Copyright in AI-Assisted Works: Human Authorship, Thaler v. Perlmutter, and the Copyright Office Guidance

In September 2022, a digital artist named Jason Allen won the digital art category at the Colorado State Fair with an image called Théâtre D'opéra Spatial, a sweeping baroque scene of robed figures before a glowing portal. He had made it with Midjourney, an image generator, over hundreds of prompts, then touched it up in Photoshop. A year later the Copyright Office told him that the part of the picture most people would call "the art" was not his, because a machine had produced it, and that his registration would cover nothing unless he disclaimed what Midjourney made. Allen refused and sued. The same fight, in a hundred quieter forms, is now sitting on the desk of every publisher, studio, agency, and software company that has let generative AI into the production pipeline. The question is old. The tools are new. The stakes are whether the thing you paid to create can be owned at all.

The short answer

Under United States law as it stands in March 2026, copyright requires a human author. Material generated autonomously by an AI system, meaning output the machine determined on its own in response to a prompt, is not protected by copyright, cannot be registered, and belongs to no one in the copyright sense. That is the holding of Thaler v. Perlmutter, 130 F.4th 1039 (D.C. Cir. 2025), affirming Thaler v. Perlmutter, 687 F. Supp. 3d 140 (D.D.C. 2023), and it is the consistent position of the Copyright Office.

But almost no one in commercial practice is making works the way Stephen Thaler claimed to. Real workflows mix human and machine contributions, and for those the answer is more useful: the human-authored parts are protected, the machine-generated parts are not, and the line between them is drawn case by case by asking whether a person actually determined the expressive elements of the output. Writing a prompt, however detailed, has so far not been enough. Selecting among outputs, arranging them, editing them substantially, or feeding the system your own human-authored work in a way that remains visible in the result, usually is. The Office requires you to disclose and disclaim AI-generated material that is more than trivial when you register, and a registration that hides it is vulnerable under 17 U.S.C. § 411(b).

Ownership, as distinct from copyrightability, is mostly a matter of contract: who the human author is, whether she was an employee, what the platform's terms say, and what the commissioning agreement assigns. The parts of the output that no one can own under copyright can still be controlled through contract, trade secret, and trademark. The rest of this article explains how all of that works, and where the law has not yet made up its mind.

Why "authorship" is the whole fight

The Copyright Act protects "original works of authorship fixed in any tangible medium of expression." 17 U.S.C. § 102(a). Copyright "vests initially in the author or authors of the work." 17 U.S.C. § 201(a). The statute uses the word "author" dozens of times and never defines it. The definitions section, 17 U.S.C. § 101, defines "work made for hire," "joint work," "compilation," and "derivative work," all of which presuppose an author without saying what one is.

That silence is why the AI debate turns on a century and a half of case law rather than on a sentence of statute. Three strands matter.

Burrow-Giles: the camera is a tool, the photographer is the author

In 1884 the Supreme Court decided whether a photograph could be a "writing" of an "author" under the Constitution's Copyright Clause. The defendant had copied Napoleon Sarony's portrait of Oscar Wilde and argued that a photograph was a mere mechanical reproduction of what stood before the lens. The Court disagreed. It defined an author as "he to whom anything owes its origin; originator; maker," and held that Sarony was the author because he had posed Wilde, selected the costume and drapery, arranged the light and shade, and "evok[ed] the desired expression." Burrow-Giles Lithographic Co. v. Sarony, 111 U.S. 53 (1884).

Burrow-Giles is the foundation of every argument on both sides of the AI question. Proponents of broad protection read it as the Court blessing a new machine-made medium because a human stood behind the machine making creative choices, which is exactly what a prompt engineer does. The Copyright Office and the D.C. Circuit read it the same way but reach the opposite result: the photograph was protected because Sarony controlled the expressive elements, not because he pressed a button on a device that then decided for itself what the picture would look like. Both readings are faithful to the opinion. The disagreement is about which part of a modern AI workflow is "Sarony arranging the drapery" and which part is "the chemistry of the plate."

Feist: originality means independent creation plus a modicum of creativity

A century later the Court held that originality is the constitutional sine qua non of copyright, and that it requires "independent creation plus a modicum of creativity." The white pages of a telephone directory flunked that test because alphabetizing names is "so mechanical or routine as to require no creativity whatsoever." Feist Publications, Inc. v. Rural Telephone Service Co., 499 U.S. 340 (1991). Feist matters in two ways for AI. First, the bar is low: a small amount of genuine human creative choice in selecting, arranging, or modifying AI output will clear it. Second, the creativity must be the author's own. A machine's "choices" are not the user's independent creation, however impressive they are.

Naruto, Urantia, and Kelley: non-human sources are not authors

Three appellate decisions established, before anyone was prompting Midjourney, that copyright does not accommodate non-human creators. In Naruto v. Slater, 888 F.3d 418 (9th Cir. 2018), the Ninth Circuit held that a crested macaque that took selfies with a wildlife photographer's camera lacked statutory standing to sue for infringement, because the Copyright Act's references to an author's "children," "widow," and "heirs" make no sense applied to an animal. The court decided the case on standing rather than on authorship as such, but its reasoning about the statute's human vocabulary was lifted almost verbatim into Thaler.

In Urantia Foundation v. Maaherra, 114 F.3d 955 (9th Cir. 1997), the plaintiff claimed that a religious text had been dictated by celestial beings. The Ninth Circuit agreed that "it is not creations of divine beings that the copyright laws were intended to protect," but held that the human members of the group who had selected and arranged the revelations into a book had contributed enough to support a copyright in the compilation. That is the template the Copyright Office now applies to AI: the non-human source gets nothing, the humans get what they actually contributed.

And in Kelley v. Chicago Park District, 635 F.3d 290 (7th Cir. 2011), the Seventh Circuit held that a living wildflower garden was not a work of authorship because its form was determined largely by nature rather than by the artist who planted it. The court's observations that "authorship is an entirely human endeavor" (quoting the Patry treatise), that "gardens are planted and cultivated, not authored," and that a garden "owes most of its form and appearance to natural forces" map uncomfortably well onto a diffusion model that draws an image out of statistical noise while the user watches.

The Compendium: the Office's long-standing position

The Copyright Office codified the human-authorship requirement in its examining manual well before generative AI went mainstream. Section 306 of the Compendium of U.S. Copyright Office Practices (Third) states that the Office "will register an original work of authorship, provided that the work was created by a human being," and section 313.2 explains that the Office will not register works "produced by a machine or mere mechanical process that operates randomly or automatically without any creative input or intervention from a human author." Compendium (Third) §§ 306, 313.2. The examples in § 313.2 include works attributed to spirits, animals, and natural forces, and, tellingly, mechanical or automated processes. The Office's position in Thaler, in Zarya of the Dawn, and in its 2023 guidance is a straightforward application of language that has been in the manual since 2014.

The legal framework in one table

Thaler v. Perlmutter, explained

Stephen Thaler is a computer scientist who has spent years trying to get intellectual property offices around the world to recognize his machines as creators. On the patent side he lost in the Federal Circuit, which held that an "inventor" under the Patent Act must be a natural person. Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022); see our companion article on AI-generated inventions. On the copyright side he applied in 2018 to register a two-dimensional image called A Recent Entrance to Paradise, listing the "Creativity Machine" as author and himself as owner by virtue of owning the machine. He told the Office, in terms, that the work had been created autonomously and that he had made no creative contribution. The Office refused, twice on reconsideration, and Thaler sued under the Administrative Procedure Act.

