Tipping the Scale: Ingestion, Output, and Authorship in the Age of Generative AI

Momo Chen

Edited by Keerthi Chalamalasetty, Jasmine Iyer, Judge Baskin, and Sahith Mocharla

The swift ascent in popularity of generative artificial intelligence (AI) paired with its growing ubiquity in recent years has taken both the creative industry and the field of copyright law by storm. In May of 2023, Hollywood writers initiated a guild strike protesting against the usage of AI models, bluntly alleging AI’s ingestion of authors’ original works as a form of theft [1]. The immediacy of the resulting legislative and institutional response aptly conveyed the scale of writers' dissatisfaction. On September 24th that same year, the Writers Guild of America's contract negotiations demanded the first collectively bargained restrictions on AI’s role in creative work [2]. Generative AI’s disruption of creative industries has now extended beyond cultural anxieties, prompting contractual retaliation. Driven by a resentment of creative destruction, creators have sought out copyright law as a tool to defend their interests, sparking the central debate at the intersection of AI and copyright law: can existing copyright laws appropriately address infringement concerns raised by AI—and, if not, should they?

In the process of resolving this debate, three constituent questions need to be addressed. 

I. First, does ingestion constitute copyright infringement? Recurringly, the way in which ingestion is explained to the masses—as the reproduction of original works used for training purposes—has led creatives to associate it with infringement. 

II. Second, should AI-generated outputs be considered infringing derivatives? 

III. Lastly, who, if anyone, gets to claim copyright ownership of AI-generated outputs? 

Each question represents a point of legal contention capable of tilting the scale between the prerogatives of copyright holders and the imperative of technological innovation. Though escalating infringement concerns presses for a stricter protection of creator incentives, that protection should not come at the expense of technological development past the point of constitutional equilibrium. Although these questions are prevailingly recognized as discrete, unresolved concerns, the three issues, when viewed in conjunction, already resemble a constitutionally balanced scale that maintains copyright law’s core imperative “to promote the progress of science and useful arts”[3]. More specifically, while the non-infringing nature of ingestion and AI outputs favors developers’ interests, the non-copyrightability of AI generated works provides the necessary counterweight to assure creator rights.

Is Ingestion Copyright Infringement?

At the heart of the discourse between generative AI and copyright law stands the issue of ingestion. When a generative AI model is initially trained, it reproduces digital files representing original works in order to extract statistical patterns from which it responds to user prompts [4]. These ingested original works are often copyrighted products ranging from images and text to sound recordings. On this basis, rights holders have sought to hold AI developers liable for infringement at the initial stage of model development, framing ingestion as a convenient scapegoat for all of generative AI's disruptive faults. Though backed by the genuine imperative of protecting copyright holder interests, such allegations risk tipping past the point of constitutional equilibrium. More specifically, if infringing ingestion were to hold broadly, the scale of financial compensation developers would be charged with would invariably pose an existential threat to AI development. Among actors within the creative industry who have expressed discontent over the loss of potential market revenue, the Times in New York Times Co. v. Microsoft Corp. (2025) spearheaded the attempt to render ingestion as infringement [5]. On December 27, 2023, the Times filed a complaint in the Southern District Court of New York alleging that Microsoft had unlawfully ingested over 16 million of the Times’ copyrighted content in its Common Crawl dataset to train the model underlying Microsoft's Copilot [6]. That scale of ingestion occurred all without authorization or compensation. The complaint not only sought monetary damages and a permanent injunction against future infringement, but also mandated the destruction of all GPT models and training sets currently incorporating Times content. Underpinning this demand was the reasoning of market harm. Specifically, the paper claimed that Copilot's outputs were capable of reproducing Times articles at length, enabling users to bypass the Times’s paywall and obviating any need to visit the publication [7]. Upon a closer reading of the precise domain of copyright law, however, an objection to this argument emerges.

