Artificial Intelligence, Autonomous Creation, and the Future Path of Copyright Law

Abstract

Most literature at the intersection of copyright and artificial intelligence (AI) has focused primarily on what copyright law is or ought to be. Frequently overlooked is the question of what copyright law will be in the AI space. Understanding this question is crucial because the path of copyright law chosen by the United States will have a major impact on the country's economic and technological future. This Article begins by scrutinizing two lines of arguments that have been advanced to deny copyright protection to AI-generated works: constitutional and incentivebased. The Article then discusses a third line of arguments- harmonization-based arguments-and identifies select instances in which Congress matched the protection offered by other jurisdictions or declined to do so. This Article further shows that global copyright law developments have slowly diverged in the AI space. In view of these growing divergences, U.S. legislators and policymakers are now confronted with a key policy choice at the intersection of copyright and AI: should the United States retain existing approaches, follow other jurisdictions, or work with these jurisdictions to develop harmonized AI-related international copyright standards? To inform the future debate on copyright and AI, the second half of this Article highlights the different areas in which substantial copyright law and policy reform may emerge in the AI space. It further discusses four options the United States can take to shape the future path of copyright law: (1) international treaty negotiations; (2) softlaw instruments; (3) a global multi-stakeholder dialogue; and (4) choice-of-law principles.

Headnote

Most literature at the intersection of copyright and artificial intelligence (AI) has focused primarily on what copyright law is or ought to be. Frequently overlooked is the question of what copyright law will be in the AI space. Understanding this question is crucial because the path of copyright law chosen by the United States will have a major impact on the country's economic and

(ProQuest: ... denotes non-USASCII text omitted.)

INTRODUCTION

The release of ChatGPT 4, Dall-E, Midjourney, Stable Diffusion, and other generative artificial intelligence (AI) tools initiated a fastexpanding public debate on AI technologies and their societal impacts.1 Since June 2023, the U.S. Senate Judiciary Committee has held ten public hearings on AI developments, covering intellectual property, human rights, regulatory issues, governance and oversight, and deepfakes during political elections.2 Building on past consultations on AI-related intellectual property issues,3 the U.S. Copyright Office and the U.S. Patent and Trademark Office have also conducted public listening sessions and new consultation processes.4 Outside the United States, there have been many notable developments, inclArtificial Intelligence Act (AI Act) in the European Union5 and the Interim Measures for the Management of Generative Artificial Intelligence Services in China6 and the continuous efforts undertaken by the World Intellectual Property Organization (WIPO) at the intersection of intellectual property and AI.7

Thus far, courts and government agencies in the United States, the European Union, and many other jurisdictions have chosen not to extend copyright protection to works autonomously generated by AI systems.8 In the United States, for example, the Copyright Office declined to register Stephen Thaler's artwork "A Recent Entrance to Paradise,"9 Kris Kashtanova's illustrations in the comic book Zarya of the Dawn,10 Jason Allen's award-winning image "Théâtre D'opéra Spatial,"11 and Ankit Sahni's India-based AIgenerated artwork, Suryast.12 Although one could debate theuding the recent adoption of the advisability of these decisions, the decisions were well supported by the Copyright Clause, which empowers Congress to "promote the Progress of Science . . . by securing for limited Times to Authors . . . the exclusive Right to their respective Writings."13 The 1976 Copyright Act also includes statutory language such as "children," "widow," "grandchildren," and "widower," suggesting that copyright protection is intended for humans, not machines.14

Constitutional arguments aside, one could muster incentivebased arguments against the introduction of greater copyright protection for AI-generated works.15 As Pamela Samuelson observed as early as four decades ago: "[I]t simply does not make any sense to allocate intellectual property rights to machines because they do not need to be given incentives to generate output. All it takes is electricity (or some other motive force) to get the machines into production."16 Other intellectual property commentators hold similar views.17 Moreover, there is thus far no empirical evidence that the AI sector has been underfunded. With active participation in AI development by Anthropic, Open AI, and Stability AI, as well as technology giants such as Alphabet, Amazon, Apple, Meta, and Microsoft, extending copyright protection to AI-generated works is more likely to provide windfalls than the economic incentives AI developers claim they badly need.

Unfortunately, legislators, policymakers, and commentators sometimes focus so much on what the law is or ought to be and ignore what the law will be, bringing to mind H.L.A. Hart's admonition about the important distinction between what the law is and what it ought to be.18 The development of copyright law is no exception. Most literature at the intersection of copyright and AI has focused primarily on what copyright law is or ought to be. This Article aims to fill the gap in the literature by exploring the frequently overlooked question of what copyright law will be in the AI space. It takes note of AI-related copyright law developments from around the world and discusses how the United States could chart a better future path of copyright law that would not only strengthen the country's economic capabilities, technological developments, and global competitiveness, but also make U.S. copyright law ready for future technological demands and challenges.

To demonstrate how courts, government agencies, legislators, and commentators have thus far focused on what copyright law in the AI space is or ought to be, as opposed to what that law will be, Part I scrutinizes two lines of arguments that have been advanced to deny copyright protection to AI-generated works: constitutional and incentive-based. This Part then discusses a third line of arguments-harmonization-based arguments-and identifies select instances in which Congress matched the protection offered by other jurisdictions or declined to do so. Showing that harmonization-based arguments have sometimes prevailed over constitutional and incentive-based arguments, this Part calls for greater policy and scholarly attention to AI developments from around the world.

Part II shows that global copyright law developments have slowly diverged in the AI space-a development that has been of great interest to the U.S. Copyright Office, legislators, and policymakers.19 While the copyrightability of AI-generated work has become a key point of departure in the international community, disputes and lawsuits over the legality of using copyrighted works to train AI models without authorization have also revealed the different approaches taken across the world.20 In view of these growing divergences, this Part argues that U.S. legislators and policymakers are now confronted with a key policy choice at the intersection of copyright and AI: should the United States retain existing approaches, follow other jurisdictions, or work with these jurisdictions to develop harmonized AI-related international copyright standards?

Part III highlights five additional areas in which substantial copyright law and policy reform may emerge in the AI space: (1) the idea-expression dichotomy; (2) access and substantial similarity; (3) fair use; (4) secondary liability; and (5) formalities. Although it is unclear which of these seven areas will encounter future legislative and policy reform, it is not far-fetched to assume that AI developments will have serious ramifications for all seven areas. This Part therefore explores how these areas will become important in the AI space and offers insights into how copyright law may change in the future.

Part IV explores how the United States could chart a better future path of copyright law. To appreciate the possibility of different approaches that are equally justifiable from both the legal and policy standpoints, this Part illustrates this possibility using a hypothetical future scenario in which an AI system that is capable of autonomously generating creative works has been used by three separate owners. This Part then discusses four options the United States can take to shape the future path of copyright law: (1) international treaty negotiations; (2) softlaw instruments; (3) a global multi-stakeholder dialogue; and (4) choice-of-law principles. Taking note of the different approaches taken by past U.S. administrations, this Part reviews options that can succeed with or without international engagement. It further reminds us that the United States could choose to partially harmonize its copyright standards with those of other jurisdictions or not alter these standards at all.

