Who owns what the AI made?

Ownership is usually decided before copyright law gets a turn

The question almost everyone asks first is whether AI output can be copyrighted. It is the right question and it has a fairly settled answer — the Copyright Office registers human authorship, and material a model produced without it is disclaimed rather than registered. But it is rarely the question that decides the outcome, because by the time it comes up the reader has usually already signed something: a provider's terms of service that says what it assigns and retains, or an agency agreement whose work-made-for-hire recital does less than it looks like it does.

So the useful move is to work backwards from what you are relying on. If the plan needs exclusivity — a brand mark, a registration, an exclusive licence to a client — then unregistrable output is a hole in the plan and the fix is either human authorship you can evidence or a different right doing the work. If the plan only needs to publish, ownership matters far less than the third-party exposure that travels with the output, and the tractable steps are scanning, provenance and knowing what your provider indemnifies.

Common questions

Can I copyright something an AI generated?

Not the machine-generated part. The Copyright Office registers works of human authorship, and its guidance on works containing AI-generated material asks applicants to disclaim the generated portions and identify what a person contributed. A work with meaningful human authorship — selection, arrangement, substantial editing — can still be registered as to that contribution. A prompt on its own has not been accepted as enough.

If it is not copyrightable, does that mean anyone can use it?

As a matter of copyright, largely yes: there is no exclusive right to assert against someone who reproduces material that nobody authored. That is why the useful question is rarely 'do I own it' and usually 'what am I relying on' — trademark for a brand mark, contract for a client deliverable, trade secret for something unpublished, and ordinary lead time for everything else.

We paid an agency and the contract says work made for hire. Are we covered?

Check it. Work made for hire reaches an independent contractor only for the categories section 101 enumerates and only where there is a signed writing, which is why well-drafted agreements pair the recital with a present-tense assignment. Even then an assignment can only convey rights that exist, so it passes nothing over material where no authorship attached. The practical fix is to also require disclosure of which parts were AI-generated, and to take the source files.

Does generated code create a licensing problem?

It can. A model trained on public repositories can emit material that matches licensed source, and the conditions attached to that source — attribution, and under copyleft the terms you distribute your own work on — do not disappear because a model produced it. Teams that ship generated code usually run licence and similarity scanning over it and keep a record of which files were machine-generated, so a later audit or diligence request has somewhere to start.

Are open model weights open source?

Usually not in the OSI sense. Most published weights ship under bespoke community licences carrying acceptable-use limits, redistribution conditions and sometimes terms that reach what you do with the output. Running the model on your own hardware does not clear any of that, so the licence and its acceptable-use annex are worth reading against what you are actually deploying.

Where do the training-data lawsuits leave us?

Unresolved, which is itself the planning input. Whether training on copyrighted works and generating output that resembles them infringes is being litigated, and the exposure sits with whoever publishes the output rather than with the model vendor by default. The tractable steps are checking generated work against the sources it most resembles before publishing, reading whether your provider offers an output indemnity and what voids it, and keeping the prompt and model version for anything published at scale.

Does this replace legal advice?

No. This is a reading of published US authority and ordinary contract practice, mapped to the description you gave. It is a starting checklist and a reading list, not an opinion on your facts, and it does not create a lawyer-client relationship.

Why does an accessibility company publish a content rights checker?

Because it is the same reader. The person working out whether the company owns the AI-generated art on its marketing site is usually the person who also owes an accessibility statement for that site and is tracking the AI obligations that attach to shipping the product. If your result says you are shipping generated output commercially, this points you at the obligations side too.