The district court (August 2023)

Judge Beryl Howell granted summary judgment to the Office. Her opinion is short and categorical: "human authorship is a bedrock requirement of copyright." Thaler v. Perlmutter, 687 F. Supp. 3d 140 (D.D.C. 2023). She traced the requirement from Burrow-Giles through Feist to the modern statute, reasoned that the Copyright Act's purpose of incentivizing creation presupposes a human who can be incentivized, and noted that every case she could find in which copyright had been recognized involved a human "originator." She also observed that the harder questions, such as how much human input is required when AI is used as a tool, were not before her, because Thaler had framed his case as one of pure machine authorship. And she held that Thaler could not rescue the application by arguing on appeal that he was the author after all, because he had told the Office the opposite and the APA confines review to the administrative record.

The D.C. Circuit (March 2025)

A unanimous panel affirmed. Thaler v. Perlmutter, 130 F.4th 1039 (D.C. Cir. 2025). Judge Millett's opinion rests on statutory text rather than on the Constitution. The Copyright Act, she wrote, is shot through with provisions that make sense only if an author is a human being: copyright lasts for the author's life plus seventy years (17 U.S.C. § 302), and machines do not have lives; authors have "widows" and "children" who inherit termination rights; transfers must be in a "signed" writing (17 U.S.C. § 204), and machines do not sign; authors have "nationality" and "domicile"; and authors have "intentions," which joint-work analysis requires. A machine, the court said, is a tool, and "the Creativity Machine cannot be the recognized author of a copyrighted work because the Copyright Act of 1976 requires all eligible work to be authored in the first instance by a human being."

Three features of the opinion matter for practitioners. First, the court was careful to say that the human-authorship rule does not prohibit copyright in works made with AI, and it cited the Office's own practice of registering such works when a human contributes enough. Second, it rejected the work-made-for-hire argument that Thaler, as the machine's owner, should be treated as its employer, because the doctrine assumes an employee who is an author in the first place. Third, it held that Thaler had forfeited any argument that he was personally the author through his design and operation of the machine, because he had told the Office the opposite throughout the administrative process. The court expressly did not decide whether such an argument would succeed if properly raised, and that reservation is the seam along which future cases will be litigated.

The Supreme Court denied certiorari on March 2, 2026. As of this writing, Thaler is the only federal appellate decision on the question. It is binding precedent in the D.C. Circuit, where many challenges to registration refusals are brought, and persuasive authority everywhere else. Jason Allen's challenge, discussed below, was filed in Colorado, where the Tenth Circuit has not spoken.

What Thaler did not decide

Because Thaler built his case as a test of pure machine authorship, neither court had to say how much human involvement is enough. That question is being answered, for now, not by courts but by the Copyright Office in its registration decisions and guidance, and by a single pending district court case brought by Jason Allen. Anyone who tells you that Thaler "held that AI art cannot be copyrighted" is overreading it. It held that art made by a machine with no human creative input cannot be copyrighted. The entire commercial question lives in the space that sentence leaves open.

The Copyright Office guidance: March 2023 and the 2025 Part 2 report

The March 2023 registration guidance

On March 16, 2023, three weeks after the Zarya of the Dawn decision discussed below, the Office published a statement of policy, Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence, 88 Fed. Reg. 16,190 (Mar. 16, 2023). It is short, and every applicant who has touched a generative tool should read it. The core points:

1. The test is control over expression. The Office asks "whether the 'work' is basically one of human authorship, with the computer merely being an assisting instrument, or whether the traditional elements of authorship in the work (literary, artistic, or musical expression or elements of selection, arrangement, etc.) were actually conceived and executed not by man but by a machine." That language comes from the Register's annual report for fiscal year 1965, which tells you how long the Office has been thinking about this.
2. Prompts alone are not enough. When an AI "receives solely a prompt from a human and produces complex written, visual, or musical works in response," the Office's view is that the traditional elements of authorship were determined by the machine. It compared a prompt to instructions given to a commissioned artist: the person giving instructions is not the author of what the artist paints.
3. Human contributions to AI-containing works can be claimed. Selecting or arranging AI-generated material in a sufficiently creative way, or modifying it to a degree that meets the originality standard, produces protectable human authorship, but protection extends only to the human-authored aspects, which are "independent of" and "do not affect" the copyright status of the AI-generated material.
4. Applicants must disclose. Applicants have a duty to disclose AI-generated content that is "more than de minimis," to describe the human contribution in the "Author Created" field, and to exclude the AI material in the "Limitation of Claim" section. The guidance ties this to the applicant's obligations under 17 U.S.C. § 409 and warns that an inaccurate registration may be cancelled or disregarded in litigation under § 411(b).
5. Existing registrations must be corrected. Owners of registrations that fail to disclose AI content were told to file supplementary registrations under 37 C.F.R. § 202.6.

The Office followed with a Notice of Inquiry on artificial intelligence and copyright in August 2023 (88 Fed. Reg. 59,942 (Aug. 30, 2023)), which drew more than ten thousand comments, and then a multi-part report.

The January 2025 Part 2 report on copyrightability

Part 2 of the Office's Copyright and Artificial Intelligence report, released January 29, 2025, is the most detailed statement of its position and is the document a court will be handed when the next registration refusal is challenged. (Part 1, from July 2024, addressed digital replicas; a pre-publication version of Part 3, on training, appeared in May 2025.) The Part 2 conclusions, in substance:

• Existing copyright principles are adequate; the Office does not recommend legislation creating a new right in AI output.
• Using AI as a tool to assist creation does not affect copyrightability, so long as a human determined the expressive elements. Using AI to brainstorm, research, outline, correct, or polish is assistance.
• Prompts, standing alone, do not provide sufficient control to make the user the author of the output, because the same prompt produces different results and the user does not determine how the system translates words into expression. The report went out of its way to say that this is true even of long, detailed, and iterated prompts, while acknowledging that a sufficiently creative prompt may itself be protected as a literary work (the protection covers the prompt's own text, not what the machine makes from it) and leaving the door open as the technology changes.
• Human-authored inputs that are perceptible in the output can be claimed. If a user feeds in her own drawing or text and the AI's output visibly incorporates it, the human expression in the output remains hers (the Office used the example of an artist's sketch rendered by an AI tool).
• Creative selection, coordination, and arrangement of AI output, and creative modification of it (inpainting, compositing, substantial editing), can support a claim in those contributions.
• The analysis is case-by-case and the Office will not adopt a bright-line percentage or word count.

The report's authority is persuasive rather than binding. Courts give Copyright Office interpretations the weight their reasoning earns under Skidmore v. Swift & Co., 323 U.S. 134 (1944), and since Loper Bright Enterprises v. Raimondo, 603 U.S. 369 (2024), no agency receives Chevron deference anyway. But the Office is the gatekeeper for registration, its examiners apply the report every day, and judicial review of a refusal is for abuse of discretion. See Atari Games Corp. v. Oman, 888 F.2d 878 (D.C. Cir. 1989) (remanding a refusal to register the video game Breakout for a fuller explanation); Atari Games Corp. v. Oman, 979 F.2d 242 (D.C. Cir. 1992) (holding on the second trip that the Register had applied too demanding a standard); Darden v. Peters, 488 F.3d 277 (4th Cir. 2007) (abuse-of-discretion review of a refusal to register maps). For practical purposes the Office's view is the law of registration until a court says otherwise.

One institutional footnote. The Register who signed the 2023 guidance and the 2025 report, Shira Perlmutter, was removed from office in May 2025 and litigated her removal through the summer and fall, see Perlmutter v. Blanche (D.C. Cir. Sept. 10, 2025). None of that affects the guidance, which remains the operative statement of Office practice and which examiners continue to apply.

Zarya of the Dawn and Théâtre D'opéra Spatial: the two decisions that set the lines

Zarya of the Dawn (February 2023)

Kris Kashtanova wrote and assembled an eighteen-page comic, Zarya of the Dawn, using Midjourney for the images. The Office initially registered the whole work in September 2022 without knowing about the AI. When Kashtanova's social media posts about the process came to its attention, it reopened the file. In a February 21, 2023 letter, the Office cancelled the original certificate and issued a new one limited to the human-authored elements: the text, and the "selection, coordination, and arrangement" of the written and visual elements. The individual images were excluded.

