Judicial Interpretation Of AI-Generated Works Under Copyright Law

Author: LawSphereX

Category: Intellectual Property

Published: 2026-08-23T20:25:33.367+00:00

Updated: 2026-08-30T07:07:07.279Z

JUDICIAL INTERPRETATION OF AI-GENERATED WORKS

UNDER COPYRIGHT LAW

Aryan Leander Wishard1

ABSTRACT

Copyright law has always had to absorb new creative technologies, from the camera to the photocopier to the computer. Generative artificial intelligence is the most disruptive of these technologies yet, because it does not merely assist human creativity but, in certain configurations, appears to substitute for it entirely. This essay examines how courts and the United States Copyright Office have interpreted copyright protection for AI-generated works against the backdrop of foundational authorship doctrine. Beginning with the human authorship requirement as settled by the Thaler v. Perlmutter litigation and the Copyright Office’s 2023 guidance, the essay examines how the partially protected Zarya of the Dawn graphic novel illustrates the emerging “sufficient human control” standard. It then turns to the fair use and training-data questions raised by Andersen v.

Stability AI and The New York Times Co. v. Microsoft, and considers the right of publicity gap exposed by AI style mimicry. The essay concludes with a comparative survey of the United Kingdom, European Union, and Indian approaches, and argues that while the bedrock requirement of human authorship is now clear, the line between mere use of an AI tool and genuine human authorship in AI-assisted works remains the central unresolved question in copyright law. Keywords: Copyright Law, AI-Generated Works, Authorship Doctrine, Sufficient Human Control, Human Authorship

1 PhD candidate Research wing Renaissance University Indore Madhya Pradesh.

I.

INTRODUCTION

In 1884, the United States Supreme Court confronted what was then a genuinely novel question: whether a photograph could qualify as the "writing" of an "author" within the meaning of copyright law. The defendant in Burrow-Giles Lithographic Co. v. Sarony argued that a photograph was nothing but a mechanical product of a machine, and that the word "author" necessarily required a human hand moving across paper. 2 The Court disagreed. It held that Napoleon Sarony, who had directed the lighting, arranged the backdrop, chosen his subject’s clothing, and determined the precise pose and expression he wanted Oscar Wilde to hold, had contributed "original intellectual conceptions" to the image that the camera merely captured. 3 The camera did not make the photographer a non-author; it made him the author of a new kind of work.

The logic of Burrow-Giles has guided courts and the Copyright Office through each successive wave of creative technology since then. It supports copyright in computer-generated typefaces, in digitally edited photographs, in algorithmically composed musical arrangements, in screenplay drafts assisted by word-processing software. The question that generative artificial intelligence now poses is whether that logic has reached its limit. When a system trained on hundreds of millions of human-authored works produces a painting, a poem, or a legal brief in response to a short text prompt, and when the human operator exercises no control over the specific expressive choices embodied in the output, is anyone the author? And if not, does the absence of an author mean the absence of copyright? These are no longer theoretical questions.

The United States District Court for the District of Columbia answered them squarely in August 2023, holding in Thaler v. Perlmutter that human authorship is a "bedrock requirement of copyright,"4 and the D. C. Circuit affirmed that holding in March 2025. 5 The Supreme Court declined review in March 2026. 6 The Copyright Office had earlier issued registration guidance drawing the line between protectable human-assisted works and unprotectable purely machine-generated output. 7 Artists have sued AI companies 2 Burrow-Giles Lithographic Co. v. Sarony, 111 U. S. 53, 58 (1884) (holding that a photograph could qualify as a “writing” of an “author” where it represented the photographer's "original intellectual conceptions").

3 Id. at 59-60.

4 Thaler v. Perlmutter, No. 22-cv-01564, 2023 WL 5333236 (D.D.C. Aug. 18, 2023) [hereinafter Thaler District Court], aff'd, No. 23-5233 (D.C. Cir. Mar. 18, 2025). 5 Thaler v. Perlmutter, No. 23-5233, slip op. at 10-12 (D.C. Cir. Mar. 18, 2025) [hereinafter Thaler Appellate].

6 Thaler v. Perlmutter, No. 25-449 (U.S. Mar. 2, 2026) (cert. denied).

7 U.S. Copyright Office, Copyright Registration Guidance

Works Containing Material Generated by Artificial Intelligence, 88 Fed. Reg. 16,190 (Mar. 16, 2023) [hereinafter AI Registration Guidance]. for training on their copyrighted works without permission. 8 The New York Times has sued OpenAI and Microsoft for allegedly reproducing its journalism verbatim through ChatGPT. 9 And a subsidiary question, whether an AI system trained on a specific artist’s distinctive style can be used to replicate that style commercially, tests the limits of both copyright and the right of publicity. 10 This essay works through these questions in order. Part II traces the historical foundations of the human authorship requirement. Part III examines the Thaler litigation in detail. Part IV addresses the partially protected "AI-assisted" category and the emerging sufficient-humancontrol standard. Part V takes up the training data and fair use question.