This Article concludes by pointing out that at this nascent stage of generative AI development, it is simply too early to tell whether AI-related copyright law developments will converge or diverge in the future. Path divergence, as we have already seen in relation to the copyrightability of AI-generated works, is not a forgone conclusion. Nevertheless, because of technological path dependency, U.S. legislators and policymakers should start embracing holistic, macro-level, and comparative thinking to facilitate the development of appropriate copyright law in the AI space.

I. AI AND THE PATH OF COPYRIGHT LAW

The debate on whether intellectual property protection should be extended to AI-generated creations first received considerable policy and scholarly attention in the patent area, not the copyright area.21 Shortly before the COVID-19 pandemic, in 2019, Stephen Thaler, with the assistance of law professor Ryan Abbott, began seeking protection for an invention developed by an intelligent machine named DABUS (Device for the Autonomous Bootstrapping of Unified Sentience).22 Patent agencies and courts around the world-including those in the European Union, the United Kingdom, and the United States-declined to grant patent protection to the invention.23 However, a court in Australia (at least initially) and another in South Africa recognized such protection.24 Commentators quickly dismissed the decisions in these two jurisdictions as outliers by noting the unique statutory language in the Australian Patents Act 199025 while pointing out that the patent reregistration system in South Africa does not require substantive examination.26 Regardless of the position taken initially by the Federal Court of Australia, that court's first decision was eventually overturned by the Full Court of the Federal Court, and the High Court of Australia (the country's highest court) declined to review the later decision.27

In the copyright area, no country has thus far extended protection to AI as an author or coauthor. The closest decision was made by the Indian Copyright Office, when it reportedly registered an AI-coauthored artwork, Suryast, in August 2021.28 Nevertheless, that office withdrew the registration three months later.29 In the past three years, however, the landscape in the area surrounding the copyrightability of AI-generated works has slowly changed, as Asian countries began to offer, or consider to offer, copyright or sui generis protection to these works.30

This Part scrutinizes two lines of arguments that courts, government agencies, legislators, and commentators have made to support the denial of copyright protection to AI-generated works. Section A focuses on constitutional arguments, and section B turns to incentive-based arguments. Although these arguments have been persuasive and could contribute to developing sound copyright policy, they alone do not determine the future path of copyright law, especially in a world where countries continue to compete with each other to achieve dominance in the AI sector.31

To help us appreciate the different future paths of copyright law, section C discusses select instances in which Congress matched the protection offered by other jurisdictions or declined to do so. This section first discusses the United States' decision not to offer sui generis database protection despite the adoption of the EU Directive on the Legal Protection of Databases32 (EU Database Directive) in 1996. It then contrasts this decision with the United States' reluctant acceptance of foreign intellectual property standards as compromises struck in the Agreement on Trade- Related Aspects of Intellectual Property Rights33 (TRIPS Agreement) and TRIPS-plus bilateral, regional, and plurilateral agreements.34 Showing that harmonization-based arguments have sometimes prevailed over constitutional and incentive-based arguments, this Part calls for greater policy and scholarly attention to AI-related copyright law developments from around the world.

A. Constitutional Arguments

Article I, Section 8, Clause 8 of the U.S. Constitution provides that "Congress shall have Power . . . to promote the Progress of Science . . . by securing for limited Times to Authors . . . the exclusive Right to their respective Writings."35 Derived from proposals introduced by James Madison and Charles Pinckney,36 the Copyright Clause was adopted in its final form without any debate.37 As a brief and ambiguous passage in The Federalist suggests,38 copyright was a "comparatively insignificant" issue in the public debate over the ratification of the proposed Constitution.39

Pursuant to this enumerated power, Congress enacted the first copyright statute in 1790 and subsequent copyright acts in 1831, 1870, 1909, and 1976.40 Although the word "authors" in the Copyright Clause can be interpreted to cover both humans and machines, subsequent case law and statutory texts clarify that the word refers to human authors. For example, in Burrow-Giles Lithographic Co. v. Sarony, the Supreme Court defined "author" as "he to whom anything owes its origin" and referred to copyright "as the exclusive right of a man to the production of his own genius or intellect."41 The 1976 Copyright Act also includes language such as "children," "widow," "grandchildren," and "widower."42 As the U.S. Court of Appeals for the Ninth Circuit explained in Naruto v. Slater (the now-famous monkey selfie case), a monkey "lack[s] statutory standing to sue under the Copyright Act" because "[t]he terms 'children,' 'grandchildren,' 'legitimate,' 'widow,' and 'widower' all imply humanity and necessarily exclude animals that do not marry and do not have heirs entitled to property by law."43

Thus, the Copyright Clause, statutory text, and prior case law have provided strong support to the argument that copyright protection is intended for humans, not machines.44 Reflecting this humanity limitation, Section 313.2 of the Compendium of U.S. Copyright Office Practices states that the Copyright 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."45 Based on this position, the Copyright Office has rejected the application for registering the copyright in AI-generated works, such as the images developed by Stephen Thaler, Kris Kashtanova, Jason Allen, and Ankit Sahni.46

Notwithstanding these rejections, the language used in Burrow- Giles, the 1976 Act, and the Compendium are arguably open to interpretation in the AI context. For instance, even though the Supreme Court defined "author" to be "he to whom anything owes its origin,"47 it clearly does not mean that female, non-gendered, and other authors will not receive copyright protection. As Arthur Miller explained more than three decades ago:

It is tempting to read [the Court's] language as requiring that a copyrighted work be created by a human author. However, it is implausible that the Court was considering that question in these isolated passages, let alone answering it. The Justices were dealing with the technology of the nineteenth century, and hardly with the question of whether a machine might be capable of "intellectual conceptions." There are limits to literal reading. By making references to "he" and "man," the Court was no more excluding machines from eligibility for authorship than it was excluding women. There simply is less than meets the eye in the language of the opinion.48

Indeed, if the 1884 language could be interpreted dynamically to cover not only "he" but also "she" and "they," one could certainly argue that the word "authors" could be interpreted more broadly in the twenty-first century to cover AI systems-or, at least, AI-generated works that involve some human creative input. After all, the Burrow-Giles Court declared the following after defining the word "authors": "We entertain no doubt that the Constitution is broad enough to cover an act authorizing copyright of photographs, so far as they are representatives of original intellectual conceptions of the author."49 To this Court, the key to copyrightability is the "[o]riginal intellectual conceptions," not whether the author is male or human.50

Read closely, Section 313.2 of the Compendium seems to suggest that the Copyright Office is open to registering works produced by a machine in the presence of "creative input or intervention from a human author."51 If so, the difficult question for the Copyright Office, and in turn the courts, will be the amount of human creative input or intervention needed to render an AI-generated work eligible for copyright protection. Courts and government agencies from around the world, including in the United States, seem to have now paid greater attention in this direction.52 A case in point is the Beijing Internet Court's decision to evaluate copyrightability based on the prompts and parameters used to create an AI-generated image.53 Part I.B will further discuss this and other decisions.54 In two recent articles, Edward Lee heavily criticized the Copyright Office's restrictive interpretation of the Constitution and the 1976 Copyright Act in determining the copyrightability of AIgenerated works.55 As he elaborates:

[The Office] imposes newfound requirements of so called "traditional" authorship-the prediction of results ahead of time, dictation of specific results, and avoidance of randomness-found nowhere in the [Copyright] Clause, Supreme Court precedent, or the Copyright Act. This misguided approach ignores the [Copyright] Clause's important goal to promote progress by incentivizing creators to use their intellectual efforts to create new works for the public's benefit.56

Calling for greater attention to the first principles of the Copyright Clause, Professor Lee believes "creators should receive copyrights for their AI-prompted works that embody a minimally creative selection or arrangement of elements."57 He nonetheless argues that the scope of protection should be "very thin" and should target only verbatim copying.58

In sum, there are strong constitutional arguments against the extension of copyright protection to AI-generated works. These arguments have paved the way for a more restrictive interpretation of the Copyright Clause and the 1976 Copyright Act, as evident in decisions made recently by the Copyright Office and the U.S.

