Using generative AI does not automatically put a work outside copyright protection. But it does not automatically make the user its author, either. In the United States, the central question is increasingly whether a person contributed enough original creative expression to claim copyright in the finished result. That shifts the AI copyright ownership debate away from the model’s training data and toward the practical realities of making an image, song, book, game asset or software product with machine assistance.

For creators and companies, the implication is straightforward: the more a work depends on an AI system’s uncontrolled expressive choices, the less certain its copyright status becomes. Human direction, editing, arrangement and transformation can matter. So can the ability to show what the human actually did.

The policy debate is moving from inputs to outputs

Litigation over whether AI developers may train models on copyrighted books, images, music and code remains consequential. But that is only one part of copyright law and generative AI. A separate question arises after a model produces something: who, if anyone, owns the result?

The U.S. Copyright Office sharpened that distinction in Copyright and Artificial Intelligence, Part 2: Copyrightability, published on January 29, 2025. The report concluded that existing U.S. copyright law can generally address AI-assisted works without a new special copyright regime. Its core position was consistent with the office’s March 2023 registration guidance: copyright protects human authorship, while material generated by a machine is not protected when the machine, rather than a human, determines the expressive elements of the output.

The report did not say that every work touched by AI is unprotectable. Instead, it focused on identifying the human contribution. A person may copyright their original selection, coordination, arrangement or modification of AI-generated material, provided those contributions meet the ordinary threshold for copyright. Protection, however, extends only to the human-authored parts.

Human authorship remains the legal baseline

Copyright protects original works of authorship fixed in a tangible medium. U.S. courts and the Copyright Office have long understood “authorship” to require a human being. That principle was reaffirmed in Thaler v. Perlmutter, a case involving an image that its applicant said had been created autonomously by an AI system.

A federal district court upheld the Copyright Office’s refusal to register the image in 2023. In March 2025, the U.S. Court of Appeals for the D.C. Circuit affirmed that decision, holding that the Copyright Act requires human authorship. The case concerned a work presented as wholly AI-generated, not a conventional AI-assisted project in which a person supplied substantial creative work. That distinction matters: the ruling did not establish that all generative-AI use defeats copyright.

Nor is there a numerical rule. U.S. law does not say that a work becomes protectable after a creator makes 20 prompts, changes 30 percent of an image or spends a particular number of hours editing. The issue is qualitative: did the human author create original expressive elements in the work being claimed?

Prompting alone is an uncertain foundation for ownership

The hardest cases sit between fully autonomous output and plainly human-made work. A detailed prompt can reflect imagination, planning and taste. Yet the Copyright Office has said that, with currently available generative systems, prompts generally do not give users enough control over the specific expressive details needed to make them authors of the resulting output.

That is not a judgment that prompting takes no skill. It reflects how most image, text and audio generators operate: users request a result, but the system determines many details of wording, composition, visual form, timing or other expression. An instruction to create a melancholy watercolor cityscape may be creative, but it is still different from personally choosing the brushwork, buildings, perspective and colors that appear in the final image.

The analysis can change when a creator does more than request output. The Copyright Office’s guidance and 2025 report point to several potentially protectable forms of human contribution:

• Selection and arrangement: choosing among outputs and combining them into an original, expressive composition, such as a book layout, comic sequence or audiovisual work.
• Human editing: revising AI-generated prose, repainting an image, altering audio or otherwise adding original expression.
• Transformation: using machine output as raw material for a substantially human-created final work.
• Traditional creative elements: writing the script, designing the characters, composing non-AI material, photographing source material or building the software architecture around an AI-assisted component.

Selection by itself is not always enough. Simply picking the best result from many generated images may be a weak claim if the images are otherwise unchanged. But selecting and arranging material into a larger expressive whole can be protected when the arrangement is original. The boundary will be fact-specific, and it will likely be tested project by project rather than resolved by a single rule about prompts.

Documentation could become part of the creative workflow

As disputes become more practical, evidence of creative control may matter almost as much as the work itself. A registration application, licensing negotiation or infringement dispute may require a creator to explain which parts were generated and which parts were authored by people.

That makes ordinary production records more valuable. Creators working on commercially important AI-assisted creative work may want to preserve:

• prompt histories and version histories;
• drafts before and after human revision;
• layered image files, project files and source code repositories;
• notes showing editorial, design or compositional decisions; and
• records identifying AI-generated material that was retained, replaced or substantially altered.

These records do not create copyright by themselves. They can, however, help establish a credible account of authorship. Disclosure also matters in Copyright Office registration. The office requires applicants to exclude AI-generated material from a claim and describe the human-authored contribution. Failing to disclose material information can jeopardize a registration.

The commercial stakes reach far beyond artists

For publishers, advertisers, film and television producers, game studios and software businesses, uncertain AI copyright ownership is a chain-of-title problem. A company buying an illustration, jingle, code module or article wants confidence that the seller can grant the promised rights and can enforce them against copycats.

If a key component is unprotected machine output, competitors may be able to reproduce that component without infringing copyright, even if other human-authored parts of the project remain protected. Contracts can allocate risk between a client and freelancer, but contracts cannot manufacture copyright in material the law does not protect.

That is why organizations using generative tools are likely to place more emphasis on disclosure policies, approval steps and records of human revision. The goal is not necessarily to ban AI-assisted creative work. It is to know what rights exist, what cannot be promised and where additional safeguards—such as trademarks, trade secrets, confidentiality or contractual restrictions—may be needed.

Output ownership does not settle the training-data fight

A creator’s ability to claim copyright in an AI-assisted creative work is separate from the unresolved question of whether a model was trained lawfully on copyrighted material. A work might contain sufficient human authorship to be copyrighted while the developer of the underlying model still faces claims about training data. Conversely, a model’s training process might be found lawful in a particular context while a largely machine-generated output remains ineligible for copyright.

Tool terms also deserve separate attention. A platform’s contract may grant users certain rights in outputs, limit uses, or reserve rights for the provider. Those terms govern the relationship between user and platform; they do not override the statutory requirement that copyright protect human authorship.

A durable rule for a changing toolset

Generative AI is forcing copyright law to ask an old question in new production environments: who made the expressive choices that matter? The U.S. answer so far is not that AI-assisted work is categorically ineligible, nor that typing a prompt is automatically authorship. It is that protection follows meaningful human creative contribution.

For anyone making work with AI, the most useful approach is to treat the system as part of a documented workflow rather than as a black-box author. Bring original judgment to the material, revise it meaningfully, preserve evidence of those choices and be candid about the machine-generated portions. The next AI copyright fight is likely to turn less on whether a tool was used than on whether a person can show what they created.

Image by StockSnap on Pixabay.