Part VI considers right of publicity issues raised by style and voice mimicry. Part VII surveys comparative approaches in the UK, EU, and India. Part VIII offers reform observations, and Part IX concludes. II. THE HISTORICAL ROOTS OF HUMAN AUTHORSHIP IN COPYRIGHT LAW Copyright protection in the United States rests on a constitutional foundation. Article I, Section 8, Clause 8 of the Constitution authorizes Congress to secure for limited times "to Authors and Inventors the exclusive Right to their respective Writings and Discoveries. "11 The word "Authors" is doing significant work in this clause. It has never been interpreted to mean anything other than human beings. The Copyright Act of 1976 operationalises the constitutional grant by protecting "original works of authorship fixed in any tangible medium of expression.

"12 The Act’s definition section contains no definition of "author," and the statute nowhere states that an author must be human. That requirement has been read in by courts, beginning with Burrow-Giles. The foundational role of Burrow-Giles in the current AI debate is worth pausing over. The Court’s holding was not simply that a photograph could be copyrightable. It was that the photograph was copyrightable because the photographer had made creative choices that were 8 Andersen v. Stability AI Ltd. , No. 3:23-cv-00201, 2024 WL 3823234 (N. D. Cal. Aug. 12, 2024) [hereinafter Andersen] (denying motions to dismiss the artists' core direct copyright infringement claims). 9 The New York Times Co. v. Microsoft Corp. , No. 23-cv-11195 (S. D. N. Y. filed Dec. 27, 2023) [hereinafter NYT v. Microsoft]. 10 Midler v. Ford Motor Co. , 849 F. 2d 460, 462 (9th Cir.

1988) (holding that deliberate commercial imitation of a distinctive voice constitutes misappropriation of identity, even though a voice itself is not subject to copyright protection because it is not a fixed work).

11 U.S. CONST. art. I, § 8, cl. 8.

12 Copyright Act of 1976, 17 U.S.C. §§ 101–1332 (2018).

his own independent mental conceptions. 13 The camera operated at his direction; it did not make the choices that gave the image its distinctive character. That reasoning maps directly onto the present question: when a generative AI system makes the expressive choices rather than the human operator, the human’s contribution may be no greater than that of the developer who designed the camera. The device created the work; the person merely pressed a button, or typed a prompt. The Supreme Court’s 1991 decision in Feist Publications, Inc. v. Rural Telephone Service Co. reinforced the human creativity requirement from a different angle. 14 The Court held that a compilation of telephone listings arranged alphabetically could not be copyrighted, because the selection and arrangement required no independent creative judgment; any competent person would have done the same thing.

15 More importantly for present purposes, Justice O’Connor’s majority opinion made explicit that copyright rewards originality, not labor. 16 The “sweat of the brow” theory, under which sufficient effort could substitute for creativity, was definitively rejected. A machine that generates an output through processing of training data may expend enormous “computational effort,” but effort by itself is not the currency copyright law accepts. What it requires is the minimum quantum of human creative expression, however modest that quantum may be. 17 Against this background, the Copyright Office had long maintained that only humans can author copyrightable works.

What the AI debate has forced is a rigorous examination of what that principle means when human and machine contributions are inextricably intertwined, and when the machine’s contribution increasingly dominates the expressive content of the result. III. THE THALER LITIGATION AND THE REQUIREMENT OF HUMAN AUTHORSHIP The Thaler litigation is the most significant judicial engagement with AI authorship to date, and its resolution at the appellate level has essentially closed the question for fully autonomous AI works. Stephen Thaler, a computer scientist, created an AI system he called the Creativity

13 Burrow-Giles Lithographic Company v. Sarony, 111 U.S. 53 (1884).

14 Feist Publ’ns, Inc. v. Rural Tel. Serv. Co., 499 U.S. 340, 345 (1991).

15 Id. at 346.

16 Id. at 363 (holding that a work must contain at minimum "a modicum of creativity" independently contributed by the author to qualify for copyright protection). 17 17 U.S.C. § 102(a) (listing categories of original works of authorship eligible for copyright protection including literary, musical, dramatic, pictorial, and audiovisual works).

Machine, which generated a visual artwork titled "A Recent Entrance to Paradise. " Thaler applied to register the work, listing the Creativity Machine as the sole author and himself as the copyright claimant by virtue of his ownership of the machine. The Copyright Office denied the application. The Review Board affirmed, finding human authorship to be a prerequisite for registration. Thaler sought judicial review in the United States District Court for the District of Columbia. Judge Beryl A. Howell granted summary judgment for the Copyright Office in August 2023. The opinion is notable for its clarity and relative brevity on the central question. The court held that human authorship is an "essential part of a valid copyright claim" and that copyright protection has never extended to works "generated by new forms of technology operating absent any guiding human hand.