District Court for the District of Columbia.59 Notwithstanding their strength, these arguments are surmountable. In the past few years, commentators have begun to identify new ways to broaden the interpretation of the Copyright Clause to extend copyright protection to AI-generated works. To the extent that Congress has found such protection to be aligned with the country's best interests, it is very doubtful that the legislature will refrain from expanding copyright protection in the AI space. After all, Congress and courts have rejected constitutional arguments in other intellectual property contexts, such as the restoration of copyright protection to works in the public domain, the extension of the copyright term in the late 1990s, and the switch of the patent system from first-to-invent to first-inventor-to-file in the early 2010s.60

B. Incentive-Based Arguments

The second line of arguments advanced to justify the denial of copyright protection to AI-generated works concerns economic incentives. The dominant justification for U.S. copyright law is the need to provide economic incentives to induce authors to participate in the creative process.61 As the Supreme Court declared in Twentieth Century Music Corp. v. Aiken: "The immediate effect of our copyright law is to secure a fair return for an 'author's' creative labor. But the ultimate aim is, by this incentive, to stimulate artistic creativity for the general public good."62 The Court's emphasis on economic incentives is easy to understand:

Without copyright protection . . . most professional authors and their investors will not [be] able to recoup the time, effort, or resources expended in the creative process, and society will suffer as a result. Copyright therefore ensures that authors participate in the creative process, rather than in other, more remunerative activities.63

Thus far, many commentators have taken the position that AIgenerated works do not deserve copyright protection due to the fact that AI systems do not need economic incentives to create.64

In response to these incentive-based arguments, those demanding stronger copyright protection for AI-generated works point out that such protection is important not because AI developers need economic incentives but because of the considerable resources they expend to develop AI. Laws should therefore be introduced to enable these developers to recoup their investments.65 With the right protection-whether through intellectual property or sui generis rights-some developers may be willing to reinvest portions of their profits to promote AI development, thereby creating a socially beneficial virtuous cycle.66

In the past few decades, countries have also offered or considered intellectual property rights to protect investments, regardless of whether economic incentives are needed. Immediately coming to mind is the sui generis protection of databases under the EU Database Directive, which provides fifteen years of protection to non-original, non-creative databases that are created as a result of "a substantial investment in either the obtaining, verification or presentation of [database] contents."67 Such protection aims to prevent the unauthorized "extraction and/or re-utilization of the whole or of a substantial part" of these contents.68 In October 2017, the European Commission also proposed to create a new data producer's right for non-personal, anonymized machine-generated data.69 Although the Commission subsequently abandoned this proposal,70 and developed the Data Act instead,71 the justifications for protecting data-driven investments are similar to those offered to support greater sui generis protection of AI-generated works.72

In addition, many countries, especially those in the developing world, have established investment protection regimes to offer special rights to foreign investors, including those in the AI sector.73 In my earlier analysis of the European Commission's proposal for the data producer's right, I discussed the possibility for data producers to use investment protection regimes to protect their business.74 Although the use of investor-state dispute settlement in the intellectual property context has dramatically slowed down since the COVID-19 pandemic,75 nothing in investment protection regimes would prevent AI developers from filing investor-state complaints to resolve disputes in this sector. Indeed, the lack of support for intellectual property rights in AI-generated works and the high costs of AI development may give foreign AI developers perverse incentives to aggressively demand protection through investment protection regimes.

So far, these counterarguments have not convinced policymakers and commentators to extend copyright protection to AI-generated works. For many, the goal of intellectual property rights is to provide economic incentives, not to protect investments. Indeed, focusing on the latter at the expense of the former can be ill-advised, if not dangerous. In many developing countries, policymakers have been affected by what I have called the "incentive-investment divide."76 They have been so focused on attracting foreign investment that they have often lost sight of the need to strike an appropriate balance in the domestic innovation system.

Since the mid-1930s, commentators have also widely criticized what Rochelle Dreyfuss has coined the "if value, then right" logic.77 Legal philosopher Felix Cohen is one of the earliest critics of such logic, calling it "transcendental nonsense."78 As he explained, this logic "purports to base legal protection upon economic value, when, as a matter of actual fact, the economic value . . . depends upon the extent to which it will be legally protected."79 In his view, the reasoning is circular. Using the value of AI technology to justify stronger copyright protection for AI-generated works runs into the same circularity problem. Whether those works deserve additional licensing fees from AI developers depends on whether copyright law grants protection in the first place.

As if these rebuttals were not strong enough, there has been no indication to date that the AI sector has been underfunded. With the developments concerning ChatGPT, Dall-E, Midjourney, Stable Diffusion, and other generative AI tools, as well as the active participation of technology giants like Alphabet, Amazon, Apple, Meta, and Microsoftin the AI sector,80 it is unclear that society will need to provide additional incentives to foster development of this sector. Should copyright protection be extended to AI-generated works, such protection would provide mostly windfalls, rather than the economic incentives AI developers claim they badly need.

In sum, there have been economic arguments both for and against greater copyright protection for AI-generated works. While arguments concerning the need for economic incentives have not been persuasive, other arguments do exist. To the extent that such greater protection is deemed to be in the country's best interests, new economic arguments will emerge to support such protection. These arguments are likely to focus on the economic benefits provided by a strong AI sector in the United States and the country' global leadership in this sector.

C. Harmonization-Based Arguments

Thus far, courts, government agencies, legislators, and commentators have focused on constitutional and incentive-based arguments. Frequently overlooked in the debate on copyright and AI is the United States' need to adjust or harmonize its domestic copyright standards to preserve global competitiveness in the AI sector. Such adjustment requires observers to understand copyright law developments in other parts of the world, especially those involving the country's major competitors. As Anu Branford reminds us in her new book Digital Empires, the United States is constantly engaging in battles with China and the European Union to exert influence over global technology regulation.81 A key battleground for such influence is AI.

There are several reasons why countries choose to harmonize their legal standards with those in other parts of the world. For example, the development of such standards would help make the business environment more predictable, thereby reducing the transaction costs of conducting business across borders.82 By establishing common rules and reducing destructive protectionism, the harmonization process would also promote trade and stability in the international community.83 In addition, such a process would help achieve economies of scale in the governance and administration of legal rights-in this case, intellectual property rights.84

Nevertheless, harmonization comes with significant drawbacks, and diversification can be beneficial in the right circumstances. For instance, diversification allows countries to avoid adopting "one size fits all" legal standards that do not sit well with local conditions.85 It also enables countries to experiment with new regulatory and economic policies.86 In addition, diversification would facilitate interjurisdictional competition that checks governmental inefficiency and abuse.87 Such standards would, in turn, make the lawmaking process more accountable to the local population.88

Because of the varying strengths and weaknesses of harmonization and diversification, in the past three decades the United States has both matched the protection offered by other jurisdictions and declined to do so. These past experiences are instructive in the AI context.