The reasoning repays attention because it is the template for everything since. Kashtanova's lawyers explained that the images came out of hundreds of prompts and many rounds of iteration for each panel. The Office accepted all of that and still found that Midjourney, not Kashtanova, "originated the traditional elements of authorship" in each picture, because the system's process of starting from noise and refining toward an image was not controlled by the user in the way a camera is controlled by a photographer. Prompts, in the Office's words, "function closer to suggestions than orders." The Office also considered Kashtanova's Photoshop edits to particular images and found them either too minor to be perceptible or insufficiently documented, while noting that substantial edits could qualify. And it noted, with evident care, that its decision did not question Kashtanova's creativity, only who determined the expressive content of the images.

Théâtre D'opéra Spatial (September 2023)

Jason Allen's case is the mirror image: a single image rather than a compilation, and an applicant who refused to disclaim. Allen told the Office that he had used at least 624 prompts and revisions in Midjourney, then edited the result in Photoshop and upscaled it with Gigapixel AI. The examiner asked him to exclude the Midjourney material. He refused, arguing that his prompting was the creative act. The Review Board's September 5, 2023 decision affirmed the refusal: the Midjourney image was not the product of human authorship; the Photoshop edits might be, but Allen had not described them with enough specificity to let the Office evaluate them; and the Gigapixel upscaling added nothing authorial. Because Allen would not limit his claim, the Office refused to register anything.

Allen sued in the District of Colorado in September 2024 (Allen v. Perlmutter, No. 1:24-cv-02665), arguing that the Office's prompt-based rule is contrary to Burrow-Giles and that his iterative process was the kind of control the law requires. As of this writing the case has not produced a final merits decision that we can report, and readers should check the docket before relying on anything about it. It is the vehicle most likely to give the courts a chance to refine Thaler.

The other Review Board decisions

Two more are worth knowing. In SURYAST (Dec. 11, 2023), Ankit Sahni fed his own photograph of a sunset into an AI tool (RAGHAV) with a Van Gogh painting as a style reference, and the Board refused registration because the tool, not Sahni, determined how the style was applied; the Board treated his photograph as a human input but found the output's expressive elements machine-determined. And in a registration decision widely discussed in early 2025, the Office registered A Single Piece of American Cheese, an image built through dozens of rounds of AI inpainting in which the user selected regions and directed each regeneration, on the theory that the selection, coordination, and arrangement of the AI-generated pieces was itself a human creative act. The registration is limited to that contribution, not to the underlying pixels, but it shows the path the Office is willing to recognize.

What "sufficient human authorship" actually looks like

Pulling the Office's decisions, the Part 2 report, and the pre-AI case law together, here is how the question is being answered in practice.

The pre-AI cases that supply the vocabulary

The Office did not invent its control-based test. The cases on what counts as authorship when a human directs a process were decided over decades.

In Aalmuhammed v. Lee, 202 F.3d 1227 (9th Cir. 2000), the Ninth Circuit held that an author is "the person who superintends the work," the "master mind" who exercises control, and that contributing ideas, suggestions, and even some expression to Spike Lee's Malcolm X did not make a consultant a joint author. A prompt engineer who offers suggestions to a system that then does the work is, on this view, in Aalmuhammed's shoes. Garcia v. Google, Inc., 786 F.3d 733 (9th Cir. 2015) (en banc), rejecting an actress's claim to a copyright in her five-second performance within a film, made a related point: copyright does not fragment a work into every contributor's separately owned piece.

On modification, Alfred Bell & Co. v. Catalda Fine Arts, Inc., 191 F.2d 99 (2d Cir. 1951), holds that a "distinguishable variation" from a prior work, even an unintentional one, can be original; that is the low bar a human editor of AI output must clear. But the Seventh Circuit in Gracen v. Bradford Exchange, 698 F.2d 300 (7th Cir. 1983), and later in Schrock v. Learning Curve International, Inc., 586 F.3d 513 (7th Cir. 2009), cautioned that derivative-work originality requires a variation that is more than trivial, and Schrock confirmed that accurate product photographs clear that bar because of the photographer's choices about lighting, angle, and exposure.

On faithful reproductions, Bridgeman Art Library, Ltd. v. Corel Corp., 36 F. Supp. 2d 191 (S.D.N.Y. 1999), denied protection to slavish photographic copies of public-domain paintings, and Meshwerks, Inc. v. Toyota Motor Sales U.S.A., Inc., 528 F.3d 1258 (10th Cir. 2008), denied protection to digital wire-frame models of Toyota vehicles that sought only to depict the cars accurately. Meshwerks is the most AI-relevant of the group: the court held that the sheer labor of building the models, and the modeler's technical skill, did not supply originality where the output's appearance was dictated by the subject and the tool. Upscaling, denoising, and "enhance" buttons sit in the same category.

On the photographer's three sources of originality (rendition, timing, and creation of the subject), Mannion v. Coors Brewing Co., 377 F. Supp. 2d 444 (S.D.N.Y. 2005), and Rentmeester v. Nike, Inc., 883 F.3d 1111 (9th Cir. 2018), supply the analytical frame that AI proponents invoke: the prompter, they say, creates the subject the way Rentmeester posed Michael Jordan. The Office's answer is that Rentmeester also controlled the rendition (the lens, the angle, the light, the instant the shutter opened), while the prompter describes a subject and then controls neither its precise depiction nor its rendition.

And Satava v. Lowry, 323 F.3d 805 (9th Cir. 2003), reminds everyone that even where protection exists, it is thin: a glass-in-glass jellyfish sculpture got protection only against virtually identical copying, because most of its features were dictated by the subject. A registration limited to "selection and arrangement" of AI images will be thin in the same way.

How to disclaim AI-generated material in an application

The mechanics are simple, and the Office's examiners now expect to see them. For the full procedure see our companion guide, Registering and Protecting Works Made With Generative AI: A Practical Guide, and the AI-Assisted Work Copyright Registration Checklist. In outline:

1. In the Author Created (or "Nature of Authorship") field, describe only the human contribution, in plain words: "text," "selection, coordination, and arrangement of images," "editing and compositing of AI-generated imagery," "human-authored musical composition." Do not write "AI-assisted artwork" there; that describes the whole, not your part.
2. In the Limitation of Claim section, in the "Material Excluded" field, state what is excluded, for example "AI-generated images" or "text generated by artificial intelligence." The Office's guidance asks for a brief description; it does not require naming the tool, though doing so is harmless and helps the record.
3. In the New Material Included field, repeat the human contribution.
4. Where the AI material is de minimis (an AI-suggested synonym, a grammar fix, a background texture no one would notice), no disclosure is required. "De minimis" here means too small to be independently copyrightable, not "small relative to the whole."
5. Keep the evidence. The Office can and does ask for prompt logs, edit histories, and layered files when a claim describes modifications. Allen lost partly because he could not or would not specify what Photoshop added.

The payoff of doing this correctly is a registration that is exactly as strong as your real contribution, which is the only registration that will survive contact with a defendant.

The duty of candor: Unicolors, Section 411(b), and what a limited registration costs you

Section 411(b) and Unicolors

A certificate of registration satisfies the suit prerequisite in 17 U.S.C. § 411(a) regardless of inaccuracies in the application unless two things are both true: the inaccurate information was included "with knowledge that it was inaccurate," and the inaccuracy, "if known, would have caused the Register of Copyrights to refuse registration." 17 U.S.C. § 411(b)(1). When a defendant raises the issue, the court must ask the Register whether the inaccuracy would have mattered. § 411(b)(2).

