Who Owns What the Machine Makes? The Copyright Office’s 2025 Copyrightability Report, Federal Courts, and What Copyright Owners Should Do About AI-Generated Content

Nearly every business we represent now uses generative artificial intelligence (GAI) somewhere in its creative pipeline to draft marketing copies, generate product imagery, write and debug software, storyboard video, or compose music beds. The questions that follows are whether the company actually owns what comes out of the tool and whether it can stop a competitor from copying it. In January 2025, the United States Copyright Office answered that question at length in Part 2 of its report on Copyright and Artificial Intelligence, titled Copyrightability. In the twenty months since, the federal courts have largely confirmed the Office’s approach, most significantly when the Supreme Court declined on March 2, 2026 to review the D.C. Circuit’s decision in Thaler v. Perlmutter.

This blog post explains what the Copyright Office concluded and why, outlines the federal decisions that now frame the issue, identifies the questions that remain open, and offers practical recommendations for copyright owners who use—or whose employees and contractors use—AI tools. For an overview of our copyright registration, licensing and enforcement services, see our Copyright Matters summary; for how we handle infringement disputes, see Litigation & Enforcement.

The Copyright Office’s Part 2 Report: What It Says

Background and scope

The Copyright Office launched its AI initiative in March 2023 and, that August, issued a Notice of Inquiry that drew more than 10,000 comments from creators, technology companies, academics and the public. The Office is publishing its findings in parts. Part 1, released in July 2024, addressed digital replicas—AI-generated imitations of a real person’s voice or likeness—and recommended new federal legislation. Part 2, released on January 29, 2025, addresses the copyrightability of AI-generated output. A pre-publication version of Part 3, on the use of copyrighted works to train AI models, followed in May 2025. This post concerns Part 2 only; the training and fair-use questions in Part 3 are a separate and much larger subject that we will address in a future entry.

Part 2 is not a regulation and does not have the force of law. It is, however, the considered position of the agency that decides whether a work will be registered. Registration is a prerequisite to filing an infringement suit and to recovering statutory damages and attorney’s fees under 17 U.S.C. §§ 411 and 412. As discussed below, the federal courts have also given the Office’s reasoning considerable weight.

The Bedrock Requirement: a Human Author

The report begins from a principle the Office describes as the “bedrock requirement” of United States copyright law: protection extends only to works of human authorship. The Office traces this principle to the Constitution’s grant of power to secure to “Authors” the exclusive right to their “Writings.” It further references the Supreme Court’s 1884 decision in Burrow-Giles Lithographic Co. v. Sarony, describing an author as “the person who translates an idea into a fixed, tangible expression entitled to copyright protection,” and to more recent decisions such as Naruto v. Slater, where the Ninth Circuit found in 2018 a monkey could not be considered an author, and Urantia Foundation v. Maaherra, in which the same court held that a book purportedly dictated by celestial beings was protectable only because of the human selection and arrangement of its contents.

The Office’s conclusion is that this requirement does not need to change to accommodate GAI. In the Office’s words, “questions of copyrightability and AI can be resolved pursuant to existing law, without the need for legislative change.” The report also declines to recommend any new sui generis form of protection for purely AI-generated material, concluding that the incentives copyright is meant to create are not needed for outputs a machine can produce at a negligible cost.

Assistance Versus Generation

The report’s most useful analytical move is to distinguish between using AI as a tool that assists a human author and using AI as the generator of the expression itself. The Office is emphatic to state the former does not affect copyrightability. Its examples include a film editor using AI to remove a crowd from a shot or to age an actor, a musician using AI to develop a beat, an illustrator using AI for color correction or rotoscoping, and any author using AI to brainstorm, outline, research or check for errors. Where the human author remains the one who determines the expressive content of the finished work, the work may be protected as a whole, with the AI-generated portions remaining unprotected on their own.

The difficulty arises when the AI system, rather than the human, determines the expressive elements. The report identifies three ways a human can contribute to AI-generated output and analyzes each.

