Two separate questions people keep mixing up
The first: do you get the image from the vendor, and what may you do with it. That is a contract question, settled by the terms of use, and at most large vendors the answer is yes — they assign to you whatever they hold.
The second: does a copyright arise in the image, meaning can you stop anyone else from using it. That is a statutory question, and the answer in most countries today is no. The two are not the same, and their practical consequences differ as well: the first lets you use the image, the second would let you defend it. What you normally get today is the first and not the second.
The American position
The United States Copyright Office published a detailed report in early 2025, and its core fits in one sentence: a prompt on its own is not enough human control for copyright to arise. However long and precise your description of what you want, the gap between the intention and the output is too wide.
The courts have backed that up: in a long-running case the federal appeals court held in 2025 that an author has to be a human being, and in early 2026 the Supreme Court declined to take the case — which made the position final. The other well-known precedent is the comic book where the Office protected the text and the arrangement of the images, but not the generated images themselves.
The first European ruling
In Europe there was no decision for a long time, only theoretical argument. In February 2026 the first one arrived: a German court refused authorship for three AI-made logos. The reasoning matters because the claimant had used a serious prompt — roughly 1,700 characters — and the court did not find that sufficient either.
The other sentence that stuck from that ruling: picking the best of four suggestions afterwards does not create authorship. The European Parliament's non-binding resolution from the spring of 2026 points the same way, treating purely AI-produced content as in principle part of the public domain. There is still no EU-level rule, though: the question is decided by national law and by the European court's standard of "free and creative choices".
Where the UK and the common-law world stand
British law still has something that almost nobody else has. Section 9(3) of the copyright act contains a special rule for computer-generated works: the author is the person who made the arrangements necessary for the work to be created. There is hardly a parallel to it anywhere in the world. In March 2026, however, the UK government proposed repealing it, on the grounds that the provision does not fit the situation it now gets applied to. In other words, the one explicitly AI-friendly rule on the books is the one heading for the exit.
The other direction is China, where a court granted authorship in an AI image, pointing to the substantial human input behind it — while a decision in September 2025 refused it, because the claimant could not produce contemporaneous evidence of his own work. The practical lesson is the same either way: wherever authorship is possible at all, documenting the process is what decides it.
Where rights can still arise
Not every image touched by AI is without protection. In four situations the picture is different.
- Your own photograph, edited with AI. The original shot is yours, and it stays yours. Editing does not extinguish an existing copyright.
- Substantial post-work. If you genuinely rework the generated image — paint over it, composite it, compose with it — the human contribution can be protectable.
- Selection and arrangement. As above: a considered assembly of many images is an achievement of its own.
- Trademark. That is an entirely separate system: a mark made with AI can be registered, because what counts there is not creation but distinctiveness and use. For the details, Logos and Brand Assets is the post to read.
The other direction: the training data
So far this has been about whether there is a right in your image. The other question is whether training the model infringed somebody else's right — and that is the litigation the industry is watching.
There is a ruling here too: a large picture agency's British case largely failed at the end of 2025, and the decision made one important technical finding — the model's weights are not in themselves an infringing copy, because they do not store the images. That does not mean training is lawful everywhere; it means the "my pictures are inside the model" argument does not stand up in that form. Country by country and case by case, the question remains open.
The indemnity promise
Several vendors advertise that they will stand behind you if there is a legal problem. That is a real thing, but the small print matters: the best-known promise of this kind applies only to the enterprise plans, only to certain outputs, and does not extend to partner models — that is, to exactly the ones the same interface offers you alongside.
The practical lesson is not that it is worthless, but that you should read what it covers. If indemnity genuinely weighs in your decision, then the plan and the model it actually applies to are what count — not the vendor's name.
What to do in practice
- Do not build on exclusivity. If the image matters commercially, count on somebody else being able to use something similar, and on you not being able to stop them.
- Whatever is a brand mark should be a trademark. Trademark protection fills the gap left by copyright — that is the workable route for a logo or a sign.
- Add your own work. Your own photograph, substantial post-work, a considered arrangement: all three raise the odds that there is something to point at.
- Document the process. In a foreign case authorship failed on exactly this: there was no contemporaneous evidence of the workflow. Keeping the prompts, the intermediate versions and the dates is cheap insurance.
Common misconceptions
- "I wrote the prompt, so it is mine." Two courts and a copyright office have rejected that, even for quite long prompts.
- "The vendor assigned me the rights, so the image is protected." A vendor can only assign what it holds. If no copyright arose, there is nothing to assign.
- "If I label it as AI-made, I lose the rights." Labelling and legal status are two separate matters: on labelling, When to Disclose AI Use has the answer, and on the technical side Watermarks and Provenance picks it up.
- "This will get sorted out, and it will apply retroactively." It may well get sorted out — but you have to plan today's work by today's rules.
What to take away
The most likely position today is that you can use the image but cannot defend it. For most marketing uses that is not a problem. Where exclusivity is the value — a brand identity, a character, an illustration series, licensed material — you either have to add creative work of your own or reach for a different protective tool. The wider picture — the training-data side, the European court cases and what you can do with your own content — is put together in copyright and AI, a separate piece.