In Unicolors, Inc. v. H&M Hennes & Mauritz, L.P., 595 U.S. 178 (2022), the Supreme Court held that "knowledge" means actual knowledge of the inaccuracy, and that a good-faith mistake of law (there, about what counts as a "single unit of publication") is as much a defense as a mistake of fact. The Court noted that willful blindness may support a finding of knowledge and that circumstantial evidence can prove it. The Eleventh Circuit had earlier read the provision the same way, requiring scienter rather than mere error, in Roberts v. Gordy, 877 F.3d 1024 (11th Cir. 2017), while the Ninth Circuit in Gold Value International Textile, Inc. v. Sanctuary Clothing, LLC, 925 F.3d 1140 (9th Cir. 2019), had held that no intent to defraud is required once knowledge is shown. Archie MD, Inc. v. Elsevier, Inc., 261 F. Supp. 3d 512 (S.D.N.Y. 2017), shows the referral procedure in action: the court asked the Register about an alleged misstatement of a work's publication status, received an answer that did not resolve the question, and refused to invalidate the registration because the applicant had not known the characterization was wrong on what was then an unsettled point of law. Courts differ on timing. The Seventh Circuit, warning that referral can become a delay tactic, lets district courts require the challenger to show an inaccuracy made with knowledge before the court involves the Register at all. DeliverMed Holdings, LLC v. Schaltenbrand, 734 F.3d 616 (7th Cir. 2013). Other courts refer as soon as the issue is properly raised, which can add months to a case.

What this means for AI. Before March 2023, an applicant who did not disclose AI content could plausibly claim ignorance of any duty to do so. After the guidance, and especially after the 2025 report, that claim gets harder every month. An applicant who used Midjourney for the images, read nothing, and claimed "artwork" has a Unicolors argument. An applicant who was advised about the guidance and chose not to disclose does not, and a court will find that the Register would have refused to register the AI material if told. The registration is then disregarded for purposes of § 411(a), the case is dismissed or stayed, and the plaintiff files again with a corrected claim, having lost the earlier effective date and, under 17 U.S.C. § 412, possibly the right to statutory damages and fees under § 504(c) and § 505. Knowingly false statements in an application are also a crime under 17 U.S.C. § 506(e), and false statements to a federal agency generally fall under 18 U.S.C. § 1001. Prosecutions are vanishingly rare; the civil consequences are not.

What a registration that is later limited is worth

Suppose you registered honestly, the Office limited the claim to your selection and arrangement, and someone copies the work. What do you have?

• Prima facie validity, under 17 U.S.C. § 410(c), but only for what the certificate covers. The defendant will point to the Limitation of Claim and argue that the images it copied are precisely what you disclaimed.
• A thin copyright. Infringement of a compilation requires copying the selection and arrangement, not just the elements. A defendant who lifts three of your forty AI images and uses them in a different order has copied nothing you own. See Feist, 499 U.S. 340, and, for the compilation infringement standard, Eckes v. Card Prices Update, 736 F.2d 859 (2d Cir. 1984). For how similarity is actually litigated, see Proving Copyright Infringement: Access, Substantial Similarity, and the Tests Courts Actually Use.
• Statutory damages and fees for infringement of the protected part, if the registration was timely under § 412. The disclaimed material carries no statutory damages because it carries no copyright.
• Access to the Copyright Claims Board for a smaller claim, where a limited registration is still a registration; see The Copyright Claims Board: Small Claims Copyright Litigation Under the CASE Act.
• Nothing against someone who copies only the AI images. That is the hard truth. If the commercial value of the work sits in the pictures, and the pictures were machine-generated, the copyright system does not protect that value. Contract, trademark, and trade secret have to do the work, as discussed below.

A defendant's playbook follows from this. Pull the registration record (the Office's public catalog shows the Limitation of Claim). Serve discovery on the creation process: prompt logs, platform account histories, layered files, time stamps. Compare what was claimed to what was done. If the claim overreaches and the plaintiff knew the rules, move under § 411(b)(2) for referral to the Register. If the claim is honest but thin, argue that what was copied is the unprotected material. Both moves are cheap relative to a merits defense.

Who owns AI-assisted work: employees, contractors, commissions, and platform terms

Copyrightability asks whether there is anything to own. Ownership asks who owns it. The second question is governed by the same rules that applied before anyone had heard of a diffusion model, applied only to the human-authored layer of the work.

The default: the human author owns the human contribution

Copyright vests initially in the author, 17 U.S.C. § 201(a), and in an AI-assisted work the author is the person whose creative choices are perceptible in the result: the writer of the text, the illustrator who painted over the output, the editor who selected and arranged. No one acquires an interest by operating the tool, paying for the subscription, or owning the computer. Thaler tried the last of these theories (he owned the machine, so he owned what it made), and the D.C. Circuit rejected it because there was no author from whom ownership could flow.

Employees: work made for hire still works, for the human part

If the human author is an employee acting within the scope of employment, the work is a "work made for hire," and the employer is both author and owner from the moment of creation. 17 U.S.C. § 101; 17 U.S.C. § 201(b). Whether someone is an employee is decided under the common law of agency, weighing the hiring party's right to control the manner and means of creation, the skill required, the source of the tools, the location of the work, the duration of the relationship, the method of payment, benefits, and tax treatment. Community for Creative Non-Violence v. Reid, 490 U.S. 730 (1989). The Second Circuit gives benefits and tax treatment special weight, Aymes v. Bonelli, 980 F.2d 857 (2d Cir. 1992), while the Ninth Circuit held that a startup's programmer was an employee despite informal arrangements because the company controlled his work and treated him as staff in substance. JustMed, Inc. v. Byce, 600 F.3d 1118 (9th Cir. 2010).

For AI-assisted work, the doctrine operates normally on the human layer and has nothing to operate on beyond it. An employee who uses an image generator to rough out a campaign, then selects, composites, and retouches the results, produces a work made for hire to the extent of the human contribution. The employer owns that layer, and the machine-generated pixels belong to no one. A company that tells its staff "everything you make with our AI tools belongs to us" is right about the copyright that exists and cannot conjure copyright that does not.

Two practical wrinkles. First, scope of employment matters more than it used to, because employees use personal AI accounts for work and work accounts for side projects. An employee's after-hours novel drafted on a company-paid chatbot seat is not obviously within the scope of employment, and the platform's terms do not answer the question. Second, policies on employee AI use should address ownership and records together, because the documentation the Copyright Office wants (what the human did) is the same documentation that establishes what the employer owns.

Contractors and commissioned works: nine categories and a timing split

Work by an independent contractor is a work made for hire only if it falls within one of nine statutory categories (a contribution to a collective work, part of a motion picture or other audiovisual work, a translation, a supplementary work, a compilation, an instructional text, a test, answer material for a test, or an atlas) and the parties sign a written instrument saying so. 17 U.S.C. § 101. A stand-alone illustration, a logo, marketing copy, and most software fit none of these, so a "work for hire" clause in a freelance contract for those deliverables fails as a work-for-hire designation and works, if at all, only as an assignment.

The circuits divide on timing. The Seventh Circuit requires the written work-for-hire agreement to precede creation of the work. Schiller & Schmidt, Inc. v. Nordisco Corp., 969 F.2d 410 (7th Cir. 1992). The Second Circuit accepts a writing signed afterward if it confirms an agreement, express or implied, reached before the work was made. Playboy Enterprises, Inc. v. Dumas, 53 F.3d 549 (2d Cir. 1995). AI-heavy pipelines generate deliverables fast and paperwork slowly, so the safe course in every circuit is the belt-and-suspenders clause: the work is made for hire to the extent the law allows, and to the extent it is not, the contractor assigns all right, title, and interest. An assignment must be in a signed writing. 17 U.S.C. § 204(a). Without one, the commissioning party usually ends up with only an implied nonexclusive license, as the producer in Effects Associates, Inc. v. Cohen, 908 F.2d 555 (9th Cir. 1990), learned after paying for special-effects footage without a written transfer. See also Asset Marketing Systems, Inc. v. Gagnon, 542 F.3d 748 (9th Cir. 2008) (implied, irrevocable license to use and modify contractor-written software).

