Who owns what it produces

Three separate questions people run together

1. Does the provider claim ownership of what you generate? Usually no. Most major providers’ terms assign whatever rights they have in the output to the user, sometimes with conditions. That is a contract between you and them, and it is the easiest of the three questions.

2. Is the output protected by copyright at all? Much harder, and it differs by country.

3. Could the output infringe somebody else’s rights? Different again, and the one with real exposure.

Conflating these produces most of the confusion in this area. A provider can perfectly well give you all their rights in something that nobody has any rights in.

Whether AI output is copyrightable

Copyright systems have generally been built around human authorship, and jurisdictions are reaching different conclusions about what that requires for machine-assisted work.

Some, the United States Copyright Office most prominently, have taken the position that material generated purely by a system in response to a prompt lacks the human authorship copyright requires, while a work containing sufficient human creative contribution can be protected as to that contribution. Some other jurisdictions, the United Kingdom among them, have long-standing provisions for computer-generated works with no human author. In March 2026 the UK government proposed removing its provision, while keeping protection for works made with AI assistance. No law has changed yet. Others have not addressed it directly.

The consequences for you are practical rather than theoretical. If you produce a logo, an illustration or a piece of text entirely by prompting, you may not be able to stop somebody else using it. If that matters (a brand mark, a book cover, anything you intend to enforce) it is worth understanding your own jurisdiction’s position before you rely on it, and worth documenting your own creative contribution.

Whether the output infringes

Two things are in dispute at once and it is worth separating them.

Whether training on copyrighted material without permission is lawful. Actively litigated in several countries, with different statutory starting points (some jurisdictions have text-and-data-mining exceptions, some do not) and no settled international answer. Cases are ongoing and settlements have been reached in some of them without establishing a general rule.

Whether a particular output infringes. Separate from the first, and more concrete. A model can reproduce training material closely, especially for material that appeared very often. Generating something in the style of a named artist is generally a different matter from reproducing a recognisable work, and trademark law applies to logos and characters regardless of what copyright says.

Several providers now offer indemnities to business customers for copyright claims arising from output, usually with conditions such as using their safety filters and not deliberately prompting for protected material. If you are producing commercial work at scale, that clause is worth finding and reading.

What to do while it is unsettled

• Do not generate a logo or a mark you intend to protect without advice about your jurisdiction.
• Do not prompt for a named living artist’s style for commercial work; it is the fact pattern most likely to attract a complaint, whatever the eventual law says.
• Keep your prompts and your edits. Evidence of human creative contribution is what several regimes turn on, and it costs nothing to keep.
• Check whether your client’s contract says anything. Increasingly they do, and a clause requiring disclosure of AI use is now common in publishing, design and journalism.
• Where the output must be clean, use tools with an indemnity and follow its conditions.

Disclosure, which is a separate matter again

Whether you must tell somebody that AI was used is not a copyright question at all. It comes from contracts, from professional rules, and from platform policies: a publisher’s submission terms, a client’s agreement, a journal’s authorship policy, an academic institution’s rules, or a marketplace’s requirement to label generated images.

These are proliferating and they vary widely, and breaching one is usually a contractual or disciplinary matter rather than a legal one. Check the specific rules of wherever the work is going. The general legal position, whatever it turns out to be, will not help you if the client’s contract said something else.

The position

This is unresolved, in different ways in different countries, and it will be resolved slowly and inconsistently. Anyone telling you that AI output definitely can or cannot be copyrighted, or that training definitely is or is not fair use, is describing a position rather than a settled rule. This lesson is not legal advice; where the answer carries money, it is a question for a lawyer in your jurisdiction, asked this year.