A leading example of the United States' decision not to harmonize its standards with those of leading trading partners is its refusal to offer sui generis database protection, notwithstanding the reciprocal protection the EU Database Directive would offer to American database producers.89 As commentators have noted, there are many reasons for the United States' refusal to offer protection in this area.90 For comparison purposes, consider the constitutional and incentive-based arguments. In Feist Publications, Inc. v. Rural Telephone Service Co., the Supreme Court declared that "[o]riginality is a constitutional requirement."91 Thus, extending copyright protection to non-original databases could be problematic from a constitutional standpoint. There were also concerns that such protection would raise serious questions under both the Commerce Clause and the First Amendment.92 From an economic standpoint, there were fears that the costs of sui generis database protection would outweigh its benefits. For instance, the protection "would confer far broader and stronger exclusive rights on database content than is necessary to provide the needed incentives for database producers."93 By granting a monopoly over collected data, such protection would also allow private actors to lock up information that was essential to basic scientific research and future creative endeavors.94 The introduction of sui generis database protection would even establish an anticompetitive environment that would make it difficult for value-adding products and services to enter the market.95 Such an environment would, in turn, make information products more expensive, thereby harming consumers and society at large.

Thus, to the extent that the debate at the intersection of copyright and AI resembles the earlier debate on sui generis database protection, the United States is unlikely to follow the lead of other jurisdictions, including China and the European Union. After all, denying copyright protection to AI-generated works can be a comparative advantage. In relation to database protection, for example, the United States has made the correct decision. As the European Commission declared in its comprehensive evaluation of the EU Database Directive shortly before the Directive's tenth anniversary:

Between 2002 and 2004, the European share decreased from 33% to 24% while the US share increased from 62% to 72%. The ratio of European/US database production, which was nearly 1:2 in 1996, has become 1:3 in 2004. . . . According to the Gale Directory of Databases, the number of EU-based database "entries" was 3095 in 2004 as compared to 3092 in 1998 when the first Member States had implemented the "sui generis" protection into national laws. More significantly, the number of database "entries" dropped just as most of the EU-15 Member States had implemented the Directive into national laws in 2001. In 2001, there were 4085 EUbased "entries" while in 2004 there were only 3095.96

Although the United States declined to follow the European Union's lead in offering sui generis database protection, the past three decades have seen many examples in which the country reluctantly accepted the standards found in other jurisdictions. Notable examples are the "importation" of intellectual property standards through international trade agreements, such as the TRIPS Agreement and TRIPS-plus bilateral, regional, and plurilateral trade agreements.97 A case in point is Section 514 of the Uruguay Round Agreements Act, which restored copyright protection to works that had fallen into the public domain due to their failure to comply with the formalities requirement under the 1976 Copyright Act.98 To comply with trade rules under the World Trade Organization (WTO), the United States also changed the term of patent protection from seventeen years from the date of grant to twenty years from the date of application.99

Within domestic law, the United States has unilaterally harmonized its laws with those of the European Union. Leading examples include the acceptance of trademark applications based on "bona fide intention to use" in 1988,100 the extension of the copyright term in 1998 following the adoption of the EC Copyright Term Directive,101 and the switch of the patent system from first-tofile to first-inventor-to-file in 2011 upon the adoption of the Leahy- Smith America Invents Act.102

To the extent that the debate at the intersection of copyright and AI resembles the debate on international trade agreements and other international harmonization efforts, the United States has proven its willingness-though at times grudgingly-to amend its intellectual property laws in the image of the laws of other jurisdictions. So far, there is no indication why the United States would not behave the same way in the AI space. Indeed, if the past is any guide, the United States tends to harmonize its laws when they are out of sync with those of its major competitors.103

In addition to the benefits of harmonization discussed above, there is a strong strategic reason why countries would follow competing jurisdictions in areas with high stakes and substantial uncertainty, like the AI space. If it is unclear whether the new intellectual property standards adopted in these jurisdictions would strengthen AI development, the best course of action might be to follow suit. Even if those standards later prove to be inferior, the follow-on effort will enable the jurisdiction at issue to become no worse offand, in turn, to avoid a potentially costly gamble.

D. Summary

There are interesting questions about whether the United States should follow the lead of other jurisdictions in undertaking copyright reform in response to challenges posed by AI technology. Although policymakers and commentators have repeatedly advanced constitutional and incentive-based arguments in this area, past experiences show that Congress has harmonized U.S. laws with those of other jurisdictions when it believes that such harmonization would be in the country's best interests. In many of these instances, harmonization-based arguments have prevailed over constitutional and incentive-based arguments.

With the strategic importance of AI developments and the strong likelihood that AI-based products and services will become a very significant part of international trade in the near future, policymakers and commentators should take harmonization-based arguments seriously. By showing the existence of a new line of arguments that complements more traditional constitutional and incentive-based arguments, this Article encourages U.S. legislators and policymakers-and, for that matter, those of other jurisdictions-to start embracing holistic, macro-level, and comparative thinking to facilitate the development of appropriate copyright law in the AI space.

II. GROWING PATH DIVERGENCE

Despite their similarities across the world,104 copyright law developments have diverged in the AI space in recent years, to varying degrees. For instance, although no jurisdiction thus far has recognized AI as an author within the meaning of copyright law, Chinese courts have now extended copyright protection to AIgenerated works, unlike the U.S. Copyright Office and the U.S. District Court for the District of Columbia.105 Similarly, even though U.S. courts have long considered the non-expressive use of copyrighted works as fair use,106 recent copyright litigation against AI developers in the United States and other jurisdictions has called into question whether one would draw the same conclusion in fair use analysis when the unauthorized use of copyrighted works for AI training would create competing works.107 Outside the United States, countries have also adopted narrower copyright exceptions for text and data mining (TDM) or computational data analysis,108 raising questions about whether the operation of these exceptions in the AI context would resemble that of the U.S. fair use provision.109

This Part discusses the growing divergence in two key areas of copyright law developments concerning AI. Section A focuses on the copyrightability of AI-generated works. Section B turns to the different legislative approaches used to facilitate the use of copyrighted works to train AI models. Although countries seem to have reached an international equilibrium for developing copyright exceptions to facilitate such training, which suggests some level of convergence, the ways in which these exceptions are designed continue to diverge.110 Both areas have therefore presented U.S. legislators and policymakers a key policy choice at the intersection of copyright and AI: should the United States retain existing approaches, follow other jurisdictions, or work with these jurisdictions to develop harmonized AI-related international copyright standards?

A. Copyrightability of AI-Generated Works111

Although copyright laws in some jurisdictions, such as Ireland, New Zealand, and the United Kingdom, have extended copyright protection to computer-generated works,112 no country thus far has recognized AI as an author or a coauthor. Nevertheless, the past few years have seen some interesting developments in Asia that suggest growing protection of AI-generated works through copyright or sui generis rights.