A substantial body of legal scholarship has attempted to address proliferating upstream infringement claims by contending that ingestion fundamentally falls outside the scope of copyright protection [8]. The governing principle is what legal scholars have come to term the “spillover principle,” where the application of copyright is strictly limited to expressive forms. According to the spillover principle, any non-expressive element in a piece of protected work should be permitted to spill over into the public domain as a positive externality [9]. As such, copyright protection does not extend to ideas within an article, genre-based principles of an artwork, or an algorithm’s systems, regardless of the form in which it is embodied [10]. In line with the idea-expression dichotomy, such restriction is a feature by design, not a bug. Following this rationale, AI developers who treat original works as “raw computational [training] material” or merely “a bag of words” should not be interpreted as engaging with the unique expression inherent in those works [10]. Infringement claims therefore lack validity––not for the reason that ingestion fulfills fair use considerations––but because it never crosses the threshold of copyrightable subject matter to begin with.

Nonetheless, assessing ingestion under the fair use framework serves to evaluate its constitutionality with respect to copyright law’s inherent aims beyond a mere examination of ingestion's mechanics [11]. To satisfy fair use assessments, Microsoft analogized its use case to Authors Guild v. Google, Inc.'s Second Circuit ruling in which the court held that Google's digitization of millions of in-copyright books constituted a “highly transformative” fair use [12]. Although Google's system—a searchable index that serves users nonsubstantial snippets in response to search queries—was a commercial product, its purpose was considered sufficiently non-expressive and distinct that it did not substitute for the original works in their intended market. Rather than displacing the market for the underlying works, Google's index made them easier to discover, facilitating greater public access to knowledge. Rejecting that analogy, the Times distinguished Microsoft’s use by highlighting Copilot’s ability to reproduce entire articles verbatim in response to targeted prompts. Unlike Google’s system, which merely directed users toward original works, Copilot produces direct substitutes that would crowd out Times journalism. This output-as-substitute argument, targeting the fourth and most consequential fair use factor, constitutes the backbone of the Times's market harm claim [13]. In essence, the Times demonstrated Copilot’s capacity to reproduce ingested original works in their entirety. This translates to the concept of “memorization,” where ingestion extends beyond an extraction of statistical patterns to retain copies of original works within the model’s parameters—including their expressive components. Though memorization would entail infringement over copied expression, its premise wrongfully conflates the act of ingestion, a development stage,with consequences that materialize only at the output stage. Market harm is a concern associated with generative AI’s outputs, not of ingestion during training. To hold ingestion infringement based on the model’s downstream products mistakenly collapses two fundamentally distinct inquiries into one. In doing so, liability would be imposed at the input stage of development before the point of generation responsible for potential market substitution even occurs. Though such a framework has yet to be adopted by courts, it is a firm corollary that follows the way ingestion is distinguished from the harm associated with AI’s downstream products.

To this end, there are two main issues with deploying these ingestion-as-infringement claims to strike an equitable balance between creator interests and technological innovation. First, copyright law as currently conceived does not apply to ingestion: the spillover principle denies the existence of prima facie infringement before the fair use analysis could ever arise. Second, even where output-level harm is genuine, it cannot be properly addressed through ingestion liability—the harm occurs downstream, and therefore so must its remedies. Perhaps the fixation on ingestion as infringement, and the resulting contortion of established copyright doctrine, stems from a deeper set of anxieties about the livelihood of artists and the value of human creativity. These concerns, however disconcerting and potentially valid, are ones that current copyright law is ill-equipped to address as ingestion ought not be held structurally liable for infringement. Instead, more targeted output-stage policy carveouts that assure sui generis (unique) protections under demonstrable market harm would better address the creative industry's grievances without stretching the subject matter of copyright law beyond its limits [14].

Outputs as Infringing Derivatives?