"18 The court rejected Thaler’s argument that the human authorship requirement was unsupported by statute, pointing to the text and history of the Copyright Act, the constitutional clause, and a consistent line of Supreme Court and circuit court precedent. 19 On the work-for-hire theory, the court was characteristically direct: that theory presupposes the existence of a valid copyright in the first instance; since no valid copyright arose, the question of who owned it was moot. 20 The D. C. Circuit affirmed in March 2025. The appellate court emphasized that the Copyright Act’s use of the word "author" throughout has been consistently interpreted as requiring a human being. 21 The court also noted that Dr.

Thaler’s argument that he should be recognized as the human author by virtue of having created and owned the Creativity Machine had been waived before the Copyright Office and was therefore not properly before the court. 22 This waiver point is significant in practice: it means that future litigants who wish to argue "I was the human author; the AI was merely my tool" must make that argument at the registration stage, not for the first time in federal court. The Supreme Court denied certiorari on March 2, 2026, leaving the D. C. Circuit’s ruling

18 Id. at *4.

19 Id. (reasoning that only human authors need copyright as an incentive to create, rendering the policy rationale for extending protection to purely machine-generated works absent). 20 Thaler District Court, supra note 9, at *4 (citing Burrow-Giles, 111 U.S. at 58, and Feist, 499 U.S. at 345, for the proposition that human creativity is the "sine qua non" of copyright).

21 Id. at 14.

22 Thaler Appellate, supra note 12, at 16 (noting that Dr. Thaler's argument that he should be treated as the human author by virtue of owning and operating the Creativity Machine was waived before the Copyright Office and therefore not available on appeal). intact. 23 The denial is unsurprising. Both the district court and the appellate court had reached the same outcome on what appeared to be solid textual and historical grounds. The Supreme Court had also previously declined to hear Thaler’s parallel patent inventorship case, in which the Federal Circuit held that the Patent Act similarly limits inventorship to natural persons.

24 Taken together, the Thaler cases establish that the human authorship and inventorship requirements are not artifacts of judicial overreach but are embedded in the statutory language Congress chose, and that the Court is disinclined to expand those categories without congressional action. What the Thaler decisions do not resolve is the question of how much human creative involvement is necessary to save a partially AI-generated work. The district court specifically noted that the case involved an "autonomous" AI with no human creative input; it did not address what would happen when a human contributes creative choices alongside the AI system. That question was taken up, at the administrative level, in the contemporaneous Zarya of the Dawn proceeding. IV.

AI-ASSISTED WORKS: THE ZARYA OF THE DAWN STANDARD The Zarya of the Dawn case occupies a different place in the doctrinal landscape than Thaler. Thaler settled the question of purely autonomous AI output; Zarya addresses the harder and more practically significant question of works where human and AI contributions are genuinely mixed. The case arose when artist Kristina Kashtanova obtained a copyright registration for an eighteen-page graphic novel created using the Midjourney image-generation platform. On discovering through social media posts that the images had been AI-generated, the Copyright Office initiated proceedings to assess whether the registration should be cancelled.

In a February 2023 letter, the Office reached a nuanced conclusion: it cancelled the registration insofar as it covered the individual Midjourney-generated images, but confirmed protection for the text of the novel, the selection of images, and the overall arrangement of text and images on each page. 25

23 Edward D. Lanquist, et.al., Supreme Court Denies Certiorari in Thaler v. Perlmutter:

AI Cannot Be an Author Under the Copyright Act, https://www.bakerdonelson.com/supreme-court-deniescertiorari-in-thaler-v-perlmutter-ai-cannot-be-an-author-under-the-copyright-act (last visited on May 6, 2026). 24 Thaler v. Vidal, 43 F.4th 1207, 1213 (Fed. Cir. 2022) (holding that the Patent Act limits inventorship to natural persons, rejecting the argument that an AI system could be listed as sole inventor), cert. denied, 143 S. Ct. 1783 (2023).

25 U.S. Copyright Office, Letter to Kristina Kashtanova Regarding Zarya of the Dawn (Feb. 21, 2023)

The Office’s reasoning turned on the degree of creative control the human author exercised over each category of expression. For the text, the answer was straightforward: Kashtanova had written it herself, and no AI system had contributed to it. 26 For the compilation as a whole, the arrangement reflected Kashtanova’s own expressive choices about which images to select, how to sequence them, and how to juxtapose them with the text; those choices were sufficiently original under Feist. 27 For the individual images, however, the Copyright Office reached the opposite conclusion. It found that Midjourney users do not exercise the kind of creative control over an image that a photographer exercises over a photograph.