In China, there are three notable cases in which creators managed to obtain protection for their AI-generated works. The first case is Beijing Film Law Firm v. Beijing Baidu Netcom Science Technology Co.113 Decided in April 2019, this case concerned the infringement on an automatically generated analytical report posted on the plaintifflaw firm's WeChat account. Although the court found that the report did not constitute a work within the meaning of the Chinese Copyright Law, it noted that the plaintiffhad some legal rights in the report. It therefore required the search engine defendant Baidu to pay the plaintiffCNY 1,560 (about USD 200) in compensation.

The second case is Shenzhen Tencent Computer System Co. v. Shanghai Yingxun Technology Co.114 Widely cited in the AI context, this case involved the infringement of a finance news report written autonomously by the plaintiff's smart writing assistant software, Dreamwriter. Although the Shenzhen Nanshan District Court did not allow AI to be an author, it recognized, for the first time, the copyright in an AI-generated work based on human contributions. As the court declared, "Based on factual findings, it is apparent that the arrangement and selection by the plaintiff's creative team of the data input, trigger condition setting, template, and choice of corpus style are intellectual activities directly related to the specific form of expression of the article involved in this case."115 The court also noted that the plaintiff's creative staffwho used the Dreamwriter to generate financial news reports included an editorial team, a product team, and a technology development team. Meanwhile, the process of generating these reports went through four stages: data service, triggering and writing, smart verification, and smart distribution. Finding that the defendant had infringed on the plaintiff's news report, the court, like the one in the previous case, required the defendant to pay the plaintiffCNY 1,500 (about USD 200) in compensation.

The most recent case is Li v. Liu.116 Decided in November 2023, this case concerned the infringement of an image generated by Stable Diffusion entitled "Spring Breeze Has Brought Tenderness," which the plaintiffposted on the Chinese social media platform Little Red Book (Xiaohongshu). Contrary to the U.S. Copyright Office's decision not to register the images in Zarya of the Dawn, the Beijing Internet Court found the AI-generated image at issue to be copyrightable under the Chinese Copyright Law. As the court explained:

The plaintiff's design of the character, its mode of presentation, and other image elements through prompts and his setup of the image's layout and composition through parameters reflect his selection and arrangement. Moreover, after inputting prompts and setting relevant parameters to obtain the initial image, the plaintiffcontinued to provide additional prompts, adjust parameters, and constantly refine the image before finally obtaining the image at issue. This process of adjustment and refinement also reflects the plaintiff's aesthetic choice and personal judgment.117

The court further explored whether the copyright holder should be the plaintiffor the designer of the AI system, considering that the Chinese Copyright Law does not allow the system itself to be an author. Finding the plaintiffto be the author and the image to be protected as a work of art under the Law, the court required the defendant to pay the plaintiffCNY 500 (about USD 70) in compensation.

Taken together, these three cases suggest that we may begin to see abroad more copyright protection for AI-generated works, especially if other courts follow the Chinese courts' approach. These cases also invite us to explore whether there is any appropriate basis for such protection and what key arguments, including those explored in Part I, exist to support such protection.118

As if the diverging court decisions from China had not made the debate on copyrightability of AI-generated works complicated enough, other Asian countries have been exploring proposals to extend protection to these works. For example, in February 2021, a proposal was submitted to the South Korean legislature to amend the copyright law to provide sui generis protection to AI-generated works. As a report from the law firm Dentons recounts:

The proposal adds the new definition of an "author of a work created by using an AI service" which is determined based on the level of contribution to the creation. By stipulating that an AI developer who created an algorithm or a human artist who provided data for AI learning may also be a copyright owner, the proposal seeks to find an equitable balance for protecting the contributors. The proposal also states that a copyright for AIcreated work lasts for 5 years after public disclosure, and for the registration process, its author needs to specify that the work was created by AI.119

Commentators in Asia have also called for the provision of sui generis protection to AI-generated works.120 It remains to be seen whether countries in that region will eventually accept such proposals-and, if so, whether other jurisdictions will follow suit.

In sum, there has been a major push for stronger protection of AI-generated works in Asia. Such a push is unsurprising for several reasons. To begin with, policymakers in many Asian countries believe that AI technology will drive their countries' economic and technological developments and global competitiveness. To the extent that these policymakers believe that stronger protection for AI-generated works will help provide the incentives for AI innovation or attract the investments needed to fuel such innovation, they will actively push for new laws and policies to offer such protection. In China, for instance, AI has featured prominently in the country's strategic plans for economic, social, scientific, and technological developments.121 Particularly ambitious is the country's goal of becoming the world's major AI innovation center by 2030,122 not to mention its ongoing AI race with the European Union, the United States, and other major technological powers.123

In addition, the past decades have seen repeated external pressure on Asian countries to strengthen the protection and enforcement of intellectual property rights. The demands for intellectual property reforms from foreign investors and their supportive governments have also created the misimpression that the key goal of intellectual property rights is to protect investments, as opposed to generating incentives for creation.124 To some extent, the origins of modern intellectual property laws in Asia, regardless of their historical roots, were somewhat different from those in Europe and the United States.125 Now that many Asian countries have made the difficult transition to implement a strong (or stronger) intellectual property regime with high (or higher) levels of protection and enforcement, mirroring the standards found in the West, it may be too difficult and too late to convince Asian policymakers that AI-generated works should stay unprotected in the public domain.

As I noted more than a decade ago, "In the near future, the economic structure of the United States-and for that matter, members of the European Union, Japan, or other existing developed countries-could depend more on innovation than existing forms of intellectual property rights."126 With the rise of generative AI and other cutting-edge technologies, many developed countries may now want a very different intellectual property system in Asia than the one they pushed Asian countries to adopt over the past few decades. It will be interesting to see whether Asian policymakers can quickly adapt to this changing mindset or will become eager to do so. If not, the system they embrace may focus less on the challenges and opportunities created by AI and other new technology than on issues raised by the more traditional protection of pharmaceuticals, computer software, and entertainment products.

B. Use of Copyrighted Works to Train AI Models

Another set of issues that the debate at the intersection of copyright and AI has frequently explored concerns the unauthorized use of copyrighted works to train AI models. Such unauthorized use has attracted some high-profile lawsuits brought by individual creators and content producers.127 In the United States, for example, plaintiffs have included Getty Images, the New York Times, and celebrity authors such as John Grisham and George R.R. Martin.128

To defend these lawsuits, U.S. AI developers rely heavily on the fair use provision in Section 107 of the 1976 Copyright Act.129 Because only a few countries such as "Israel, Liberia, Malaysia, the Philippines, Singapore, South Korea, Sri Lanka, and Taiwan have adopted the fair use regime or its close variants,"130 most countries across the world do not have an open-ended system of copyright limitations and exceptions. Instead, they have a closed system that allows them to provide copyright limitations and exceptions, such as a copyright exception for TDM or computational data analysis.131

For instance, the United Kingdom and the European Union have introduced specific copyright exceptions for TDM. Section 29A(1) of the UK Copyright, Designs and Patents Act 1988 permits "[t]he making of a copy of a work by a person who has lawful access to the work" in order to "carry out a computational analysis of anything recorded in the work for the sole purpose of research for a non-commercial purpose."132 Articles 3 and 4 of the EU Directive on Copyright and Related Rights in the Digital Single Market (DSM Directive) also allow for TDM,133 which is defined as "any automated analytical technique aimed at analysing text and data in digital form in order to generate information which includes but is not limited to patterns, trends and correlations."134 In addition, Article 9.2 of the TRIPS Agreement states: "Copyright protection shall extend to expressions and not to ideas, procedures, methods of operation or mathematical concepts as such."135 This provision helps facilitate the nonexpressive use of copyrighted works around the world, including the training of AI models.136

In the debate on copyright exceptions for AI training, it is often overlooked that Japan introduced the world's first TDM exception in 2009,137 long before the adoption of similar exceptions in the United Kingdom and the European Union. The UK and EU TDM exceptions were not introduced until 2014 and 2019, respectively.