On January 16th, 2023, Getty Images initiated proceedings against Stability AI in the UK regarding the development and deployment of the Stable Diffusion image-generation model [15]. The class-action complaint in Anderson v. Stability AI (2023) further asserts that all images produced by Stable Diffusion should be considered infringing derivative works, on the grounds that all are derived from Getty’s image database on which the model was trained [16]. Getty anchors its claim by characterizing Stable Diffusion as a “collage tool” whose outputs compete directly against the artists’ own works and thereby harm their markets [17]. Users of Stable Diffusion can moreover submit prompts requesting images of a particular subject in the style of a specific named artist, further blurring the line between inspiration and replication. As with the concerns raised at the ingestion stage, the downstream proliferation of AI-generated outputs poses an equally formidable threat to the interests of human creators. In order to establish copyright infringement at the output level, copyright owners must satisfy two conditions: they must show both that the AI program had access to their original work––demonstrated by proving that a copy was scraped and ingested––and that the output appropriated a substantial portion of that work’s original expression [18]. This second condition addresses the plaintiff's infringement claims.

The substantial similarity test is difficult to define. The Ninth Circuit, under whose jurisdiction Anderson falls, employs the two-part extrinsic/intrinsic test. It draws on expert testimony to assess similarity in protectable expressive elements and, more subjectively, asks whether “an ordinary reasonable person” finds a “substantially similar total concept or feel” [19]. On the other hand, the Third Circuit that governs Getty Images v. Stability AI utilizes the abstraction-filtration comparison, where the work is broken down into its structural components with non-protectable elements filtered out at each level of abstraction [20]. Regardless of the governing standard in use, the mere claim or even finding that a second work was based upon an earlier one does not automatically render it an infringing derivative. In particular, one study found a significant amount of copying in only “2% of images generated by Stable Diffusion” [21]. Though the authors of the study have stated that their methodology “likely underestimates the true rate” of copying, the model’s tendency to introduce noise to image elements during the encoding process nevertheless supports a low likelihood of similar output. Corroborating this improbability, OpenAI has publicly commented that infringing outputs are “unlikely accidental outcomes” of well-constructed generative AI systems, noting that diffusion models assemble outputs through probabilistic inference and generation rather than memory-based retrieval [22]. The class-action complaint itself acknowledges, by the abstraction-filtration comparison, that even “in the style of” claims encounter limitations. After all, copyright does not protect broad techniques and features, only expressive elements specific to a particular work [23]. In the categorical sense, AI-generated outputs are hence unlikely to constitute infringing derivative works under established copyright laws.

Market dilution concerns, however, challenge the validity of existing copyright doctrine and shift the focus of subsequent analysis back to fair use. Specifically, the U.S. Copyright Office has contended that even when outputs are considered technically non-infringing, flooding the market with stylistically similar outputs may constitute cognizable harm, warranting a fair use assessment under market dilution concerns [24]. Such reasoning surfaced in Kadrey v. Meta Platforms, Inc. (2025), where Meta used books pirated from “shadow libraries” to train its AI model Llama [25]. The court ultimately ruled in favor of Meta’s fair use defense, finding the use highly transformative given the plaintiff’s failure to demonstrate sufficient evidence of market harm. This result, however, was not inevitable. Had such evidence been prepared ahead of time, a potentially winning market dilution argument would have been preserved, as the Court was already evaluating transformativeness through the lens of market harm. This possibility finds grounding in Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith (2023), involving Andy Warhol’s use of Lynn Goldsmith's photograph of Prince, which notably narrowed the conception of fair use [26]. In particular, the case found that the silkscreen incorporating photographs of Prince for a magazine cover had similar purposes with Goldsmith’s original collection. Thus, by recognizing both works as commercially licensed images of Prince, Warhol compressed the scope of transformative use and marked a doctrinal shift toward prioritizing the market value of original works. The ruling entailed that it no longer sufficed for a challenged work to offer new meaning or value. Instead, the threshold was elevated to whether the use has a fundamentally different purpose from the original. To this end, Warhol's narrowing of transformative use, read alongside Kadrey's market harm dicta, significantly diminishes the prospects of a successful fair use defense. 