A photographer chooses subject, framing, lighting, depth of field, and exposure; all of these choices are made before the shutter clicks, and they are what distinguish one photographer’s work from another’s. 28 A Midjourney user types a prompt, and the system generates four possible images, each determined by the model’s internalized representation of billions of training examples. The user can iterate and refine, but cannot directly control which specific expressive elements appear in which specific configuration. 29 The principle that emerges from Zarya, amplified by the Copyright Office’s March 2023 AI Registration Guidance30 and its January 2025 Copyrightability Report31, is that a human author can obtain copyright in AI-assisted works to the extent that identifiable human expressive choices are embedded in the registered work.

The “sufficient human control” standard asks whether the human’s contribution determined the expressive elements of the output, as opposed to merely setting in motion a process over which the human had no meaningful creative direction. 32 This standard is explicitly calibrated to the nature of the specific tool used: the Copyright Office has acknowledged that different AI systems afford different degrees of [hereinafter Zarya Letter], https://www. copyright. gov/docs/zarya-of-the-dawn. pdf.

26 Id. at 2-3.

27 Id. at 9 (concluding that the selection, coordination, and arrangement of AI-generated images by the human author sufficed to protect the overall compilation, even though the individual images themselves lacked human authorship). 28 Id. at 5 (comparing a Midjourney user's relationship to the tool with that of a photographer to a camera, but finding the comparison inapt because Midjourney users lack comparable degrees of control over the expressive elements of the output). 29 Burrow-Giles, 111 U. S. at 58 (the photographer selected the subject's clothing, backdrop, pose, and expression; those choices constituted original intellectual conceptions). 30 U. S. Copyright Office, Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence (March 16, 2023), https://www. copyright. gov/ai/ai_policy_guidance. pdf. 31 U. S.

Copyright Office, Copyright and Artificial Intelligence, Part 2: Copyrightability 18-22 (Jan. 29, 2025) [hereinafter Copyrightability Report].

32 Id. at 24-27.

creative control to users, and that the analysis must account for these differences. 33 The practical implications are significant. An author who uses a word processor retains full copyright in the resulting text; the word processor makes no expressive contributions. An author who uses AI to generate background images, selects among them according to aesthetic criteria, and arranges them with her own written narrative may receive copyright in those human-authored elements. An author who types a single descriptive prompt and accepts the first output without modification likely receives nothing. The line between these categories is genuinely uncertain in the middle range.

Where a human has generated hundreds of iterations of a prompt, carefully selected among outputs, combined elements from multiple generations, and substantially modified the result, how much human control is "sufficient" is a question courts have not yet squarely resolved. 34 The Copyright Office has been candid that further case-by-case development will be necessary. 35 The Report’s recommendation that Congress may ultimately need to legislate on this point reflects the Office’s awareness that flexible administrative guidance, however thoughtfully constructed, cannot fully substitute for clear statutory text in an area where creative and economic stakes are this high. 36 V. FAIR USE AND THE AI TRAINING DATA PROBLEM A.

The Transformative Use Framework The fair use doctrine, codified in Section 107 of the Copyright Act,37 provides an affirmative defence to infringement based on a four-factor analysis: the purpose and character of the use; the nature of the copyrighted work; the amount and substantiality of the portion used; and the 33 Copyrightability Report, supra note 24, at 29-30 (observing that the "sufficient human control" standard necessarily requires case-by-case analysis because different generative AI tools differ in the degree of control they afford to human users). 34 Id. at 30 (noting that where human contributions constitute the "dominant" source of expression, copyright may extend to those contributions even when the work also contains AI-generated elements).

35 Copyrightability Report, supra note 24, at 40 (observing that legislative clarification may ultimately be necessary to resolve the question of how much human creative involvement is required, but that the existing framework is capable of accommodating most cases through flexible application of the originality and authorship requirements). 36 See Pamela Samuelson, supra note 47, at 1731 (arguing that Congress should address the AI authorship question legislatively rather than leaving it to courts to stretch existing doctrine in ad hoc ways, because the systemic economic questions involved are too large for case-by-case adjudication). 37 17 U. S. C. § 107 (setting out the four-factor fair use test and illustrating its application to purposes including criticism, comment, news reporting, teaching, scholarship, and research). effect on the potential market for the original.

38 The first factor has become the dominant analytical locus since the Supreme Court’s 1994 decision in Campbell v. Acuff-Rose Music, Inc. , which established that a “transformative” use, one that adds new expression, meaning, or message to the original, weighs heavily in the defendant’s favour. 39 AI companies have argued that training a machine learning model on copyrighted works is transformative because the purpose of training is to extract statistical patterns, not to reproduce or distribute the content of specific works. The trained model, on this view, learns relationships among words, colours, or musical sequences and uses that learning to generate new outputs, much as a human reader learns from books she has read. 40 The strongest precedent for this argument is Authors Guild v. Google, Inc.

