A business runs a prompt and gets a finished image, a paragraph of marketing copy, or a block of code. The natural assumption is that whoever paid for the tool owns what came out. The harder truth is that under United States copyright law, content lacking sufficient human authorship may belong to no one, and a separate contract decides what the vendor lets the user keep.
Businesses, marketers, and creators reach the question of who owns AI-generated content from the same place: they have produced something with a generative tool and they want to know whether it is theirs to use, to sell, and to stop others from copying. A marketing team has generated a campaign image and wants to register it. A founder has used a model to draft product copy and assumes the output is a company asset like any other. An agency has delivered AI-assisted work to a client and now has to say, in writing, who owns it. In each case the instinct is that paying for the tool buys the rights to its output. That instinct is only partly right, and the part it misses is where the exposure sits.
This article addresses one narrow question — AI-generated content ownership — and stays on it. It is a focused companion to the firm's broader treatment of AI law, contracts, copyright, and governance, which covers the wider field; this piece does not repeat that ground. It explains the human-authorship requirement that sits at the center of AI copyright, what the requirement means commercially when output may be unprotectable, the role that human selection, arrangement, and modification can play, the separate contract layer in a vendor's terms of service, the distinct infringement exposure that comes from inputs and training data, and the governance a business can put in place where copyright protection is thin.
It is general information, not legal advice, and one caution runs through all of it: the law governing copyright in AI-generated material is unsettled and moving quickly, and what is described here is general and stated as of 2026. The Copyright Office has issued guidance, courts are deciding cases, and the rules that apply to a particular work depend on its specific facts, the tool used, and the law in force when the question is asked. The strongest position a business can take is to treat ownership as a question for counsel and a matter of records, not as something a tool's marketing page resolves.
Who owns AI-generated content, and the human-authorship requirement
The starting point for who owns AI-generated content in the United States is a requirement older than the technology: copyright protects works of human authorship. A copyright is the bundle of exclusive rights the law grants the author of an original work fixed in a tangible form — the right to copy it, to distribute it, to make new works from it, and to stop others from doing those things. The Copyright Office has taken the position, in its published guidance, that this protection extends only to material that owes its origin to a human being, and that purely machine-generated output, produced without sufficient human authorship, is not the work of an author the statute recognizes. The question of ownership therefore begins before ownership: a thing that copyright does not protect at all has no copyright owner to assign or to defend.
The case law has reinforced that position in what is often called the Thaler line of authority, which addressed an attempt to register a work asserted to have been generated autonomously by a machine, with the machine named as the author. The registration was refused, and the refusal was upheld, on the ground that a human author is required and that material lacking human authorship is not copyrightable. Stated generally, the principle that emerges is that a work produced entirely by a system, with no meaningful human creative contribution, falls outside copyright. This area is actively evolving and litigated as of 2026, and the contours of how much human involvement is enough, and of what counts as the relevant contribution, are exactly what courts and the Copyright Office are still working out.
What this means for the ownership question is direct. Where AI output lacks sufficient human authorship, there may be no copyright in it for anyone to own, and the absence is not cured by paying for the tool, by being the person who ran it, or by calling the result a company asset. Copyright AI questions are, at bottom, authorship questions, and authorship under current United States law is a human attribute. A business that wants to own its content as protectable property has to understand where the human contribution lives in what it produces, because that is what the next two sections turn on — first what the gap costs commercially, then what human work can do to close it.
A thing that copyright does not protect at all has no copyright owner to assign or to defend.
What unprotectable output means commercially
The commercial consequence of the human-authorship requirement is the part a business feels, and it is worth stating plainly. If a piece of AI-generated content is partly or wholly unprotectable by copyright, then to the extent it is unprotectable, the firm that made it generally cannot use copyright to stop a competitor from copying it. A marketing image, a block of generated text, or a generated design that falls outside copyright can, as a copyright matter, be reproduced by others, and the business that commissioned it may have no copyright claim to assert. Content a company assumed was a defensible asset can turn out to be something anyone is free to copy, and the company learns this at the moment it tries to enforce a right it does not have.
Registration sharpens the issue rather than avoiding it. Copyright registration is the formal filing with the Copyright Office that records a claim and, under United States law, is generally a prerequisite to bringing an infringement suit for a domestic work. The Office's guidance calls for applicants to identify and disclose the AI-generated portions of a work and to claim only the human-authored material, which means a business seeking to register AI-assisted content has to be candid about what the machine produced. An application that conceals AI-generated material, or that claims authorship of output the applicant did not create, risks the validity of the registration. The honest path is to register the human contribution for what it is, which requires knowing what that contribution was.
None of this means AI-generated content has no commercial value, and the point is not to discourage its use. It means the value may not rest on copyright, and a business should not price, license, or defend AI output as if a copyright stands behind every part of it when it may not. The practical posture is to treat protectability as uncertain until the human-authorship analysis is done, to avoid representing to clients or buyers that AI-assisted work is fully owned when that has not been established, and to lean on the other protections described later in this article where the copyright is thin. The exposure here is not abstract; it is the difference between an asset a company can defend and one it merely possesses.