The applicability of the Warhol precedent however may be limited, since AI outputs typically serve expressive and informational purposes distinct from the training data on which it was developed. In instances where similarity in purpose is shown, disagreements persist over whether plaintiffs can bypass the substantial similarity test entirely and proceed directly to a fair use assessment on the strength of market harm evidence alone. Seeking to foreclose fair use arguments at the entry threshold, some scholars invoke the same no-infringement, no fair-use logic applied at the ingestion stage [27]. Given that latent representations of works used to train models do not constitute copies, and that outputs rarely meet the bar of substantial similarity, the argument proceeds that outputs should not be considered infringing derivatives at all. Instead, market harm claims unaccompanied by a showing of infringement are better understood as a form of virtuous competition that advances copyright law’s progress-promoting goals. Ironically, increasing the volume and diversity of a model’s training data reduces rather than amplifies the likelihood of substantially similar outputs. Though these considerations do not rehabilitate the prima facie infringement misconception, they do expose a blind spot in established copyright law: the lack of an adequate tool to address probabilistic market dilution at scale.

Unlike the mislabeling of ingestion as infringement, the discourse surrounding AI outputs stems from a genuine disagreement over tangible market dilution. Where the letter of copyright law renders output technically non-infringing, its spirit acknowledges how the saturation of markets with AI-generated content collectively erodes economic incentives and protections for creators. Even when no individual output crosses the substantial similarity threshold, each AI-generated image that displaces a commission inflicts market harm that stifles incentives for human creativity. However, relying upon copyright law alone to resolve this tension risks a seismic adjustment of established doctrinal thresholds that have evolved to serve fields far beyond generative AI. As previously proposed, a more suitable measure would be a calibrated policy carve-out targeting demonstrable market substitution. Furthermore, such a carve-out does not need to be developed from scratch. The Digital Millenium Copyright Act (DMCA) of 1998, for instance, protects intellectual property online while shielding service providers from liability for user infringement. More specifically, the DMCA’s safe harbor structure could offer liability protection conditioned upon developers compliance with procedural requirements [28]. Specific requirements can consist of the EU’s Article 4 text-and-data-mining exception under the Directive of Copyright in the Digital Single Market (CDSM), which offers rightsholders the choice to opt-out of commercial ingestion [29]. Subsequently, transparency mandates would be enforced to monitor consensual ingestion of original works. These measures would address dilution at scale without distorting copyright doctrine in ways that produce adverse consequences across creative fields. Though copyright law may be the first line of defense that creatives resort to, it is not the appropriate tool to resolve their concerns, regardless of how resonant such calls to establish infringement may be.

Copyright ownership of AI-generated outputs

On the flip side of the generative AI discourse, many users of the technology have been advocating for copyright ownership over AI outputs. If ingestion falls outside copyright's domain and outputs are unlikely to constitute infringing derivatives, who—if anyone—would get to claim the fruits of AI-assisted creation? Legal theorist Ioan-Radu Motoarca notably identifies two relevant conceptions of authorship: the ‘basic’ author as the proximate cause of a work, and the ‘overseer’ as the one who directs its creation [30]. Ordinary usage involving trivial human creative input and generic prompting points authorship away from human users. The now typical user expression “I got ChatGPT to write me a story,” for instance, implicitly acknowledges the AI model as the primary overseer of the story’s creation and the “effective cause of a work produced” by the basic conception [31]. Under both accounts, the AI produced the work; the user merely directed the occasion of its creation. 