, in which the Second Circuit held that Google’s scanning of millions of books to build a searchable corpus was transformative because the scanning served a fundamentally different purpose than reading or publishing the books. 41 Legal scholars sympathetic to AI companies have pointed to this case and to the broader category of "copy-reliant technology" as grounds for treating AI training similarly. 42 The argument has not yet received definitive judicial resolution. In Andersen v. Stability AI, artists filed a class-action lawsuit alleging that Stability AI, Midjourney, and related companies had trained their image-generation models on billions of images scraped from the internet without authorization.

The Northern District of California allowed the direct copyright infringement claims against Stability AI to proceed past a motion to dismiss, finding that the plaintiffs had plausibly alleged that Stability AI stored compressed copies of their works in the training datasets incorporated into Stable Diffusion. 43 This is a notably narrow holding; it does 38 Campbell v. Acuff-Rose Music, Inc. , 510 U. S. 569, 579 (1994) (holding that commercial parody may qualify as fair use; the Court confirmed that the "purpose and character of the use" factor turns primarily on whether the new work is transformative, adding new expression, meaning, or message). 39 Matthew D. Bunker, et. al. , The Jurisprudence of Transformation: Intellectual Incoherence and Doctrinal Murkiness Twenty Years After Campbell v. Acuff-Rose Music 12 Duke Law & Technology Review 92-128 (2014).

40 See Matthew Sag, The New Legal Landscape for Text Mining and Machine Learning, 66 J. COPYRIGHT

SOC'Y U. S. A. 291, 315-18 (2019) (arguing that training large language models on publicly available text is analogous to reading and is therefore unlikely to constitute infringement absent memorization and reproduction). 41 See Authors Guild v. Google, Inc. , 804 F. 3d 202, 215-16 (2d Cir. 2015) (holding that Google's scanning and indexing of millions of copyrighted books to enable snippet searches was transformative and constituted fair use, because the purpose was fundamentally different from the purpose of the originals). 42 See generally Matthew Sag, Copyright and Copy-Reliant Technology, 103 NW. U. L. REV. 1607, 1612-15 (2009) (arguing that machine-learning processes that ingest copyrighted works in order to produce qualitatively different outputs may be considered transformative in the same sense as prior technologies found lawful in analogous cases). 43 Id.

at 5 (finding that allegations that Stability AI stored compressed copies of plaintiffs' copyrighted works in its training datasets were sufficient to plead direct infringement at the pleading stage). not decide the merits, but it does establish that courts will not dismiss AI training data cases on the pleadings on the strength of the transformativeness argument alone. 44 B. Market Harm and Verbatim Reproduction The fourth fair use factor, market harm, is where the AI training question becomes most contested. The Supreme Court emphasized in Harper & Row Publishers, Inc. v. Nation Enterprises that market substitution is the most important of the four factors, and that transformativeness does not override clear evidence of market harm. 45 The New York Times lawsuit against Microsoft and OpenAI sharpens this point directly.

The Times’ complaint included documented examples in which ChatGPT had allegedly reproduced lengthy verbatim passages from Times articles in response to user prompts, effectively providing the content of the articles without requiring users to visit the Times’ website or pay for a subscription. 46 If a user can obtain the expressive content of a Times article from ChatGPT rather than from the Times, that is market substitution in the most direct sense, and it is difficult to reconcile with a fair use finding under Harper & Row. The verbatim reproduction problem is analytically distinct from the training problem. Even if one accepts that training a model on copyrighted works is transformative and therefore fair use, the model’s output at inference time is a separate act.

If that output reproduces substantial portions of a specific copyrighted work, it may infringe even if the training was lawful. This argument is cognate to the doctrine of contributory infringement: a tool that has substantial non-infringing uses may still generate liability if the operator specifically directs it to infringe. 47 The Times argues that ChatGPT’s reproduction of its articles is a foreseeable and insufficiently prevented result of how OpenAI deployed the model, not a mere incidental effect of training. 48 44 Andersen v. Stability AI Ltd. , No. 3:23-cv-00201, Order Re: Motions to Dismiss at 7 (N. D. Cal. Oct. 30, 2023) (largely granting motions to dismiss but allowing direct infringement claims against Stability AI to proceed and granting leave to amend).

45 Harper & Row Publishers, Inc. v. Nation Enters., 471 U.S. 539, 566-67 (1985) (holding that market

substitution by the infringing work is the most important of the four fair use factors; a defendant cannot rely on transformativeness when the use supplants demand for the original). 46 Id., Complaint ¶¶ 50-60 (documenting instances in which ChatGPT allegedly reproduced lengthy verbatim passages from Times articles when prompted by users). 47 Sony Corp. of Am. v. Universal City Studios, Inc., 464 U.S. 417, 442 (1984) (formulating the staple article of commerce doctrine in copyright, holding that the sale of a device capable of substantial non-infringing uses cannot support secondary liability). 48 NYT v. Microsoft, supra note 32, Complaint ¶¶ 75-90 (alleging that ChatGPT's verbatim reproduction of Times content, when prompted, demonstrates that the AI functions as a market substitute for the original articles, directly harming the Times' subscription and advertising revenue).