AI adds a diligence problem the old forms never contemplated. A contractor may deliver work that is mostly machine-generated, and an assignment of "all copyright" in it may convey very little. Commissioning parties now ask contractors to disclose AI use, to identify which elements are human-authored, and to keep process records, and they price the deliverable accordingly. The Work Made for Hire Determination Checklist walks through the categories and the timing rules, and the Generative AI Authorship Toolkit supplies model clauses.

Collaborators: AI is never a joint author, and prompting may not make you one

A "joint work" is one prepared by two or more authors "with the intention that their contributions be merged into inseparable or interdependent parts of a unitary whole." 17 U.S.C. § 101. A machine cannot form that intention, which is one of the reasons the D.C. Circuit gave in Thaler, so listing an AI system as a co-author on an application is an error the Office will reject.

The more interesting question is what happens among the humans. The Second and Seventh Circuits require each joint author to contribute independently copyrightable expression and to share an intent to be co-authors. Childress v. Taylor, 945 F.2d 500 (2d Cir. 1991); Thomson v. Larson, 147 F.3d 195 (2d Cir. 1998) (dramaturg on Rent was not a joint author); Erickson v. Trinity Theatre, Inc., 13 F.3d 1061 (7th Cir. 1994). The Ninth Circuit requires the same copyrightable contribution and then asks about control, objective manifestations of shared intent, and whether the work's audience appeal turns on each contribution. Aalmuhammed v. Lee, 202 F.3d 1227 (9th Cir. 2000). And the Second Circuit has refused to let a contributor to a unitary work carve a separate copyright out of his piece of it. 16 Casa Duse, LLC v. Merkin, 791 F.3d 247 (2d Cir. 2015).

Put those rules together and a team member whose only contribution was writing prompts or running the generator may have contributed nothing that counts, and may therefore own nothing, while the teammate who wrote the text and arranged the pages is a sole author. That result surprises people, so teams should settle ownership in writing rather than by assumption. For the general rules, see Copyright Ownership, Joint Authorship, and Termination of Transfers.

Platform terms: "you own the output" is a promise, not a copyright

Every major generative platform addresses output ownership in its terms, and most do it in roughly the same way: the provider assigns to the user whatever right, title, and interest the provider has in the output, often with the telling qualifier "if any," and reserves a license to use inputs and outputs to operate (and sometimes to improve) the service. Some tie the user's rights to a paid tier or to company size, and some warn that similar outputs may be generated for other users. Enterprise agreements commonly add a commitment not to train on customer data and an indemnity against third-party infringement claims, conditioned on the customer using the provider's safety filters and not deliberately prompting for infringing material. Terms change often, so read the version in force when the work was made and keep a copy.

Three consequences follow. First, the assignment cannot create copyright. If the output is unprotected, the user receives a contractual promise that the provider will not claim it, and nothing enforceable against strangers. Second, the platform's license-back and data-use terms matter more than its ownership clause for anyone hoping to keep prompts, uploaded drafts, or unreleased outputs confidential, because a secret shared under terms that permit broad provider use is harder to defend as a secret. Third, the indemnity, not the ownership clause, is where the money is, and its conditions are where it fails. We cover those terms in Licensing Data and Models for Artificial Intelligence: Training Rights, Outputs, and Indemnities and the negotiation in Negotiating an AI Data or Model License: A Practical Guide.

Duration and termination, briefly

The human layer of an AI-assisted work lasts for the author's life plus seventy years or, for a work made for hire, ninety-five years from publication or 120 years from creation, whichever ends first. 17 U.S.C. § 302. Grants by individual authors remain terminable after thirty-five years under 17 U.S.C. § 203, which does not reach works made for hire, and termination recaptures only the human-authored expression, which is all the author ever had. See Copyright Termination and Reversion: Sections 203 and 304 and Taking Back a Grant.

Protecting what copyright will not: contracts, trade secrets, and trademarks

When the commercially valuable layer of a work is machine-generated, the owner needs other tools. None is a perfect substitute, and each has a limit that a careful opponent will find.

Contracts bind the parties, and only the parties

A license or terms of use can forbid a counterparty from copying, redistributing, or training on material whether or not it is copyrighted. The Seventh Circuit enforced exactly that kind of restriction on an uncopyrightable telephone database in ProCD, Inc. v. Zeidenberg, 86 F.3d 1447 (7th Cir. 1996), reasoning that contract rights, which bind only those who agree to them, are not "equivalent" to copyright's rights against the world and so escape preemption under 17 U.S.C. § 301. The Second and Sixth Circuits reach similar results under the "extra element" test when the contract adds something such as a promise to pay. Forest Park Pictures v. Universal Television Network, Inc., 683 F.3d 424 (2d Cir. 2012); Wrench LLC v. Taco Bell Corp., 256 F.3d 446 (6th Cir. 2001).

The limit is preemption when a contract claim is a copyright claim in disguise. Section 301 reaches state-law rights in works "within the subject matter of copyright," and the Second Circuit has held that this subject matter includes the unprotectable elements of works that fall within copyright's categories, so that states cannot grant copyright-like protection to what federal law leaves free. NBA v. Motorola, Inc., 105 F.3d 841 (2d Cir. 1997). Applying that principle, the Second Circuit held preempted a lyrics website's contract claims against Google over copied transcriptions of lyrics the site did not own, ML Genius Holdings LLC v. Google LLC, No. 20-3113 (2d Cir. Mar. 10, 2022) (summary order), and a federal court in California held preempted a platform's attempt to use its terms of service to stop the scraping of user posts, X Corp. v. Bright Data Ltd., No. 23-cv-03698-WHA (N.D. Cal. May 9, 2024). Whether uncopyrightable AI output is "within the subject matter of copyright" for preemption purposes has not been decided, but the logic of Motorola suggests it is, because an AI-generated image is a pictorial work in form even when it is unprotected in fact. The lesson: contracts protect AI output best as negotiated, bilateral licenses with payment, confidentiality, or use restrictions tailored to a relationship, and worst as clickwrap attempts to stop the whole world from copying. On scraping and contract claims generally, see Data Scraping After hiQ v. LinkedIn: Copyright, Contract, and Computer Fraud Claims.

Trade secrets protect the process, not the published output

Prompts, prompt libraries, system prompts, fine-tuning datasets, model weights, evaluation sets, workflow configurations, and unreleased outputs can all qualify as trade secrets if they derive independent economic value from not being generally known and the owner takes reasonable measures to keep them secret. 18 U.S.C. § 1839(3). The Defend Trade Secrets Act provides a federal civil action with injunctions and damages for misappropriation. 18 U.S.C. § 1836. For a studio whose edge is a carefully engineered pipeline rather than any single output, trade secret law is often a better fit than copyright: it protects the method that produces the look rather than the look itself, and it does not care whether a human authored anything.

Its limits are just as clear. Anything published is no longer secret, so the output drops out of trade secret protection the moment it ships. Secrecy has to be real: access controls, confidentiality agreements with staff and contractors, enterprise platform terms that bar training on inputs, and logs showing who had access. And a competitor who independently builds a similar pipeline, or reverse engineers a published output, has misappropriated nothing. For how the secret must be identified once litigation starts, see Identifying Trade Secrets in Litigation: Reasonable Particularity, Sequencing Discovery, and Protecting the Secret While You Sue, and for building the program beforehand, Building a Trade Secret Protection Program from Scratch.