Under current copyright law, however, AI models are incapable of holding property rights under the human authorship requirement of the Copyright Act [32]. In 2018, Stephen Thaler, an AI developer who built a generative system called the “Creativity Machine,” attempted to register copyright for an artwork his system had generated [33]. The Copyright Office denied the application on the ground that the work lacked human authorship—and denied it again upon two requests for reconsideration. The Office’s persistence was well-founded: numerous structural provisions of the Copyright Act presuppose a human author [34]. The stipulation of protection duration after “the author's death,” the assumption of inheritance by the “widow or widower,” and written signature requirements for the transfer of rights all confirm this understanding [35]. Though AI outputs made with minimal user creative input have been rendered non-copyrightable, the question of copyright ownership in cases of meaningful human involvement remains.

Tracing an outputted work’s creative origins points to the model’s developers as human authors with meaningful creative contribution, by virtue of their configuration of training data and program code. When Thaler later sued the Register of Copyrights, Shira Perlmutter, in federal courts in D.C., the work-for-hire argument was invoked in support of his claims. In Thaler v. Perlmutter (2023), under the Copyright Act, a “work made for hire” is a creative output produced by an employee under the direction of an employer [36]. In such cases, the employer—analogous here to the AI developer or user—is treated as the initial copyright owner rather than the person who actually performed the creative act. Thaler, who is simultaneously the model’s developer and its configurer, can be compared to the director of a film studio who commissions a score. In this comparison, he specifies the theme and mood, and the composer independently executes the creative work within the scope of his employment. Though appealing, the work-for-hire argument cannot adequately circumvent the constitutional requirement that copyright vest exclusively in human creators. Ruling in favor of Perlmutter, the D.C. Circuit held in March 2025 that the Copyright Act “requires all eligible work to be authored in the first instance by a human being” [37]. The analogy is further weakened by the prominence of stochastic randomness in generative AI models. Unlike the work-for-hire context, which presumes similar outputs given consistent creative instruction, AI outputs may vary drastically depending on the length and specificity of the prompt as well as the model in use. Nevertheless, because Thaler involved a fully autonomous model without meaningful human creative contribution, its holding only pertains to cases without human input specific to the creation of individual artworks. Thaler's ambiguous reading of “first instance” creation thus preserves the possibility of users fulfilling human authorship requirements through extensive initial prompting and creative direction.

Currently, U.S. copyright law looks to the Copyright Office’s March 2023 AI Guidance document for working answers [38]. In practice, copyright attaches exclusively to aspects of a work where a human exercised sufficient creative control over expressive elements. Works primarily generated by autonomous AI spill over into the public domain, while modified AI drafts are examined on a case-by-case basis. The guidance document found concrete application that same year when artist Kristina Kashtanova applied for copyright in a graphic novel, Zarya of the Dawn, created using Midjourney, an AI image-generation platform [39]. Kashtanova had iterated through hundreds of descriptive prompts over the course of a year to coax Midjourney toward images matching her creative vision. She then combined those AI-generated images with her own written text before registering the full work with the Copyright Office. The comic ultimately received partial copyright protections. The text Kashtanova authored, along with her selection, coordination, and arrangement of images—authorial judgements reflecting human creative contribution—were deemed copyrightable; the images generated by Midjourney were not [40]. Though the overseer conception seems to support Kashtanova’s claim for authorship, her role more closely mirrors that of an anthology editor who may claim ownership of her editorial choices in curating materials but not of any underlying contributions. As such, the Zarya of the Dawn proceeding adopts a narrow reading of the aforementioned “first instance” human creation, affirming that the initial creative act must be performed by a human. In so doing, the Copyright Office effectively establishes an authorship gradient where the extent of protection a work receives is set proportional to the degree of human authorial contribution. To this end, some scholars have further proposed treating prompts as copyrightable creative assets in their own right [41]. However, any such determination requires case-by-case examination of the prompt's extensiveness and specificity, as well as variability of its resulting outputs. 