C. Judicial Caution and Unresolved Questions

Courts have been deliberately cautious about extending existing fair use doctrine to AI training. The Copyrightability Report noted that no court has yet squarely decided whether AI training constitutes fair use, and that the question requires full evaluation of all four factors in specific factual contexts. 49 The pending NYT litigation, the ongoing Andersen matter, and related cases by book authors against Meta and other AI companies will ultimately produce that full evaluation, but the timeline is measured in years. In the meantime, the Copyright Office has acknowledged that the economic stakes are large and that the outcome of these cases will shape the relationship between generative AI and the creative industries. VI. STYLE IMITATION, VOICE MIMICRY, AND THE RIGHT OF PUBLICITY Copyright law does not protect style.

An author’s distinctive narrative voice, a painter’s recognisable brushstroke technique, a musician’s characteristic harmonic language, none of these are copyrightable, because copyright protects specific original expression rather than the generalised creative approach that produced it. 50 This principle has been settled since the Supreme Court’s decision in Baker v. Selden,51 and it has never been successfully challenged. The practical consequence for AI is that a generative image model trained on a specific illustrator’s work can, in principle, produce images in that illustrator’s style without infringing the copyright in the training images themselves, provided it does not reproduce specific protected expression. But the absence of copyright protection does not mean the absence of legal protection entirely. The Ninth Circuit’s 1988 decision in Midler v. Ford Motor Co.

52 established that a celebrity’s distinctive voice, though not itself copyrightable, is protected from deliberate commercial imitation as an aspect of the right of publicity. 53 The court held that when a professional singer’s voice is widely known and is deliberately imitated to sell a product, the sellers have “appropriated what is not theirs. ”54 The court drew an analogy between a voice and a face: both 49 Copyrightability Report, supra note 24, at 45 (acknowledging that no court has yet decided whether AI training on copyrighted works constitutes fair use, and that the question will require evaluation of all four factors with particular emphasis on market harm). 50 17 U. S. C. § 106 (enumerating the exclusive rights of a copyright owner: reproduction, distribution, preparation of derivative works, public performance, and public display).

51 Baker v. Selden, 101 U.S. 99 (1879).

52 Midler v. Ford Motor Co., 849 F.2d 460 (9th Cir. 1988)

53 Id.

54 Id. at 463 ("A voice is not copyrightable. The sounds are not fixed. What is put forward as protectible here is are uniquely personal attributes over which the individual retains a proprietary interest, even though neither is a fixed work of authorship. Waits v. Frito-Lay extended this analysis to cover Lanham Act false endorsement claims. 55 Generative AI intensifies the Midler problem significantly. A voice synthesis system trained on a specific singer’s recordings can produce new “performances” in that singer’s voice without reproducing any specific recording. An image generator trained on a specific illustrator’s catalogue can produce new images that are unmistakably in her style. The Copyright Office noted in its 2025 Copyrightability Report that these cases raise right of publicity questions that fall outside the direct scope of copyright law.

56 Several of the artistplaintiffs in Andersen have alleged Lanham Act violations alongside their copyright claims, precisely because the style-imitation harm is not fully captured by copyright infringement doctrine. 57 The legal gap this creates is real and underappreciated. An artist whose distinctive style is commercially replicated by an AI system suffers a recognisable economic harm: the market for her unique style, which she developed through years of labor, is partly satisfied by a machine that offers similar aesthetic output at near-zero marginal cost. Whether this harm is cognizable under existing law depends on the jurisdiction’s right of publicity doctrine and on how courts interpret the commercial imitation cases in the AI context. 58 This is a question that neither copyright doctrine, standing alone, nor the Midler line of cases, standing alone, fully resolves. VII.

COMPARATIVE PERSPECTIVES A. The United Kingdom The United Kingdom has the most explicit statutory provision for computer-generated works of any major jurisdiction. Section 9(3) of the Copyright, Designs and Patents Act 1988 provides more personal than any work of authorship. ").