Trademarks protect source, not expression

An AI-generated logo, mascot, or character name can function as a trademark, because trademark rights come from use of a designation to identify the source of goods or services, not from authorship. 15 U.S.C. § 1127; 15 U.S.C. § 1051. For a brand that designed its logo with an image generator, trademark is the main protection available, and it can be strong. But trademark law polices confusing uses of a source identifier, not copying of an image as art, and it cannot be bent into a substitute for copyright. In Dastar Corp. v. Twentieth Century Fox Film Corp., 539 U.S. 23 (2003), the Supreme Court held that "origin" in 15 U.S.C. § 1125(a) refers to the producer of the tangible goods, not the author of the ideas or communications embodied in them, so copying uncopyrighted creative material and selling it under one's own name is not, by itself, false designation of origin. A business built on AI-generated characters should register the names and key designs as trademarks where they function as brands and accept that the artwork may be free for others to use in non-confusing ways. The Trademark Registration Guide covers the filing.

Notices, credits, and copyright management information

A copyright notice is permitted and useful, 17 U.S.C. § 401, and a notice on a work that contains both human and AI material is not improper, because a notice identifies the owner of the work rather than claiming every pixel. A notice on material that is purely AI-generated is different. With fraudulent intent, it is a misdemeanor under 17 U.S.C. § 506(c); more commonly it hands a defendant a misuse or unclean-hands theme and costs the claimant credibility with the judge. Conversely, the human layer's copyright management information is protected: removing or falsifying it with the required knowledge violates 17 U.S.C. § 1202. For form and placement, see Copyright Notice: Form, Function, and Best Practices.

Outside the United States: the UK, the EU, China, and the rest

Copyright is territorial. Each country decides for itself what it protects, and the Berne Convention requires member states to treat foreign works as they treat their own, not to share a definition of authorship. The result is a patchwork, and a work that is unprotected in the United States may be protected in London or Beijing.

The United Kingdom's computer-generated works provision

Since 1988, UK law has provided that for a literary, dramatic, musical, or artistic work "which is computer-generated, the author shall be taken to be the person by whom the arrangements necessary for the creation of the work are undertaken." Copyright, Designs and Patents Act 1988, s. 9(3). A computer-generated work is one generated "in circumstances such that there is no human author of the work," s. 178, and it is protected for fifty years from the end of the year it was made, s. 12(7), without the moral rights of attribution and integrity. The leading decision, Nova Productions Ltd v. Mazooma Games Ltd [2007] EWCA Civ 219, held that the person who made the arrangements for frames displayed by a video game was its programmer, not the player pressing the buttons. Who makes the "arrangements" for a Midjourney image (the user, the developer, or both) has never been decided, and commentators doubt the provision fits the originality standard that UK courts absorbed from EU law. The government's December 2024 consultation on copyright and AI asked whether to keep, reform, or remove it. Until something changes, a UK business has an argument that does not exist in the United States, and a U.S. business with UK operations should not assume that what is unprotected at home is unprotected there.

The European Union: "the author's own intellectual creation"

EU copyright law has no computer-generated works provision. The Court of Justice defines originality as the "author's own intellectual creation," Infopaq International A/S v. Danske Dagblades Forening, Case C-5/08 (2009), achieved when the author makes "free and creative choices" that stamp the work with a "personal touch," Painer v. Standard VerlagsGmbH, Case C-145/10 (2011); see also Cofemel, Case C-683/17 (2019). That vocabulary of personality and choice is generally understood to require a human, and the early national decisions agree: in 2023 the Municipal Court in Prague refused protection to an image the claimant had generated from his own prompt. The EU AI Act, Regulation (EU) 2024/1689, leaves authorship rules alone. It requires providers of generative systems to mark synthetic output in a machine-readable way and deployers to disclose deepfakes (Article 50, scheduled to apply from August 2026), and it requires providers of general-purpose models to adopt a copyright-compliance policy and publish a summary of their training content (Article 53). For those obligations, see AI Governance and Compliance: Building a Program for the EU AI Act, State AI Laws, and Everything After.

China: the Beijing Internet Court goes the other way

In Li v. Liu (Beijing Internet Court, Nov. 27, 2023), the plaintiff generated a portrait with Stable Diffusion, posted it online, and sued a blogger who used it without permission and stripped the watermark. The court held the image protected and the plaintiff its author, reasoning that his choice and ordering of prompts, his selection of parameters, and his repeated adjustment of the output reflected intellectual input and personalized expression, and that the model was a tool much like a camera. It awarded 500 yuan and an apology. The decision is the clearest judicial rejection anywhere of the U.S. Office's view of prompts, but Chinese courts decide these cases on their facts and have treated the user's documented creative process as decisive rather than presuming authorship from use of the tool. Chinese courts have also been willing to hold AI services liable for infringing outputs, as the Guangzhou Internet Court did in early 2024 over generated images of the Ultraman character.

Elsewhere, and why it matters to a U.S. client

Japan's Agency for Cultural Affairs issued guidance in 2024 treating the copyrightability of AI-generated material as a case-by-case question of creative intention and creative contribution, with factors such as the specificity of instructions, the number of attempts, selection among outputs, and later human edits. Australian courts denied protection to computer-generated directories and source code years before generative AI, in Telstra and Acohs, because no identifiable human author had produced the expression. And the SURYAST image that the U.S. Office refused was registered in Canada, and briefly in India, with the AI tool listed as co-author, registrations that have since drawn a court challenge in Canada and a withdrawal notice in India. For the parallel story in patent law, where the DABUS applications failed almost everywhere, see Artificial Intelligence and Inventorship: Global Perspectives on Machine Contributions to Innovation.

For a multinational business, the practical point is that protection follows the place of exploitation. In a U.S. court, ownership of a foreign work may be decided under the law of the country most closely connected with it, Itar-Tass Russian News Agency v. Russian Kurier, Inc., 153 F.3d 82 (2d Cir. 1998), but whether protectable expression exists and whether it was infringed in the United States are questions of U.S. law. A UK computer-generated work therefore gets no U.S. protection for its machine-made layer, and a Chinese judgment for an AI image means nothing in New York.

Strategy for each side, and what it costs

If you make or commission AI-assisted work and want it protected

1. Put the valuable expression in human hands. Decide at the start of a project which layer carries the commercial value (story, characters, composition, code architecture, layout) and make sure a human authors that layer. The Office's case-by-case test rewards workflows in which the machine drafts and humans decide.
2. Keep the record as you go. Version histories, layered source files, prompt logs, and short contemporaneous notes on what the human changed are the evidence the Office asks for and the evidence a court will want. Reconstructing them later is expensive and unpersuasive. The Generative AI Authorship Toolkit includes a contribution-log template.
3. Register the human layer early. Registration before infringement, or within three months of publication, preserves statutory damages and attorney's fees under 17 U.S.C. § 412, and registration within five years of publication carries the presumption of validity under 17 U.S.C. § 410(c). A limited registration is still worth having.
4. Disclose accurately and narrowly. Claim what the humans did, exclude what the machine did, and do not try to slip AI material through. The price of overclaiming is the registration itself.
5. Paper the ownership. Employment and contractor agreements should cover AI-assisted work, require disclosure of AI use, and assign rights. Commissioning agreements should allocate the risk that a deliverable is partly unprotectable and say who bears it.
6. Layer the other protections. Trade secret for the pipeline, trademark for the brand, contract for counterparties, and speed to market for everything else.