Taken together, the present state of U.S. copyright law navigates the attribution of copyright in AI-assisted works with considerable care. By extending protection exclusively to human-authored expressive elements, the Copyright Office forecloses the alternative in which copyright provisions would incentivize the proliferation of low-effort AI outputs. Catastrophically unleashing the floodgates to low-effort protected works shielded from competition would further erode economic incentives for human creators, exacerbate market dilution concerns, and ultimately invert copyright's constitutional purpose of promoting original human authorship. The non-copyrightability of AI-authored components of a work thus represents a principled equilibrium and a much needed counterforce to the liability protection endowed to AI development. This preserves incentives for human creativity and tampers concerns over the sweeping displacement of original works, all without impeding technological progress

Conclusion

The three doctrinal questions examined in this article—ingestion, infringing outputs, and copyright ownership of AI-generated works—are not independent inquiries. Rather, they are better understood as successive weights placed upon the same scale, each adjusting the balance between the imperative of technological innovation and the prerogatives of human authorship. Ingestion falls outside copyright's legitimate domain under the spillover principle and ought to remain neutral without tipping that scale at all [43]. Holding ingestion as infringement would impose liability at the initial stage of AI development before any harm has materialized. Output liability, though animated by genuine anxieties over market dilution, is a concern that copyright as currently conceived is ill-suited to address. A calibrated policy carve-out, targeted at demonstrable market substitution, would accommodate the specific challenges posed by generative AI without distorting doctrine access across the broader range of fields copyright law serves. Finally, the non-copyrightability of AI-generated works provides the balancing counterweight: by constraining the incentives for AI usage, it preserves economic incentives for human creativity without foreclosing the technology itself [44]. This tripartite framework thus demonstrates how the resolution of each doctrinal question dictates the range of permissible answers to the other. Ultimately, it is only by acknowledging this interdependence that lawmakers can navigate the landscape of AI copyright law without mistaking intervention for disruption.

---

[1] Molly Kinder, Hollywood Writers Went on Strike to Protect Their Livelihoods from Generative AI. Their Remarkable Victory Matters for All Workers., Brookings Inst. (Apr. 12, 2024), https://www.brookings.edu/articles/hollywood-writers-went-on-strike-to-protect-their-livelihoods-from-generative-ai-their-remarkable-victory-matters-for-all-workers/.

[2] Know Your Rights - Artificial Intelligence, Writers Guild of American West, https://origin.www.wga.org/Content/Page (last visited Apr. 16, 2026).

[3] U.S. CONST. art. I, § 8, cl. 8. 

[4] Pamela Samuelson, Generative AI Meets Copyright, 381 Science 158 (2023), https://doi.org/10.1126/science.adi0656.

[5] Complaint #1 - The New York Times Company v. Microsoft Corporation, No. 1:23-Cv-11195 (S.D.N.Y., Dec. 27, 2023), CourtListener, https://www.courtlistener.com/docket/68117049/1/the-new-york-times-company-v-microsoft-corporation/.

[6] See [5].

[7] See [5].

[8] Oren Bracha, The Work of Copyright in the Age of Machine Production, SSRN Journal (2023), https://www.ssrn.com/abstract=4581738.

[9] “What Is Spillover Theory? Simple Definition & Meaning.” LSD Law, https://definitions.lsd.law/spillover-theory (last visited Apr. 22, 2026).

[10] 17 § 102

[11] See [8].

[12] Authors Guild, Inc. v. Google, Inc., 804 F.3d 202 (2d Cir. 2015).

[13] 17 U.S.C. § 107

[14] SUI GENERIS INTELLECTUAL PROPERTY PROTECTION FOR COMPUTER SOFTWARE, https://cyber.harvard.edu/property/protection/resources/phillips_unedited.html (last visited Apr. 22, 2026).

[15] Getty Images -v- Stability AI, Courts and Tribunals Judiciary (Nov. 4, 2025), https://www.judiciary.uk/judgments/getty-images-v-stability-ai/.