55 Waits v. Frito-Lay, Inc., No. 90-55981 (9th Cir. 1992).

56 U. S. Copyright Office, Copyrightability Report, supra note 24, at 35 (noting that AI systems trained on a specific artist's style and prompted to generate "in the style of" that artist raise right of publicity questions not directly addressed by copyright law). 57 Andersen v. Stability AI Ltd. , No. 3:23-cv-00201, First Amended Complaint ¶¶ 104-115 (N. D. Cal. Nov. 29, 2023) (adding allegations that outputs generated in response to prompts using artists' names mimicked their styles in ways constituting potential Lanham Act violations). 58 See generally Ryan Vacca, AI as Inventor and Author, 101 OR. L. REV. 793, 820-25 (2023) (analyzing the inadequacy of copyright and right of publicity frameworks to address the harm suffered by artists whose distinctive styles are replicated at scale by generative AI).

that the author of a computer-generated work is "the person by whom the arrangements necessary for the creation of the work are undertaken. "59 This provision vests copyright in the human operator or developer even when the work contains no human creative expression in the conventional sense. The UK approach deliberately avoids the question of whether AI outputs are "original" in the sense that copyright usually requires; instead, it identifies a human party as the author by operation of law. This is a pragmatic rather than a principled solution. Critics have argued that it stretches the concept of authorship beyond coherence: calling the person who "arranges for" the creation of a work its "author" is at odds with the intuition that authorship involves creative expression.

60 Defenders point out that it generates legal certainty, ensures that commercially valuable AI outputs receive protection, and provides an incentive framework for investment in creative AI development. The UK Intellectual Property Office launched a consultation in 2022 on whether the provision should be retained, extended, or abolished; it ultimately decided to retain it pending further development, while acknowledging the deep conceptual difficulties it presents. B. The European Union The EU has not enacted specific legislation on AI authorship, but the general framework of EU copyright directives generally requires that a work be the "author’s own intellectual creation" reflecting "free and creative choices. "61 This phrase, derived from the CJEU’s Infopaq and Painer decisions, has been understood to require a human author who has expressed her personality in the work.

On this analysis, a fully autonomous AI output would not qualify for copyright, because there is no human personality expressed in it. The EU’s current position therefore broadly parallels the US requirement of human authorship, though the doctrinal path is different. The EU has, however, addressed the training data question more directly than US courts have 59 Copyright, Designs and Patents Act 1988, c. 48, § 9(3) (UK) (defining the "author" of a computer-generated work as "the person by whom the arrangements necessary for the creation of the work are undertaken," thereby vesting copyright in the human operator even where no human creative expression is involved in the output). 60 See generally Martin Senftleben & Laurens Buijtelaar, Robot Creativity: An Incentive-Based Neighbouring Rights Approach, 12 EUR. INTELL. PROP. REV.

797, 800-04 (2020) (examining the UK computer-generated works doctrine and arguing that it offers a more pragmatic response to AI authorship than requiring courts to stretch the human authorship requirement). 61 See Jane C. Ginsburg, People Not Machines: Authorship and What It Means in the Berne Convention, 49 IIC 131, 134-36 (2018) (arguing that the Berne Convention's reference to "author" has been consistently understood across member states to require a human creator, even though the Convention does not define the term). yet done. The 2019 Copyright in the Digital Single Market Directive introduced text and data mining exceptions that permit research organizations to mine copyrighted works for scientific purposes, while allowing rights holders to opt out for other commercial mining.

62 These provisions are specifically relevant to AI training and represent a legislative compromise, rather than a judicial one, between the interests of rights holders and those who seek to build AI systems using copyrighted data. Whether this framework is adequate to address the scale of commercial AI training remains actively debated. C. India India’s Copyright Act, like the UK statute, contains a provision addressing computer-generated works. Section 2(d)(vi) of the Indian Copyright Act defines the "author" of such works as "the person who causes the work to be created," vesting authorship in the human arranger of the creative process. 63 The Supreme Court of India’s decision in Eastern Book Co. v. D. B.

Modak held that copyright requires the author’s own skill, judgment, and labour, and that a mechanical or automated process without independent human judgment does not satisfy this standard. 64 This suggests that a purely autonomous AI output would likely not receive copyright protection under Indian law, while the outputs of AI-assisted human creativity might, depending on the degree of human judgment involved. The WIPO’s 2019 Issues Paper noted that jurisdictions across the world approach the AI authorship question with significant variation, but that most systems find some nexus to a human creative decision-maker as a precondition for protection. 65 The divergence is primarily in how direct and substantial that nexus must be.

The US and EU require genuine creative expression from a human; the UK and India create a deemed-authorship mechanism that can apply even in the absence of such expression. This divergence will generate cross-border 62 See Directive 2006/116/EC of the European Parliament and of the Council of 12 December 2006 on the Term of Protection of Copyright and Related Rights, art. 1(3), 2006 O. J. (L 372) 12 (EU) (providing for copyright in computer-generated works through a special neighbouring right, with ownership vesting in the person who made necessary arrangements for the work's creation). 63 Indian Copyright Act, No. 14 of 1957, § 2(d)(vi) (India) (defining "author" to include "the person who causes the work to be created" with respect to computer-generated works, paralleling the UK approach and vesting authorship in the human arranger of the creative process). 64 See Eastern Book Co. v.