Prompts Alone Are Not Enough

The most consequential conclusion in the report is that, “[b]ased on the functioning of current generally available technology, prompts [alone do not] provide sufficient [human] control” to make the user the author of the output. The Office reasons that “[p]rompts essentially function as instructions that convey unprotectible ideas.” Even a lengthy and detailed prompt describes what the user wants—it does not control how the system produces it. The report points to the fact that the same prompt submitted repeatedly produces different outputs, that the system fills in every element the prompt does not specify, and that even the developers of these systems cannot predict what a given prompt will yield. The Office rejected analogies to Jackson Pollock’s drip paintings and to nature photography, observing that in those cases the artist controlled the execution—the choice of paint, the movement of the brush, the framing and timing of the shot—in a way an AI user does not.

Two related arguments were also rejected. The first is that iterative prompting—revising and resubmitting a prompt hundreds of times until the output matches the user’s vision— amounts to authorship. The Office concluded that it does not. Repeatedly re-running a prompt is essentially “re-rolling the dice,” and the effort expended does not convert the system’s expressive choices into the user’s, as the “sweat of the brow” theory has been extensively rejected by modern copyright jurisprudence. The second is what the Office called “authorship by adoption”: the idea that selecting a preferred output from among many candidates is itself an authorial act. The report concludes that choosing among options the machine generated is not the same as creating the expression, though it acknowledges that the selection and arrangement of multiple outputs into a larger work may be protectable as a compilation.

The Office was careful to tie this conclusion to “current generally available technology” and stated that it will revisit the question if systems evolve to give users genuine control over the expressive elements of the output.

Expressive Inputs

A user who inputs his or her own copyrightable work—a sketch, a photograph, a draft text—and directs the AI to modify it stands on different footing. Where the human’s original expression remains perceptible in the output, the human is the author of that portion. The report’s illustration is a hand-drawn sketch of a character that was fed into an image generator with a prompt; the outline, positioning and arrangement of the elements in the resulting image derived from the sketch were protectable, while the AI-generated rendering, lighting and three-dimensional effects were not. This principle matters a great deal in commercial practice, because most professional workflows begin with a human-created asset rather than a bare text prompt. Therefore, only the humanly created, protectable elements of the work will qualify for copyrightability.

Modification, Selection and Arrangement

Finally, a human who edits AI-generated material substantially enough to satisfy the originality standard, or who creatively selects, coordinates and arranges AI-generated elements, is the author of those contributions. The Office cautioned that whether a given set of modifications rises to the minimal level of creativity that copyright requires “will depend on a case-by-case determination,” and that the resulting protection extends only to the human-authored expression, not to the underlying AI-generated material. The Office’s registration practice, described below, reflects this framework: it will register the human-authored elements of a work that contains AI-generated material, provided the applicant discloses and disclaims the AI-generated portions.

How the Office Applies the Framework: Three Registration Decisions

Three registration decisions bracket the report and show where the Office draws the line. Each is a matter of public record and each has become a reference point in the litigation.

Zarya of the Dawn (February 2023). This decision illustrates the Office’s modification, selection, and arrangement of AI-generated content approach. Kristina Kashtanova wrote a graphic novel and illustrated it with images generated in Midjourney. The Office registered the text and the selection, coordination and arrangement of the images and text, but cancelled the registration as to the individual images themselves, concluding that Midjourney—not Ms. Kashtanova—determined their expressive content, notwithstanding the hundreds of prompts she used.

Théâtre D’opéra Spatial (September 2023). This decision rejects the technique of iterative prompting as an insufficient effort on its own to show authorship. Jason Allen’s image, which won the digital-arts category at the 2022 Colorado State Fair, was generated in Midjourney through at least 624 prompt iterations, then upscaled with Gigapixel AI and touched up in Adobe Photoshop. The Office’s Review Board refused registration of the work as a whole, finding that the Midjourney output was not human-authored and that Mr. Allen had declined to limit his claim to the Photoshop modifications. That refusal is now before a federal court, as discussed below.

A Single Piece of American Cheese (January 2025). This registration is a practical demonstration of the report’s compilation principle and of how a human can build a protectable claim on top of AI output by exercising region-by-region control. Shortly after the report was issued, the Office registered an image created by Kent Keirsey, the chief executive of the AI company Invoke, that was produced through roughly 35 rounds of “inpainting”—selecting specific regions of an AI-generated image and regenerating them with new instructions, then compositing the results. The Office registered the selection, coordination and arrangement of the AI-generated material but not the individual generated elements.

