Part one: who owns it
The United States: no human author, no copyright
The US Copyright Office has been consistent. In 2023 it partly cancelled the registration of *Zarya of the Dawn*, a comic illustrated with Midjourney: the writer's text and her arrangement of panels stayed protected, the images themselves did not. A prize-winning image, *Théâtre D'opéra Spatial*, was refused. In 2025 a federal appeals court confirmed that a machine cannot be an author.
The Office's 2025 report addressed the question everyone actually cares about. Prompts alone do not make you an author, however long, however original, however many iterations you ran. But your own expressive contributions can be protected: a drawing you fed in, edits you made afterwards, the selection and arrangement of generated pieces into a larger work.
The UK and its family: an author by legal fiction
UK law has an unusual provision from 1988. Where a literary, dramatic, musical or artistic work is computer-generated with no human author, the author is deemed to be the person who made the arrangements necessary for its creation, with a fifty-year term. Ireland, New Zealand, Hong Kong, South Africa and India have similar provisions. The UK has consulted more than once on removing it.
India's own experience shows how unresolved this is. In 2020 the Indian Copyright Office registered an artwork listing an AI as co-author, then issued a withdrawal notice. The registration and the retraction are both on the record.
China: the opposite conclusion
In November 2023 the Beijing Internet Court held that an AI-generated image *was* protected, because the user's prompts, parameter choices and repeated refinement showed personalised intellectual investment. Similar facts, opposite result.
So
The same picture may be unprotectable in Washington, protected in Beijing, and owned by legal fiction for fifty years in London. Anyone telling you this is settled is selling something.
What to do while it stays unsettled
- Treat the prompt as worthless as property. It is not a moat, and in the US it is not authorship.
- If you need exclusivity, add human authorship you can point to, and keep the working files: your own photograph underneath, your edits, your compositing, your arrangement.
- Read platform terms as contract, not copyright. "You own your outputs" means the company will not claim them. It cannot manufacture a right that does not exist, and it usually does not stop another user generating something near-identical.
- For logos and brand marks, route around the problem. Trademark protects a mark through use and registration regardless of how it was drawn.
- The bigger commercial risk is rarely "do I own this". It is "does this infringe something": a trademarked character, a recognisable person, a style so specific it names a living artist.
Part two: who has to say it is AI
Here the law is moving faster, and in one direction.
China went first and hardest. Labelling rules in force since 1 September 2025 require generated content to carry both an explicit label a person can see or hear and an implicit label inside the file's metadata, and require platforms to check declarations and label what they detect.
The European Union. The AI Act's transparency article requires providers to mark generated output in machine-readable form, and requires anyone deploying a deepfake to disclose it. It was scheduled to apply from August 2026, and that timetable has been amended more than once — so confirm the current position rather than trusting any date, including this one.
India amended its IT Rules for synthetically generated information after a late-2025 consultation. The striking part is that it prescribes *visibility*: a label covering a defined portion of the frame and the opening portion of audio, plus a duty on platforms to verify user declarations. It is an answer to a real problem. A label nobody sees is not a label.
The United States has no general federal labelling law. It has the 2025 TAKE IT DOWN Act, requiring platforms to remove non-consensual intimate imagery including synthetic imagery, and a patchwork of state election-deepfake laws, several challenged on free-speech grounds.
Denmark moved to give people a right in their own likeness and voice.
The technical layer, and why it is not enough
Content Credentials, built on the C2PA standard, attach a cryptographically signed manifest to a file recording what created it and what edited it. Some cameras and most major generators support it. Invisible watermarking — patterns embedded in the pixels or the audio — is a second layer.
Both are real and both are fragile. A screenshot destroys them. Many re-encodes and social uploads strip them. And detectors claiming to identify AI content from pixels alone are unreliable in both directions: they clear synthetic images and they accuse real photographers, which has already cost people work and reputation.
The practical consequence: provenance from the source beats detection after the fact, and a visible label beats both, because it is the only part that survives a screenshot.
The rule that outlives the statutes
If a reasonable viewer could mistake your output for a record of something that happened, say that it is not — in the frame, at the start, where they will see it before they believe it. Not in the credits. Not in the alt text. Not in a caption they will scroll past.
Before you move on