If you are accused of copying an AI-assisted work

1. Pull the registration and the deposit. The public catalog shows the Limitation of Claim, and the deposit copy can be obtained for litigation. Compare what the plaintiff claimed with what you used.
2. Filter before you compare. Courts remove unprotectable elements before comparing works, whether through the abstraction-filtration-comparison method developed for software in Computer Associates International, Inc. v. Altai, Inc., 982 F.2d 693 (2d Cir. 1992), or the Ninth Circuit's extrinsic test, Skidmore v. Led Zeppelin, 952 F.3d 1051 (9th Cir. 2020) (en banc). AI-generated material belongs on the unprotectable side of that line, alongside ideas, scènes à faire, and public-domain material, and what remains may be thin, protected only against virtually identical copying as in Satava. The mechanics are in Litigating Substantial Similarity in a Copyright Case: A Practical Guide.
3. Take discovery on the process. Prompt logs, platform account histories, layered files, and timestamps are discoverable under Fed. R. Civ. P. 26(b)(1) and Fed. R. Civ. P. 34, and a plaintiff who let them disappear after a dispute arose faces sanctions under Fed. R. Civ. P. 37(e). Platform exports must be authenticated under Fed. R. Evid. 901.
4. Consider independent creation. Two users giving similar prompts to the same model can get similar outputs without either copying the other, and similarities traceable to a model's defaults are evidence of the tool, not of copying. Expert testimony on how the model behaves must satisfy Fed. R. Evid. 702.
5. Use § 411(b) when the facts support it. If the plaintiff knew the rules and claimed AI material anyway, seek referral to the Register. If the claim is honest but limited, argue that what you took is what was disclaimed.
6. Do not overplay it. Using an AI tool does not make a work unprotected. A defendant who copies a human-written novel with an AI-generated cover has copied a protected novel, and judges notice when a defense rests on the cover.

Costs and timing

Fees are set in 37 C.F.R. § 201.3 and change from time to time, so check the current schedule before filing. The cheapest item on the list, by a wide margin, is a careful first application.

Mistakes we see on both sides

• Listing the AI tool or its developer as an author or co-author.
• Describing the human contribution as "AI-assisted artwork" or "prompt engineering," which tells the examiner nothing claimable.
• Filing the Single Application, which has no way to exclude material, for a work that contains AI material that must be excluded.
• Skipping disclosure because the human contribution feels larger than the machine's.
• Disclaiming too much, for example excluding human-edited material along with the raw output.
• Deleting prompts, drafts, and layered files once the work ships.
• Relying on a freelancer's "work for hire" clause for logos, illustrations, or copy that fit no statutory category.
• Treating a platform's "you own the output" clause as a copyright.
• Sending takedown notices for purely AI-generated material. A notice must come from the owner of an exclusive right or its agent, 17 U.S.C. § 512(c)(3), and a knowing material misrepresentation exposes the sender to damages under § 512(f); see Lenz v. Universal Music Corp., 815 F.3d 1145 (9th Cir. 2016) (sender must consider fair use before sending). The mechanics are in How to File a DMCA Takedown Notice (And Respond to One).
• On the defense side, assuming that any AI involvement defeats the claim.

Worked examples: a graphic novelist, a game studio, and a marketing agency

The graphic novelist

Mara Okafor is finishing Saltglass, a 140-page graphic novel about a lighthouse keeper on a drowned coast. She writes the script and all of the dialogue. She draws the four principal characters by hand on a tablet, in a style she developed over a decade. For backgrounds (seascapes, storm skies, ruined harbors) she uses an image generator, typically running thirty or forty prompts per panel and keeping one result. For a dozen splash pages she feeds her own pencil sketches into the generator's image-to-image mode and then paints over the result for hours. She letters, lays out, and paces every page herself.

What she owns. Her script and dialogue are fully protected. Her character drawings are protected pictorial works, and the characters, as consistently and distinctively drawn, are protectable elements of the book. The selection, coordination, and arrangement of text and images on every page is protected compilation authorship, exactly as in Zarya. The splash pages are protected to the extent her sketches remain perceptible in the result and her overpainting is more than trivial, under the Part 2 report's treatment of human inputs and modifications. The prompted backgrounds, one by one, are not hers.

How she registers. A Standard Application naming Mara as author and claimant. Author Created: "text; two-dimensional artwork (character illustrations and modifications to splash-page imagery); selection, coordination, and arrangement of text and images." Material Excluded: "AI-generated background images and AI-generated portions of splash-page imagery." She keeps the layered files for the splash pages and a one-page process memo saying which elements came from where, so that an examiner's question can be answered in a day rather than a month.

What happens when someone copies. A pirate who reprints the book wholesale infringes her text, characters, art, and arrangement, and she can sue on her registration. A print shop selling posters of three of her AI backgrounds infringes nothing of hers, and she should not send takedown notices about those posters. A writer who lifts her lighthouse keeper's look and personality for a new comic presents a character-infringement claim based on human-drawn work. And her publishing contract needs attention before any of that happens: a standard warranty that "the work is wholly original and the author owns all rights in it" should be amended to describe the AI elements accurately, or Mara is in breach of warranty on the day she signs.

The game studio

Copperline Games is a forty-person studio building a narrative adventure game. Its concept artists use image generators to explore ideas and then paint final assets by hand. Its environment team generates tileable textures with an AI tool and ships most of them as generated. Its writers draft the branching story dialogue, while a language model produces several thousand lines of ambient chatter for non-player characters, which a narrative designer reviews and edits lightly. Its engineers use a coding assistant and accept suggestions ranging from single lines to whole functions. Twelve of the forty are contractors working from three states.

What Copperline owns. The hand-painted art, the written story dialogue, the human-written code, and the game's audiovisual presentation as a whole, if the paperwork is right. Employees' contributions are works made for hire. Contractors' contributions qualify as works made for hire only through a statutory category (a contribution to an audiovisual work fits) and a signed agreement, which argues for signing before work starts, as the Seventh Circuit requires, with an assignment as backup everywhere. The AI textures and the ambient chatter, as generated, are not protected. The lightly edited chatter is a judgment call: fixing grammar and tone is unlikely to be enough, while rewriting lines to carry a character's voice may be. Code accepted verbatim from the assistant is unprotected, though much short functional code would be unprotectable anyway under the idea-expression line that runs back to Baker v. Selden, 101 U.S. 99 (1879).

How Copperline registers. One Standard Application for the game, claiming the human-authored audiovisual material, text, and computer program code, plus the selection and arrangement of all elements, and excluding "AI-generated textures, AI-generated non-player-character dialogue, and AI-generated source code." The exclusion need not inventory every texture, but it must be accurate. For the separate considerations in registering games and software, see Copyright Registration of Games and Copyright Registration of Computer Programs.

How Copperline protects the rest. The prompt libraries and the fine-tuned texture model are trade secrets, protected by access controls, confidentiality agreements, and an enterprise platform agreement that bars training on Copperline's inputs. The game's title and lead character's name are trademarks. The end-user license agreement forbids asset extraction, an enforceable promise against players who accept it under ProCD, though not against the world. If a competitor rips the textures from the game files, Copperline's best claims are contract (if the competitor accepted the license) and, if the look has become source-identifying, unfair competition. Copyright in the textures is not available, and a demand letter that pretends otherwise invites a declaratory judgment action.

The marketing agency

Brightwater Creative is a twelve-person agency producing a product-launch campaign for Halvorsen Outdoor. Brightwater's employees generate dozens of hero images with an image model, choose six, composite them with photographs shot by a staff photographer, and add typography and copy written by a staff copywriter. They also develop a new logo for the product line, refining it through many prompts and then redrawing it by hand as clean vector art. Halvorsen's master services agreement says every deliverable is a "work made for hire" owned by Halvorsen and that Brightwater warrants the deliverables are original and infringe no third-party rights.

What exists to transfer. Brightwater's employees create works made for hire for Brightwater, which owns the human layer: the photographs, the copy, the compositing and arrangement, and the redrawn logo to the extent the redrawing added original expression. The raw hero images are owned by no one. The MSA's work-for-hire language does not make the deliverables works made for hire for Halvorsen, because advertising materials fit none of the nine categories and Brightwater's employees are not Halvorsen's employees, so the clause works only if it also contains an assignment. The logo is mainly a trademark question: Halvorsen's rights in the mark will come from use and registration, and the copyright question matters chiefly to whether others may copy the artwork for non-trademark purposes.