The Federal Courts: Thaler, Allen, and the Cases That Remain

Thaler v. Perlmutter — The Human Authorship Rule is Settled

The leading judicial decision is Thaler v. Perlmutter. Dr. Stephen Thaler applied to register a two-dimensional artwork titled A Recent Entrance to Paradise, listing as its author an AI system he calls the “Creativity Machine” and describing the work as “autonomously created by a computer algorithm running on a machine.” He claimed ownership of the copyright as the machine’s owner under a work-made-for-hire theory. The Office refused registration, and Dr. Thaler sued under the Administrative Procedure Act.

Judge Beryl Howell of the District Court for the District of Columbia granted summary judgment to the Office in August 2023, holding that “human authorship is a bedrock requirement of copyright” and that a work generated autonomously by a machine, with no human involvement, is not eligible for protection. The D.C. Circuit affirmed on March 18, 2025, in an opinion by Judge Millett joined by Judges Wilkins and Rogers. Rather than resting on constitutional grounds, the court held that the Copyright Act of 1976 itself “requires all work to be authored in the first instance by a human being.” The court walked through the statute’s provisions and found that they presuppose a human author at every turn: (1) copyright “vests initially in the author,” which presupposes the legal capacity to own property; (2) the term of protection runs for “the life of the author and 70 years after the author’s death”; (3) the Act refers to the author’s children, widow and widower, domicile and nationality; (4) transfers require the author’s signature; and joint authorship requires an intention that machines cannot form. The court also noted that the Act treats machines as tools, defining a “computer program” as a set of statements “to be used” to bring about a result.

Two aspects of the D.C. Circuit’s opinion are as important as the holding. First, the court expressly limited the rule to fully autonomous machine output: “The human authorship requirement does not prohibit copyrighting work that was made by or with the assistance of artificial intelligence. The rule requires only that the author of that work be a human being—the person who created, operated, or used artificial intelligence — and not the machine itself.” The court acknowledged that the Office “has allowed the registration of works made by human authors who use artificial intelligence” and that line-drawing disputes about how much AI contribution is too much would arise in other cases. Second, the court held that Dr. Thaler had waived the argument that he himself was the author by virtue of having built and operated the machine, because he had not adequately presented it to the Office and offered only a “bare and conclusory assertion” on appeal. The D.C. Circuit therefore never decided whether a human who designs and runs an AI system can claim authorship of its output.

The D.C. Circuit denied rehearing en banc, and Dr. Thaler petitioned for certiorari in October 2025 (No. 25-449), asking the Supreme Court to decide whether AI can be an author under the Copyright Act. The Solicitor General opposed review, emphasizing the narrowness of the question actually presented on Dr. Thaler’s own framing and the clarity of the statutory text. On March 2, 2026, the Supreme Court denied the petition. That denial is not a decision on the merits, but it leaves the D.C. Circuit’s opinion—the only appellate decision squarely addressing AI authorship—as the governing authority, and it follows the Court’s earlier refusal to review Thaler v. Vidal, in which the Federal Circuit held in 2022 that an AI system cannot be named as an inventor on a patent.

Allen v. Perlmutter — the case that will test iterative prompting

The question Thaler left open—how much human involvement to be considered an AI-assisted work is enough—is squarely presented in Allen v. Perlmutter, No. 1:24-cv-02665, pending in the District of Colorado. Mr. Allen filed suit in September 2024 under the Administrative Procedure Act challenging the refusal to register Théâtre D’opéra Spatial. He argues that the 624 prompt iterations, the use of Midjourney’s variation and upscaling tools, and his Photoshop refinements together reflect sufficient creative control to make him the author of the image, and that the Office’s approach is arbitrary and inconsistent with the technological neutrality of the Copyright Act. The Office responds that: (1) prompts are ideas and instructions; (2) the AI system rather than the user determines the expressive output; (3) that time, effort and iteration do not substitute for authorship and (4) that Mr. Allen could have—but refused to—register the work with a claim limited to his post-generation edits.