[16] Andersen v. Stability AI Ltd., Class Action Complaint, No. 3:23-cv-00201 (N.D. Cal. Jan. 13, 2023), https://ipwatchdog.com/wp-content/uploads/2023/02/Andersen_et_al_v._Stability_AI.pdf.

[17] See [15]

[18] Andersen v. Stability AI: The Landmark Case Unpacking the Copyright Risks of AI Image Generators - NYU Journal of Intellectual Property & Entertainment Law, https://jipel.law.nyu.edu/andersen-v-stability-ai-the-landmark-case-unpacking-the-copyright-risks-of-ai-image-generators/ (last visited Apr. 7, 2026).

[19] Christopher T. Zirpoli, Generative Artificial Intelligence and Copyright Law, Cong. Rsch. Serv., LSB10922 (2026), https://www.congress.gov/crs-product/LSB10922.

[20] See [15]

[21] Gowthami Somepalli et al., Diffusion Art or Digital Forgery? Investigating Data Replication in Diffusion Models, arXiv (Dec. 7, 2022), https://arxiv.org/abs/2212.03860.

[22] Comment of OpenAI, Request for Comments on Intellectual Property Protection for Artificial Intelligence Innovation, Docket No. PTO-C-2019-0038, 84 Fed. Reg. 58,141 (Dec. 16, 2019), https://www.uspto.gov/sites/default/files/documents/OpenAI_RFC-84-FR-58141.pdf.

[23] See [16]

[24] See [19]

[25] Kadrey v. Meta Platforms, Inc., No. 3:23-cv-03417-VC, 2025 WL 4123456 (N.D. Cal. June 25, 2025), https://law.justia.com/cases/federal/district-courts/california/candce/3:2023cv03417/415175/598/.[26] Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith, 598 U.S. 508 (2023), https://www.supremecourt.gov/opinions/22pdf/21-869_87ad.pdf. 

[27] See [25]

[28] Section 512 of Title 17: Resources on Online Service Provider Safe Harbors and Notice-and-Takedown System | U.S. Copyright Office, https://www.copyright.gov/512/ (last visited Apr. 22, 2026).

[29] Eleonora Rosati, The Exception for Text and Data Mining (TDM) in the Proposed Directive on Copyright in the Digital Single Market — Technical Aspects, Eur. Parl. Doc. IPOL_BRI(2018)604942 (2018), https://www.europarl.europa.eu/RegData/etudes/BRIE/2018/604942/IPOL_BRI(2018)604942_EN.pdf.

[30] Luca Motoarca, AI, Copyright, and Pseudo Art, 26 Yale J.L. & Tech. 430 (2024), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4456742.  

[31] U.S. Copyright Office, Compendium of U.S. Copyright Office Practices ch. 300 (3d ed. 2021).

[32] See [31]

[33] 17 U.S.C. §§ 101 et seq.

[34] Thaler v. Perlmutter, No. 23-5233 (D.C. Cir. Mar. 18, 2025), https://media.cadc.uscourts.gov/opinions/docs/2025/03/23-5233.pdf.

[35] See [31]

[36] See [34]

[37] See [34]

[38] Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence, 88 Fed. Reg. 16,190 (Mar. 16, 2023), https://www.federalregister.gov/documents/2023/03/16/2023-05321/copyright-registration-guidance-works-containing-material-generated-by-artificial-intelligence.

[39] Zarya of the Dawn: How AI Is Changing the Landscape of Copyright Protection, Harvard Journal of Law & Technology (Mar. 6, 2023), https://jolt.law.harvard.edu/digest/zarya-of-the-dawn-how-ai-is-changing-the-landscape-of-copyright-protection.

[40] U.S. Copyright Off., Zarya of the Dawn: Registration Decision (Feb. 21, 2023), https://www.copyright.gov/docs/zarya-of-the-dawn.pdf.

[41] See [30]

[42] See [44]

[43] See [8]

[44] See [30]