output among several generated options, with no further edits
• Accepting the output as-is, with no revision

The practical guideline that comes out of every source is the same: the more visible, traceable, and substantial your involvement, the more likely the final work is protected, and the more it pays to be able to prove it.

Getty Images v. Stability AI: the question still unanswered

There's a second legal front, separate from who owns the output: what happens when an AI is trained on copyrighted work without permission? The most significant case is British. Getty Images sued Stability AI, claiming its Stable Diffusion model had been trained on over 12 million Getty images taken without authorization.

On November 4, 2025, the UK High Court dismissed the main claim, the one for secondary copyright infringement. The judge's reasoning: the model doesn't store or reproduce the original images, only numerical parameters derived from them, so it stopped being an "infringing copy" once it no longer contains the protected work (Ropes & Gray). Getty did win, but only partly, on trademark infringement, because some generated images still showed the Getty watermark.

It's a ruling that clarifies barely more than it complicates. It doesn't say whether training an AI on copyrighted work is legal, only that, in this specific case, the finished model isn't itself a copy. The core question, about the training itself, remains open in other jurisdictions.

What this means in practice, if you write with AI's help

A few concrete takeaways emerge from all this, and they hold more or less everywhere. If you're wondering not just who owns a text written with AI but also where AI actually helps in a novel and where it risks flattening it, we covered that question in another article: Using AI to Write a Novel: Where It Helps (and Where It Doesn't).

Document your process. Save the intermediate versions of your work, not just the final result. If you ever need to prove your creative contribution, a revision history is worth more than a thousand words.

Don't stop at the prompt. If you generate a draft with AI and publish it as-is, you're taking a real risk of having no protection on that text at all. Rewrite it, reorganize it, add your own voice.

Track your substantial edits. Structure, characters, plot, style: these are the elements almost every jurisdiction recognizes as yours, if you can show you're the one who decided them.

Read the terms of service of the tool you're using. Some AI tools contractually grant you usage rights over the output regardless of copyright. Those are two different things: copyright is what the law grants you, terms of service are what the company promises you.

Expect differences from country to country. If you publish or sell content internationally, remember the same work could be protected in China and unprotected in the US, or the other way around.

Frequently asked questions

Can I sell an image generated with AI?

Yes, you can sell it. But if it was generated without any substantial creative input from you, anyone else can copy it and resell it too, because you have no copyright to enforce.

If I publish a book written with ChatGPT's help, who owns the rights?

It depends on how much you rewrote, reorganized, and edited the generated text. If your creative contribution is substantial and provable, the rights stay yours over the part you actually wrote or edited.

Does the EU AI Act change who owns the content?

No, not directly. The AI Act mainly governs transparency around training data and labeling of AI content. It doesn't set the rules for who owns the output, that's still a matter of national copyright law.

What happens if someone copies content I generated entirely with AI?

In most countries, including Italy and the US, you have no legal way to stop them, because that content isn't protected by copyright. It's one of the most concrete practical reasons your creative contribution matters, not just as a principle, but to actually protect what you publish.