Article 30-4 of the Japanese Copyright Act, which was adopted in 2018 and consolidated and expanded from the original TDM exception in Article 47-7,138 states that "[i]t is permissible to exploit a work, in any way and to the extent considered necessary, in any of the following cases, or in any other case in which it is not a person's purpose to personally enjoy or cause another person to enjoy the thoughts or sentiments expressed in that work."139 The "following cases" language in this provision refers to (1) technology testing for development or "put[ting technology] into practical use"; (2) "data analysis"; and (3) "computer data processing."140 Because using copyrighted works to train AI models arguably does not involve the enjoyment of the thoughts or sentiments expressed in those works, there should be no copyright liability in most circumstances.

When Japan revised the Copyright Act in 2018, it also added new copyright exceptions to support the "exploitation of works incidental to the [use] of works on a computer" and the "minor exploitation incidental to computerized data processing and the provision of the results thereof."141 These exceptions work in tandem with Article 30-4 of the Japanese Copyright Act.142

Since the arrival of ChatGPT and other generative AI tools, legislators and policymakers have actively explored ways to tighten Article 30-4, due to concerns about the challenges these tools have posed to the copyright system. In May 2024, the Japan Copyright Office published the General Understanding on AI and Copyright in Japan, which provides guidance on the application of Article 30-4 in the AI context.143 The document suggests that businesses could be held liable if they knowingly collect training data from infringing sources.144

Notwithstanding this recent document, the Japanese exception for TDM and other uses that do not result in the enjoyment of "the thoughts or sentiments expressed in" the copyrighted work remains one of the broadest copyright exceptions available throughout the world for AI training.145 To the extent that scholars are interested in legal transplants or legal diffusion,146 it is intriguing that this exception has diffused, if not transplanted, abroad in the opposite direction of past intellectual property law transplants-that is, from Asia to the West, not the other way around.147

Singapore, which revised its Copyright Act in November 2021, provides another illustrative example. Section 244 of the new statute provides a copyright exception for making a copy of the copyrighted work for computational data analysis and for preparing that work for such analysis.148 Although this new exception is not limited to the non-commercial use of the copyrighted work and cannot be overridden by a contract, it contains some conditions not found in the Japanese exception.149 First, the copy made can only be used for the purposes specified in this provision.150 Second, the user cannot supply that copy to other people, except for the purpose of "verifying the results of the computational data analysis" or "collaborative research or study relating to the purpose of [such] analysis."151 Third, the user needs to have lawful access to the copyrighted work that is to be used.152 Finally, it is expected that the first copy used in the analysis is noninfringing. 153 The statute nonetheless introduces exceptions to cover situations where the user does not know or have reason to know the illicit nature of the copy used or where computational data analysis necessitates the use of an infringing first copy.154

To help with interpretation, Section 243 defines "computational data analysis" as "using a computer program to identify, extract and analyse information or data from the work or recording" or "using the work or recording as an example of a type of information or data to improve the functioning of a computer program in relation to that type of information or data."155 On the fact sheet released by the Intellectual Property Office of Singapore, possible uses listed are "sentiment analysis, text and data mining, and training machine learning."156 Section 243 further provides as an illustration "the use of images to train a computer program to recognise images,"157 while Section 244 states that "making a copy includes . . . storing or retaining the copy."158

Compared with Article 30-4 of the Japanese Copyright Act, Section 244 of the Singapore Copyright Act is more restrictive. Yet the latter provides clarity by offering carefully drafted definitions and helpful illustrations. The provision also contains detailed language to facilitate a more politically palatable compromise between copyright holders and their supportive industries on the one hand and AI developers and users of generative AI tools on the other. In addition, Sections 243 and 244 of the Singapore Copyright Act were adopted more than a decade after the introduction of the Japanese exception for TDM and three years after the expansion of that exception to cover uses that do not result in the enjoyment of "the thoughts or sentiments expressed in" the copyrighted work. Given the late adoption of Sections 243 and 244 and the additional conditions attached, some policymakers may find the Singapore exception more appealing.159 Thus, when all of the features have been taken into consideration, the new Singaporean copyright exception for computational data analysis may provide a more attractive model for legal transplant or emulation, especially among countries in the Asia-Pacific region that continue to worry about pushbacks from local copyright industries, foreign investors, and their powerful and supportive governments.160

While the copyright law revision in Singapore is important in the AI context, it is also interesting from a comparative standpoint. Like the U.S. Copyright Act, the Singapore Copyright Act includes a fair use provision. Codified originally in Section 35 (now Sections 190-194),161 the provision was adopted as a broad fair dealing exception in 2005 after the United States-Singapore Free Trade Agreement and its intellectual property chapter had entered into force.162 Before the latest revision, Singapore had what I have called "a fair dealing provision in name but a fair use provision in effect."163 After the revision, however, Singapore now has a fair use provision both in name and effect.

This new fair use provision enabled Singapore to become one of the few countries that has transplanted the four fair use factors from the United States virtually verbatim.164 To the extent that Singaporean and U.S. courts offer similar interpretations of those factors to support AI development-not a given considering the potential jurisdictional differences-the law in Singapore will likely be more supportive of AI development due to the existence of both the fair use provision and an explicit copyright exception for computational data analysis, with the latter covering issues that may fall outside the scope of fair use. The United States, by contrast, has only the fair use provision.165

The country that has been widely discussed in the AI debate, yet that has not provided a copyright exception for either TDM or computational data analysis, is China. In November 2020, amid the COVID-19 pandemic, China adopted the Third Amendment to the Copyright Law, which took effect on June 1, 2021. Included in this statute was Article 24, which provides the list of enumerated circumstances in which a copyrighted work may be used without authorization or remuneration.166 The Third Amendment added to that list Clause 13, which covers "[o]ther circumstances provided for by laws and administrative regulations."167 Because the language mentions laws and administrative regulations specifically, Article 24(13) does not convert the system of copyright limitations and exceptions in China from a closed system to an open-ended one.168 Nevertheless, the provision can still be very supportive of AI development once the appropriate regulations have been introduced. Thus far, China has not yet updated the Regulations for the Implementation of the Copyright Law.169 Should a new copyright exception for TDM or computational data analysis be introduced through those implementing regulations or a new set of AI-specific regulations, that exception could be read into Article 24(13) of the Copyright Law.

In July 2023, China adopted the Interim Measures for the Management of Generative Artificial Intelligence Services,170 pioneering legislation that aims to give China a similar early-mover advantage as the then-proposed EU AI Act.171 While the Interim Measures mention intellectual property protection, the language is vague and open to interpretation. Article 4 states that "[t]he provision and use of generative AI services shall . . . [r]espect intellectual property rights and commercial ethics [and] protect commercial secrets."172 Article 7 stipulates further that "[t]he providers of generative AI services . . . [must not infringe] the intellectual property rights that are lawfully enjoyed by others."173 How this language is to be interpreted will, of course, depend on the Chinese courts' determination of the legality of using copyrighted works to train AI models. Should such use be deemed legal, Articles 4 and 7 will not provide additional protection to copyright holders.