The parties filed cross-motions for summary judgment beginning in August 2025, and briefing was completed in January 2026. As of this writing the court has not ruled. Whatever the outcome, the decision is likely to be appealed to the Tenth Circuit, and it will be the first judicial ruling on whether extensive iterative prompting can amount to authorship under the framework the Copyright Office adopted in Part 2. Businesses whose AI workflows depend on prompting rather than on expressive inputs and post-generation editing should follow the case closely.

What the Courts Have Not yet Decided

It bears emphasis that no federal court has yet held that any AI-generated output is protectable, and no court has rejected the Office’s prompt-based analysis. The Eleventh Circuit, whose decisions govern the federal courts in Florida, has not addressed the question at all, and the Middle and Southern Districts of Florida have not issued a reported decision on it. The existing appellate authority in this circuit on the human-authorship requirement is limited to older cases involving works attributed to non-human sources, and the Eleventh Circuit would likely treat Thaler and the Office’s registration decisions as persuasive. The open issues include:

• How much post-generation modification is enough to support a registration;
• Whether a user who trains or fine-tunes a model on her own works and then prompts it stands in a different position from a user of a public tool;
• Whether a sufficiently constrained system (such as inpainting or region-by-region control) gives the user the kind of control the Office says prompts lack; and
• How the human-authored and AI-generated portions of a work are to be separated for purposes of an infringement analysis.

Each of these questions will be litigated in the next several years, and the answers will depend heavily on the record each claimant builds.

Recommendations for Copyright Owners

The practical upshot of the report and the case law is that copyright in an AI-assisted work is not lost or won at the moment of creation; it is built or forfeited by the way the work is made and documented. The following recommendations are the ones we most often make to clients across industries—from software and marketing to entertainment, publishing and product design—and they are organized around the workflow itself.

1. Design the workflow so that a human determines the expression

The single most important step is structural. If AI is used to produce a final asset from a text prompt, the asset is very likely unprotectable in the United States regardless of how many prompts were used. If, instead, the workflow begins with a human-created asset (a sketch, a photograph, an outline, a melody, a draft) and uses AI to develop or refine it, and a human then makes substantive creative decisions about the result, the human contributions are protectable and may often extend to the overall work. In practice this means: start from original inputs where possible, use tools that allow region-by-region or element-by-element control rather than whole-image generation, and build in a human editing stage that makes real creative changes rather than cosmetic ones.

2. Document the human contributions as the work is created

Because the Office and the courts decide these questions on a case-by-case basis, the record is everything. For any commercially significant AI-assisted asset, retain the original human inputs, the prompts and settings used, the intermediate outputs, the edits made and by whom, and the reasons for creative choices. Tools that keep layered files or version histories (Photoshop, Figma, DAWs, code repositories) make this easy or screenshots and a short creation memo suffice where they do not.

This documentation allows the registration application to be completed accurately, supports the claim if the Office asks questions or refuses, and becomes the evidence of authorship in any later infringement suit, where the defendant will argue that the elements it copied were generated by the machine and belong to no one.

3. Register early and accurately claim what is human

Registration remains the gateway to enforcement, and the Office’s rules require applicants to disclose AI-generated material that is more than de minimis and to exclude it from the claim. The Office’s March 2023 registration guidance and its Compendium of practices explain how to do this: identify the human-authored material in the “Author Created” field, describe the AI-generated material in the “Limitation of Claim” section, and where the claim rests on selection, coordination and arrangement, say so.

Two cautions follow from the case law. First, do not overclaim. An application that lists AI-generated material as human-authored is inaccurate, and under 17 U.S.C. § 411(b) a knowingly inaccurate application can invalidate the registration in litigation. Second, do not underclaim by refusing to limit the claim, as Mr. Allen did: a registration for the human-authored portions is worth far more than a refusal for the whole. Our Copyright Matters page describes our registration services, and we regularly prepare applications for AI-assisted works with the limitation language the Office expects.

4. Fix ownership by contract, because copyright may not

Where an asset is created by an employee within the scope of employment, the employer owns whatever copyright exists. Where it is created by a contractor, agency or freelancer, ownership passes only by written assignment or a qualifying work-made-for-hire agreement. However, if the asset is AI-generated and unprotectable—there may be no copyright to assign at all.