What should change. The MSA should say that deliverables may include AI-generated material, that Brightwater assigns whatever rights it has, and that Brightwater does not warrant copyright ownership of AI-generated elements. The non-infringement warranty should be tied to the tools Brightwater uses and to the indemnities those vendors provide, because the realistic infringement risk in AI imagery comes from training data and memorized output, not from Brightwater's own copying. Halvorsen, for its part, should ask for a list of AI-generated elements, cooperation in registering the campaign's human layer, and a commitment that confidential launch materials go only into enterprise tools with no-training terms. For the infringement side of that risk, see Copyright Infringement Claims Against Generative AI: The New York Times, Getty, and What Comes Next, and for vendor diligence, the AI Vendor Procurement and Governance Checklist.

What is settled and what is still open (as of March 2026)

Two cautions about the right-hand column. First, the Office's position on prompts is expressly tied to "the functioning of current generally available technology," and tools that let users dictate composition, pose, and layout with precision (sketch-to-image, region-by-region inpainting, pose controls) are pushing workflows toward the control the Office says is missing. Second, a court that eventually disagrees with the Office will be deciding an APA case about one applicant's record. Do not reorganize a business around a single district court opinion.

Frequently asked questions

If I write a very detailed prompt, do I own the image?

Not under current Copyright Office practice. The Office's view, applied in Zarya, Allen, and the Part 2 report, is that a prompt of any length conveys ideas that the system then expresses, so the prompter does not author the output. Your prompt text may itself be protected if it is creative enough, but that protects the prompt, not the picture. The question is being tested in Allen v. Perlmutter, so watch that docket.

Do I have to disclose that I used AI to fix grammar or suggest a title?

No. Disclosure is required for AI-generated material that is more than de minimis, meaning material that would be independently copyrightable if a human had written it. Spelling and grammar correction, brainstorming that you then write up yourself, and similar assistance do not trigger the duty. If you are unsure, a short note to the examiner costs nothing.

If the AI-generated parts are not protected, can anyone copy them?

As far as copyright is concerned, yes, if they can separate the AI material from your protected expression. In practice that separation is often hard, and other tools may still apply: contract against anyone who accepted your license terms, trademark against confusing uses, and trade secret for anything you never published.

Does the AI company own what I make?

Usually not. Most terms assign the provider's rights, "if any," to the user and reserve a license for the provider. That assignment is only as valuable as the rights that exist, and the license and data-use terms deserve more attention than the ownership clause.

I registered a work before 2023 without mentioning AI. What should I do?

File a supplementary registration under 37 C.F.R. § 202.6 to add an accurate limitation of claim. The 2023 guidance told applicants to do exactly that. Correcting the record before a dispute is the best evidence of good faith you will ever have, and it narrows the § 411(b) argument a defendant can make later.

Can I list the AI tool as a co-author?

No. The Office will not accept a non-human author, and joint authorship requires an intent to merge contributions that a machine cannot form. Name the human authors and describe the AI material in the exclusion.

Is code written with an AI coding assistant copyrightable?

The human-written code is, and so is the human selection and arrangement of the program as a whole. Code accepted verbatim from an assistant is treated like any other AI output. Because much short, functional code is unprotectable regardless of who wrote it, the bigger practical protections for software are trade secret and contract. See Legal Protection of Software: Copyrights, Patents, Trade Secrets, and Contracts.

Do other countries protect AI-generated works?

Some do. The UK protects "computer-generated" works for fifty years, and the Beijing Internet Court has protected a Stable Diffusion image. The EU's originality standard is generally read to require a human. Protection is decided country by country, so a foreign right does not carry into the United States.

Will this change soon?

Probably at the margins rather than at the core. The human-authorship requirement now rests on an appellate decision the Supreme Court declined to review. The open questions concern how much human control is enough, and those will be answered by registration decisions, the Allen case, and technology that gives users more precise control over what the machine produces.

Practical takeaways

• Copyright protects the human layer of an AI-assisted work and nothing else. Design your workflow so that the valuable layer is human.
• Prompts alone have not been enough at the Copyright Office. Selection, arrangement, substantial modification, and perceptible human inputs have been.
• Disclose more than de minimis AI material in every application, use the Standard Application, and describe the human contribution in plain words.
• Keep contemporaneous records of who did what. They win registrations, defeat § 411(b) challenges, and establish ownership.
• A registration limited to the human contribution is still valuable, but it is thin. Expect defendants to argue that they took only what you disclaimed.
• Ownership follows the usual rules for the human layer. Get assignments in writing, and do not rely on "work for hire" labels outside the nine categories.
• Platform terms allocate rights between you and the provider; they cannot create copyright.
• Use contract, trade secret, and trademark for what copyright will not protect, and know the limits of each.
• If you are defending, pull the registration record, take discovery on the creative process, and filter AI material out of the comparison before arguing similarity.
• Treat this analysis as current to March 2026 and check Allen and the Copyright Office's latest guidance before relying on it.

This document is general information, not legal advice for any particular situation; the law changes, and the facts matter.

Related Documents

Related Articles

• Copyright Infringement Claims Against Generative AI: The New York Times, Getty, and What Comes Next: the other half of the AI copyright story, covering training data, fair use, and infringing outputs.
• Proving Copyright Infringement: Access, Substantial Similarity, and the Tests Courts Actually Use: how a thin, partly disclaimed copyright is enforced in court.
• The Copyright Claims Board: Small Claims Copyright Litigation Under the CASE Act: a lower-cost forum for modest claims on limited registrations.
• AI-Generated Inventions: Who Owns What the Machine Creates?: the patent-law counterpart to Thaler v. Perlmutter.
• Artificial Intelligence and Inventorship: Global Perspectives on Machine Contributions to Innovation: how the DABUS cases fared around the world.
• Copyright Ownership, Joint Authorship, and Termination of Transfers: the ownership rules that govern the human layer.
• Licensing Data and Models for Artificial Intelligence: Training Rights, Outputs, and Indemnities: output ownership, license-backs, and indemnities in AI contracts.
• AI Governance and Compliance: Building a Program for the EU AI Act, State AI Laws, and Everything After: labeling and transparency obligations that sit alongside copyright.
• Identifying Trade Secrets in Litigation: Reasonable Particularity, Sequencing Discovery, and Protecting the Secret While You Sue: protecting prompts and pipelines once a dispute begins.
• Data Scraping After hiQ v. LinkedIn: Copyright, Contract, and Computer Fraud Claims: where contract claims run into copyright preemption.
• Copyright Notice: Form, Function, and Best Practices: notices for works that mix human and machine material.

Related Guides

• Registering and Protecting Works Made With Generative AI: A Practical Guide: the step-by-step companion to this article.
• Negotiating an AI Data or Model License: A Practical Guide: negotiating the platform and vendor terms discussed above.
• Litigating Substantial Similarity in a Copyright Case: A Practical Guide: filtration and comparison when part of the work is unprotected.

Related Checklists

• AI-Assisted Work Copyright Registration Checklist: A Practical Checklist: every step of an AI-assisted application, in order.
• Work Made for Hire Determination Checklist: A Practical Checklist: employees, the nine categories, and assignments.
• AI Vendor Procurement and Governance Checklist: A Practical Checklist: diligence on the tools that generate the material.

Related Toolkits

• Generative AI Authorship Toolkit: Disclosure Statements, Contribution Logs, and Ownership Clauses: templates for applications, records, and contracts.
• AI Licensing Toolkit: Data Provenance Records, License Terms, and Indemnity Clauses: provenance records and indemnity language for AI vendors.
• Copyright Licensing and Clearance Toolkit: Permissions, Rights, and Chain of Title: chain-of-title tools for works with mixed origins.