C. Summary

This Part has shown that the developments of copyright law in the AI space have slowly diverged, with the United States taking a rather different approach from those embraced by its trading partners in Asia. While the copyrightability of AI-generated work has become a key point of departure at the international level, disputes and lawsuits over the legality of using copyrighted works to train AI models have revealed the differing approaches taken across the world even though these approaches converge in their end goals. In view of these growing divergences, U.S. legislators and policymakers are now confronted with a key policy choice at the intersection of copyright and AI: Should the United States retain existing approaches, follow other jurisdictions, or work with these jurisdictions to develop harmonized AI-related international copyright standards?

III. CHANGING COPYRIGHT LAW DEVELOPMENTS

Part II has discussed two areas in which substantial legal and policy reform may emerge at the intersection of copyright and AI- namely, copyrightability and AI training. This Part turns to five additional areas: (1) the idea-expression dichotomy; (2) access and substantial similarity; (3) fair use; (4) secondary liability; and (5) formalities. Although it is unclear which of these seven areas will encounter legislative reform in the future, it is not far-fetched to assume that AI developments will have serious ramifications for all seven areas. This Part therefore explores how these areas will become important in the AI context and offers insights into how copyright law may change in the future.

A. Idea-Expression Dichotomy

The idea-expression dichotomy is a foundational principle in copyright law, whether in the United States or other parts of the world. As Matthew Sag observes, "[T]he idea-expression distinction is a central element of the balance between the interests of authors in preventing the exploitation of their writings and society's competing interest in the free flow of ideas, information, and commerce."174 Due to this important distinction, some commentators have asserted that using copyrighted works to train AI models does not need copyright holders' authorization because the use is non-expressive and copyright law does not extend beyond expressions.175 In a forthcoming article, Daryl Lim and I also argue that from a pro-competitive standpoint, the ideaexpression dichotomy may be even more important than fair use in the AI context.176

Outside the United States, legislatures and courts have taken a similar position. As early as 2009, Japan adopted a pioneering copyright exception that allows initially for TDM and later also for other uses that do not result in the enjoyment of "the thoughts or sentiments expressed in" a copyrighted work.177 Having been around for about a decade and a half, this provision predates the adoption of similar provisions in the United Kingdom, the European Union, and Singapore.178

Shortly after the arrival of ChatGPT and other generative AI tools, Mark Lemley wrote a highly provocative article discussing how these tools have "turn[ed] copyright upside down."179 As he explains, generative AI technologies have made the idea (the prompt) more important than the expression, which was autonomously generated by an AI system based on that idea. Such technologies have therefore called into question the traditional analysis in copyright law based on the idea-expression dichotomy. Section 102(a) of the 1976 Copyright Act stipulates that "[c]opyright protection subsists . . . in original works of authorship fixed in any tangible medium of expression."180 Section 102(b) states further that copyright protection does not "extend to any idea, procedure, process, system, method of operation, concept, principle, or discovery."181 In addition, the merger doctrine denies copyright protection to expressions "[w]here an idea and the expression 'merge,' or are 'inseparable.'"182

While Professor Lemley wrote his article shortly after the arrival of ChatGPT and other generative AI tools, future uses and later generative AI tools are likely to change the interplay between an idea and its resulting expression.183 Increasingly, a user needs to input a large number of prompts, parameters, and other information into an AI system to generate creative outputs. The selection and arrangement of all this information could arguably rise to the level of substantial "creative input or intervention from a human author,"184 even when the AI system has played an important role in generating outputs. As a result, courts in different jurisdictions may disagree over whether the human effort involved is sufficient to sustain copyright protection-in particular, whether that effort has caused the generation of the creative output at issue. For instance, the U.S. Copyright Review Board did not consider sufficient the more than six hundred inputs Jason Allen claimed to have provided to develop his artwork "Théâtre D'opéra Spatial."185 By contrast, the Beijing Internet Court, in Li v. Liu, was willing to recognize the copyright in an image generated from a small number of prompts.186 The latter court went even further to highlight the creative contributions of these prompts by listing them in the opinion-the first judicial decision to do so (as far as I am aware) and a rarity in any jurisdiction.187

As prompts, parameters, and other inputs into an AI system become more complex, it is only a matter of time before the combination of this information will resemble a computer program consisting of multiple lines of codes-subject matter that our copyright system has frequently protected.188 Should there be a greater causal link between the human creative input and AIgenerated output, one could certainly view these prompts, parameters, and other inputs as elements making up a new computer language.

Moreover, companies have begun to actively hire prompt engineers to deploy generative AI tools to develop texts, images, and audio and video clips. The more engineering the inputs involve, the more likely the AI-generated output will fall on the side of expression in the idea-expression dichotomy. The merger doctrine may still apply in some circumstances, but it is unlikely to do so in many others.

B. Access and Substantial Similarity

The two standard elements in a prima facie case of copyright infringement are access and substantial similarity. However, with the wide use of AI systems in the creative process, it remains unclear how courts will approach each element, including the weight given to that element.189 While independent creation has provided a longstanding defense in copyright law,190 making access an important element of a copyright infringement case, it is unclear what will constitute independent creation when individuals create with the help of AI systems. How courts determine this issue will likely vary from jurisdiction to jurisdiction.

Over the years, policymakers, commentators, and intergovernmental organizations have underscored the need to separate AI-generated works from AI-assisted works.191 The use of AI systems for assistance in the creative process is usually referred to as "intelligence augmentation," not autonomous creation.192 Yet regardless of whether one uses the AI system to generate a new work or merely as a tool or assistant, that the system has been trained with preexisting copyrighted works suggests that any user of that system will have access to the creative elements originating from the training data. Such access partly explains why the U.S. Copyright Office now requires the disclosure of AI-generated content in a copyright registration application.193 To provide more information about which copyrighted works have been used to train AI models, the newly adopted EU AI Act has gone even further to require AI developers to "draw up and make publicly available a sufficiently detailed summary about the content used for training of the general-purpose AI model, according to a template provided by the AI Office."194

Whether in the United States or across the Atlantic, the demand for greater disclosure also underscores the lack of transparency over what data-both copyrighted and otherwise-have been used to train AI models. As if this problem were not bad enough, it can be quite difficult to figure out what data have been used for training purposes in a complex supply chain that involves multiple actors from around the world.195 There is also a good chance that some data used for training purposes may have been unknowingly created with an AI system, due to the lack of proper labeling and disclosure of training data.

Like analyzing access, determining substantial similarity will become more complicated in the AI context. One may recall that in Ty, Inc. v. GMA Accessories, Inc., Judge Richard Posner noted the need to compare the defendant's stuffed toy with not only the plaintiff's bean-bag pig but also a real pig and fictional pigs in the public domain.196 Such comparison is needed because the defendant might not have copied the design from the plaintiff's, but instead from either a real or fictional pig. With comparison, the court could quickly eliminate these possibilities.