Agreements with creative vendors should therefore (a) require disclosure of AI use and identification of AI-generated components; (b) assign all copyright and, separately, all other rights in the deliverables, including any rights arising under foreign law; (c) warrant that the human-authored components are original and that the vendor has complied with the AI tool’s terms of use; (d) allocate the risk that a deliverable proves unprotectable or infringes a third party’s work; and (e) require delivery of the documentation described above. The same analysis applies to the AI tool’s own terms of service: most major providers assign or license output rights to the user, but the terms differ, some restrict commercial use, and the provider cannot convey a copyright that does not exist.

Our technology and start-up and entertainment law practices routinely draft these provisions.

5. Use the other tools in the box

For assets that are likely to be treated as AI-generated, copyright is not the only or necessarily the best form of protection. A logo or product name generated with AI can be protected as a trademark once it is used in commerce, because trademark law asks whether the mark identifies source, not who created it; see our Trademark Matters summary. The ornamental design of an article can be protected by a design patent, which requires a human inventor but not human “authorship” of every line. Proprietary prompts, fine-tuned models, training datasets and the workflows that produce consistent output can be protected as trade secrets if reasonable measures are taken to keep them confidential. And access to AI-generated content can be controlled by contract, through terms of use, licenses and technical measures, even where copyright would not prevent copying.

6. Reassess enforcement positions on both sides

Owners contemplating an infringement suit over an AI-assisted work should expect the defendant to attack the registration and the ownership of the copied elements at the outset, and should be prepared to prove which elements were human-authored. Conversely, businesses accused of infringing a competitor’s AI-generated content should investigate how that content was made. If the copied elements were generated by a machine, the plaintiff may own nothing in them. The same investigation applies to AI-generated infringement of your own works—an output that reproduces your protected expression is actionable regardless of whether the output itself is copyrightable.

Our Litigation & Enforcement team handles both sides of these disputes.

7. Think internationally

Other jurisdictions have taken different approaches. The United Kingdom, Ireland, New Zealand and several other countries have statutory provisions granting a form of copyright in computer-generated works to the person who made the arrangements necessary for their creation, and a Chinese court has recognized copyright in an AI-generated image based on the user’s prompt selection and adjustments. The Copyright Office acknowledged these divergent approaches and declined to follow them. The practical consequence is that a work may be protected abroad and unprotected here. Companies distributing AI-assisted content internationally should map the protection available in each key market and structure their workflows for the most demanding regime—which, for now, is the United States.

8. Adopt a written AI-use policy

Finally, every organization that creates content should have a short written policy, reviewed by counsel, covering (1) which AI tools are approved; (2) what may and may not be input into them (client confidential information and third-party copyrighted works are the usual prohibitions); (3) how AI-generated material is to be labeled and documented (4) who is responsible for the human creative review, and (5) how AI use is disclosed to clients, licensees and the Copyright Office. This type of policy protects the company’s own copyrights, reduces the risk of infringing others’, and makes the documentation described above routine rather than exceptional.

Conclusion

The Copyright Office’s Part 2 report and the Thaler litigation have settled the outer boundaries of the debate. A work created autonomously by a machine has no author and no copyright in the United States, and the Supreme Court has declined to disturb that rule. A work in which a human merely uses AI as a tool while retaining control over the expression is protected as fully as any other. Between those poles, which is where most commercial content now lives, the answer depends on what the human actually contributed and whether it can be proved. Prompts alone will not do it under the Office’s current view, but expressive inputs, meaningful modification, and creative selection and arrangement will. Allen v. Perlmutter will tell us whether the courts agree about prompting, but it will not change the practical advice: design the workflow so that a person determines the expression, document it, register what is human, and use contracts and other forms of intellectual property to cover the rest.

If your business is creating, licensing or enforcing rights in AI-assisted content, or has received a claim involving such content, the ADD+G attorneys can evaluate the workflow against the Copyright Office’s framework and the current case law and recommend a protection strategy. Please visit our Copyright Matters and Litigation & Enforcement pages, or contact us at 305-374-8303.