D. B. Modak, (2008) 1 SCC 1, ¶ 30 (India) (Supreme Court of India holding that a work qualifies for copyright if it reflects the author's own skill, judgment, and labour, and that a mechanical or automated process without independent human judgment does not satisfy this standard). 65 WIPO, Draft Issues Paper on Intellectual Property Policy and Artificial Intelligence ¶¶ 18-27 (2019), https://www. wipo. int/edocs/mdocs/mdocs/en/wipo_ip_ai_ge_19/wipo_ip_ai_ge_19_1. pdf (canvassing the questions of whether AI-generated works should receive protection and, if so, under what legal theory and in whom ownership should vest). complications as AI-generated works circulate globally. VIII. THE PATH FORWARD: REFORM CONSIDERATIONS The human authorship requirement, as currently interpreted, is both necessary and insufficient.

It is necessary because the constitutional and statutory design of copyright presupposes a human creator whose incentives and welfare are affected by the availability of protection. A machine needs no incentive to create; it generates output whenever instructed to do so, regardless of whether its outputs are protected. 66 Extending copyright to purely autonomous AI works would therefore fail the foundational incentive rationale without generating any corresponding social benefit in the form of additional creative production. The requirement is insufficient, however, because it leaves the AI-assisted category in genuine uncertainty. The Copyright Office’s "sufficient human control" standard is the right conceptual framework, but it requires case-by-case application to an almost infinite variety of human-AI working relationships.

Different AI tools afford their users very different degrees of creative control; the same tool may afford high control in one use scenario and virtually none in another; and the creative choices that matter for copyright purposes may be spread across many iterations of prompting, selecting, and refining that are hard to document and evaluate after the fact. Legislators considering the AI authorship problem would do well to consider whether a disclosure-and-registration system, requiring authors to specify which elements of a work are human-authored and which are AI-generated, could provide the evidentiary basis for principled case-by-case adjudication while also ensuring that the public record reflects the true creative history of registered works.

67 On the training data question, the balance between incentivising investment in AI development and protecting rights holders from unauthorized exploitation of their work is one that courts are not well-positioned to strike alone. The EU’s legislative solution, using sector-specific text and data mining exceptions with opt-out rights, is one model worth examining. The ongoing NYT litigation may produce a result that renders legislation unnecessary in the near term; 66 See Robert Brauneis & Robert Gomulkiewicz, Measuring the Immeasurable: The Philosophical and Legal Case Against Copyright Valuations, 77 WASH. L. REV. 721, 762-65 (2002) (arguing that the incentive rationale for copyright assumes a human creator whose behaviour can be influenced by the prospect of exclusivity; this rationale has no direct application to a machine that creates regardless of legal incentive).

67 AI Registration Guidance, supra note 16, at 16,193 (explaining the disclosure obligation and noting that applicants who fail to disclose AI-generated content risk cancellation of their registrations). alternatively, it may produce a result so disruptive to one side or the other that legislative correction becomes politically irresistible. 68 IX.

CONCLUSION

The Thaler litigation and its denouement in the Supreme Court’s denial of certiorari in March

2026 have put one question to rest: an AI system cannot be an author for copyright purposes,

and a work generated entirely by an autonomous AI, without meaningful human creative contribution, is not copyrightable. The Zarya of the Dawn proceeding and the Copyright Office’s subsequent guidance have sketched the beginnings of an answer to the harder question of AI-assisted works, establishing that sufficient human control over the expressive elements of a work remains a precondition for copyright even when AI tools are used heavily in the creative process. The training data cases are further from resolution, but the judicial instinct, visible in both Andersen and the NYT litigation, is to subject AI training to ordinary copyright analysis rather than to recognize a blanket exemption on transformativeness grounds.

What connects these doctrinal threads is a single underlying problem: copyright law was designed by and for human beings, to protect the specific kind of creative labor that human beings perform. Generative AI does not perform that labor. It does something different, something that draws on the accumulated creative labor of vast numbers of human authors, and something that produces outputs which are, in many contexts, functionally indistinguishable from human creative work. The difficulty is not that copyright law is fundamentally inapplicable to this situation; it is that its key concepts, authorship, originality, expression, and transformativeness, were calibrated for a world where machines assist but do not create. Recalibrating them for a world where machines create but humans operate is the central challenge that legislators, administrators, and courts now face.

The historical record offers reason for cautious optimism. Photography was once thought incompatible with copyright; now it is one of the most litigated categories of protected work. Computer programs were once thought to be mere mechanical instructions; now they are routinely registered and licensed. Copyright law has a record of adapting to new creative technologies without losing its core commitments. Whether the adaptation required for

68 Martin Senftleben, The TDM Opt-Out in the EU – Five Problems, One Solution

https://legalblogs.wolterskluwer.com/copyright-blog/the-tdm-opt-out-in-the-eu-five-problems-one-solution/ (last visited on May 9, 2026). generative AI can be accomplished within the existing statutory framework, or whether it will require legislative intervention, is the most important open question in intellectual property law today. It is unlikely to remain open for long.