Copyright and AI Output: Who Owns It

“Who owns AI output?” is three questions bundled together, answered by three different bodies of rule, with three different answers. Separating them takes a paragraph and removes most of the argument.

Engineering and product guidance, not legal advice — and this is an area where the law differs sharply by country and is actively developing. Anything below that matters to a commercial decision needs confirming against the current position where you operate.

Three questions in a trench coat

• As between you and the provider, who has the rights? A contract question, and usually the easiest to answer definitively.
• Is the output protected by copyright at all? A question of law, and one where the answer in several jurisdictions turns on human authorship.
• Could using the output infringe someone else’s rights? A risk question, independent of the other two — output you clearly own can still be output you cannot safely publish.

People typically get a comfortable answer to the first and assume it settles the third. It does not, and the third is the one that produces incidents.

Ownership is mostly a contract question

Whatever rights exist in an output, the allocation between you and your provider is set by their terms. Find the clause — it is usually near the definitions of Input and Output — and read it for four things rather than for the headline sentence:

Read your own vendor’s current terms rather than a summary of anybody’s. These clauses have changed repeatedly across the market and the version you signed under may not be the version published today.

Protectability is a law question

A contract can allocate rights; it cannot create rights that do not exist. Several jurisdictions condition copyright on human authorship, and on that basis material generated without sufficient human creative contribution may not attract copyright protection at all — meaning nobody owns it, including you, and anyone may copy it.

The line is drawn around human contribution rather than around tool use, which is the part worth internalising. A work in which a human made the creative choices — selection, arrangement, substantial editing — and used a model as an instrument is treated differently from output accepted as generated. The US Copyright Office has published guidance on registering works containing AI-generated material, and other jurisdictions have taken varying positions, some more accommodating. Read the current guidance for the places that matter to you rather than assuming the answer travels.

The practical consequence for a product: if a generated asset is commercially important — a logo, a character, a piece of code you intend to license — human creative involvement is not merely a nice story about your process, it is the thing that may determine whether you have anything to enforce. Structure the workflow so that involvement is real and documented.

Infringement is a risk question

Independent of ownership: a model can produce output substantially similar to protected material, particularly for content that was distinctive and heavily repeated in training data. Logos, well-known characters, distinctive prose styles and widely copied code snippets are the usual suspects, and asking a model directly for something in the manner of a specific creator raises the risk rather than transferring it.

Trade marks are a separate and frequently forgotten axis. A generated image containing a recognisable brand mark is a trade mark problem whatever the copyright analysis says.

Proportionate controls, in the order they usually pay off: do not prompt for named living creators or specific protected works in production paths; run similarity checks on high-exposure assets before publication; keep a human in the approval path for anything public-facing; check the licence terms attached to code suggestions your tooling surfaces; and know what your provider’s indemnity actually covers before you rely on it.

The records worth keeping

If a question arises later, the answer lives in records you either kept or did not. For any asset that matters commercially, keep the prompt or brief, the model and version, the date, what a human changed afterwards and who, plus any similarity or clearance check performed. A short structured note attached to the asset is enough, and it costs nothing at creation time.

The reason to bother is narrow but real: every one of the three questions above turns partly on facts about how the thing was made, and those facts are unrecoverable a year later. The team that can answer “what did the human actually contribute” is in a materially better position than the team that cannot, whichever way the law settles.

Two adjacent habits are worth adopting at the same time, because they cost nothing once the note exists. Keep the record with the asset rather than in a separate tracker — a sidecar file, a database column, a field in your asset manager — since a tracker maintained by discipline alone stops being maintained the first busy week. And keep it for code as well as for images and prose: generated code raises the same three questions, the licence position on training-derived suggestions is the same open area, and code is the asset your company is most likely to be asked about in a due diligence process.

Finally, be precise in customer-facing language. There is a real difference between “we own this”, “we have the rights we need to license this to you” and “this is not protected by copyright and anyone may use it”. A contract or a marketing page that overstates the first when the accurate answer is the second is a claim somebody can check, and this is a field where people increasingly do.