The role of human authorship and why prompts alone fall short
If the absence of human authorship is the problem, then the presence of human authorship is the route to protection, and this is where a business has room to act. Stated generally, the human contributions that courts and the Copyright Office have treated as capable of supporting copyright are creative choices a person makes: the selection of which elements to use, the arrangement of those elements into a particular expression, and substantial human modification of machine output into something the person meaningfully authored. Where a human selects, arranges, edits, and reworks AI output enough that the result reflects original human expression, copyright may protect the human-authored portions of the work, even if the raw machine contribution standing alone would not be protected.
The limit that businesses most often misjudge concerns prompts. As a general matter, entering a prompt — describing what the user wants and letting the system generate it — has been treated as insufficient on its own to make the user the author of the output, because the prompt directs the machine rather than fixing the specific creative expression the way an author does. A detailed prompt is an instruction, and the system supplies the expressive choices that copyright cares about. This is a general statement of a developing position, not a fixed rule for every situation, and how prompts interact with other human contributions is among the questions still being worked out. The safe assumption for now is that prompting alone is unlikely to carry the authorship weight a business needs.
The practical lesson is that protection follows demonstrable human creative work, and a business that wants protectable content should structure its process so that real human authorship exists and can be shown. That means treating AI output as raw material that a person then selects from, arranges, and substantially revises, rather than as a finished product the user merely accepted. It also means keeping a record of what the human did, because the protectable portion is defined by the human contribution and the human contribution is a question of fact. Human authorship is not a formality to recite; it is the thing that, where it genuinely exists, gives the business something copyright will protect.
A detailed prompt is an instruction, and the system supplies the expressive choices that copyright cares about.
The separate contract layer in vendor terms
Copyright is only one of the two layers that decide what a user may do with AI output, and the second is contract. Whatever the copyright analysis says about authorship, the relationship between the user and the AI vendor is governed by the vendor's terms of service — the contract a user accepts when using the tool. Those terms decide, as between the two of them, what rights in the outputs the user receives and what the vendor retains. A vendor may grant the user broad rights to use, sell, and modify the output; it may keep certain rights for itself; it may impose restrictions on particular uses; and different vendors, and different service tiers, take different positions. The contract operates independently of copyright and can give a user permission to use output that copyright would not protect, or withhold permission that a user assumed came with the subscription.
Reading those terms before relying on a tool for commercial work is the practical safeguard, and the questions are specific. The terms address who owns or may use the outputs, whether the user's inputs may be retained or used to improve the model, what the user is permitted and forbidden to do with what it generates, and — importantly — whether the vendor offers any indemnity. An indemnity is a contractual promise by one party to cover certain losses of the other; some AI vendors have offered limited indemnities against certain third-party intellectual-property claims arising from use of their outputs, subject to conditions and exclusions. Whether such a promise exists, what it covers, and what conditions attach are matters a business should determine from the actual terms rather than assume, because an indemnity is only as good as its scope.
The two layers can point in different directions, and a business should hold both in view at once. The contract can grant generous usage rights while the copyright in the output remains weak or absent, so that the user is free to use the content but cannot stop others from doing the same. The contract can also restrict uses that copyright would otherwise have allowed. Neither layer answers the other's question, and a confident statement about ownership requires reading the terms and running the authorship analysis together. This is a setting where the documents control the outcome, and where the difference between what a business believes it owns and what it actually holds is found in language it may not have read.
Input infringement exposure and practical governance
Ownership of the output is a distinct question from whether using the tool creates infringement exposure on the input side, and the two should not be confused. Separate from who owns what a model produces, there is the risk that the material a model was trained on, or that a particular output reproduces, infringes someone else's copyright. Whether and when AI training and AI outputs implicate the rights of third parties whose works were used is the subject of active litigation as of 2026, and it remains unsettled. The point for a business is that even content it could not own as a copyright matter can still carry a risk that the content infringes another party's rights, and that risk is one of the reasons the indemnity question in the prior section matters. Output ownership and input exposure are two different problems that travel together.
Given a copyright picture this uncertain, the sensible response for a business is governance rather than guesswork. A human-in-the-loop process — one in which a person meaningfully selects, edits, and revises AI output rather than publishing it untouched — both improves the case for human authorship and reduces the chance of shipping problematic material unreviewed. Keeping records of the human contribution to a given work, so the business can later show what a person actually did, supports any later claim that the human-authored portions are protectable. Reading and retaining the vendor's terms, and knowing what they say about ownership, training use, and indemnity, closes the gap between what a business assumes and what it agreed to. None of these steps is difficult; what they require is that the business decide them on purpose.
Where copyright protection is thin, a business is not without recourse, and the closing point is to use the protections that do not depend on copyright. Trademark can protect the brand names, logos, and source identifiers under which content is offered, regardless of whether the underlying content is copyrightable. Trade secret can protect valuable information a business keeps confidential and takes reasonable steps to guard. Contract can bind the parties a business deals with — clients, vendors, contractors — to terms about use, ownership, and confidentiality that copyright alone would not supply. For a New York business in particular, where the firm's clients build and sell, the practical plan is to combine careful human authorship, disciplined records, and a clear reading of vendor terms with trademark, trade secret, and contract, and to route the specific facts to counsel because the law here is changing fast and a general article can frame the questions but cannot answer them for a given work.
Output ownership and input exposure are two different problems that travel together.