In the AI context, courts will have to compare not only the plaintiff's and defendant's works-texts, images, audio clips, or videos-but also works that users can generate from AI systems. When these systems are widely deployed in the creative process, it will be no surprise to find similar creative outputs-not because they copied each other, but because the creators in both instances used the same AI system, or two or more AI systems, that had been trained with the same underlying copyrighted works. While no copyright infringement will take place in a jurisdiction that denies copyright protection to AI-generated works, the analysis will become more complicated in jurisdictions offering such protection. Even in jurisdictions denying copyright protection to AI-generated works, complications can arise when the defendant has used an AI system in the creative process to help generate an early draftof the contested text, image, audio clip, or video before providing substantial human creative input or intervention that warrants copyright protection.

C. Fair Use

Fair use is often offered as defense in a case involving the unauthorized use of copyrighted works to train AI models, including in the ongoing copyright litigation against AI developers.197 Going beyond model training, however, the greater use of AI tools, including generative AI tools, will still create many new issues concerning fair use. To be sure, fair use is a unique feature of the U.S. copyright system,198 thereby raising questions about whether those issues will arise in other jurisdictions. Nevertheless, a number of jurisdictions have now transplanted the U.S. fair use provision or its close variants.199 Some jurisdictions that have fair dealing provisions-a closed system of narrowly defined copyright limitations and exceptions-have also incorporated the four Section 107 factors or their variants through statutory or case law.200 Thus, the new fair use issues arising in the AI context may affect more countries than one anticipates.

In the U.S. fair use analysis, courts have traditionally focused on both the first factor ("the purpose and character of the use") and the fourth factor ("the effect of the use upon the potential market for or value of the copyrighted work").201 As Barton Beebe observes, "[T]he first and fourth factors are shown each to exert an enormous amount of influence on the outcome of the test, with the fourth very much in the driver's seat."202 Particularly notable is the transformative use doctrine,203 which has played a dominant role in fair use analysis.204 The interpretation of this doctrine also raises interesting questions in view of the recent U.S. Supreme Court decisions in Google LLC v. Oracle America, Inc.205 and Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith.206

Compared with the first and fourth factors, the second factor ("the nature of the copyrighted work") often plays an insignificant role, if it is not overlooked at all. As the Second Circuit reminded us in Authors Guild, Inc. v. HathiTrust, the second factor "may be of limited usefulness where . . . the creative work . . . is being used for a transformative purpose."207 More than a year later, that same court declared in Authors Guild v. Google, Inc. that "[t]he second factor has rarely played a significant role in the determination of a fair use dispute."208 In the AI context, however, this factor will become more important for two reasons. First, AI technology involves computer software, the functional aspects of which have always figured more significantly in the second-factor analysis. A leading case in this area is Sega Enterprises Ltd. v. Accolade, Inc., a copyright infringement case involving the defendant's use of the disassembly process to reverse-engineer the computer code needed to facilitate compatibility with the plaintiff's game console.209 As the Ninth Circuit explained:

[T]he fact that computer programs are distributed for public use in object code form often precludes public access to the ideas and functional concepts contained in those programs, and thus confers on the copyright owner a de facto monopoly over those ideas and functional concepts. That result defeats the fundamental purpose of the Copyright Act-to encourage the production of original works by protecting the expressive elements of those works while leaving the ideas, facts, and functional concepts in the public domain for others to build on.210 The societal harm created by overprotection of the functional aspects of computer software is significant. As Marshall Leaffer observes, "If defendant's reverse engineering through disassembly or decompilation were illegal, the copyright owner would have a de facto monopoly over [those] aspects of the work."211

Second, in a jurisdiction offering copyright protection to AIgenerated works, the use of these works in the creative process will become important in the second-factor analysis.212 Because some courts may not consider the infringement on an AI-generated work as harmful as the infringement on a human-created work, this analysis may favor the fair user more when an AI system generated the plaintiff's work. As Clark Asay reasons, "[M]uch of the creativity found in the AI output may owe more to the AI's internal workings than human efforts, in which case the limited human creativity involved in generating the output means that the second factor pushes in favor of a finding of fair use."213

Like the second factor, the third factor ("the amount and substantiality of the portion used in relation to the copyrighted work as a whole") could raise new fair use issues in the area of AI training. Indeed, the analysis of this factor is likely to depend on the technical details of the training process involved, which may vary from one AI developer to another. At first glance, the thirdfactor analysis seems to favor copyright holders, due to the use of the entirety of their copyrighted works for training purposes.214 Nevertheless, if the training process has used only portions of each work, which have been converted to "tokens,"215 courts could find the third factor favoring AI developers.

D. Secondary Liability

Most litigation against AI developers, such as those involving the unauthorized use of copyrighted works to train AI models, has focused on direct infringement. Nevertheless, if those lawsuits prove difficult, copyright holders may start advancing liability claims based on theories of contributory and vicarious infringement. Those theories were used in litigation against filesharing services in the late 1990s and early to mid-2000s.216 It will therefore be logical for litigants to use these theories again in the AI context. After all, individuals using generative AI systems have provided the prompts, parameters, and other inputs into the systems to generate creative outputs.217

Should secondary liability be pursued, courts will inevitably entertain questions about both volitional conduct and proximate cause. As the late Justice Scalia reminded us in his strong dissent in ABC, Inc. v. Aereo, Inc.:

The distinction between direct and secondary liability would collapse if there were not a clear rule for determining whether the defendant committed the infringing act. The volitional-conduct requirement supplies that rule; its purpose is not to excuse defendants from accountability, but to channel the claims against them into the correct analytical track.218

In Perfect 10, Inc. v. Visa International Service Association, the Ninth Circuit also highlighted the importance of proximate cause.219 In an earlier case, Perfect 10, Inc. v. Amazon.com, Inc., that court found that "Google could be held contributorily liable if it had knowledge that infringing Perfect 10 images were available using its search engine, could take simple measures to prevent further damage to Perfect 10's copyrighted works, and failed to take such steps."220 Drawing on the previous case, the Ninth Circuit explained:

The salient distinction is that Google's search engine itself assists in the distribution of infringing content to Internet users, while Defendants' payment systems do not. The Amazon.com court noted that "Google substantially assists websites to distribute their infringing copies to a worldwide market and assists a worldwide audience of users to access infringing materials." . . . Defendants [in this case] do not provide such a service. They in no way assist or enable Internet users to locate infringing material, and they do not distribute it. They do, as alleged, make infringement more profitable, and people are generally more inclined to engage in an activity when it is financially profitable. However, there is an additional step in the causal chain . . . .221

There may also be questions concerning whether the technology would fit within the safe harbor provided by Sony Corp. of America v. Universal City Studios, Inc.222 Even though the use of generative AI tools may result in the creation of infringing outputs, it is also "capable of substantial noninfringing uses."223 To the extent that AI developers have not induced users to commit copyright infringement,224 AI technology should receive some substantial protection under the Sony safe harbor.

In addition, it would be interesting to see if Congress will step in to limit the liability of AI developers, especially after it has realized the high value of AI technology from both an economic and a strategic standpoint. Indeed, there is a strong precedent. More than two decades ago, Congress introduced Section 512 of the Copyright Act to limit the liability of online service providers.225 Since its adoption, courts, policymakers, commentators, the copyright industries, and consumer advocates have actively explored the appropriate contours and effectiveness of this safe harbor, including the type of knowledge needed to trigger protection,226 whether fair use should be considered when filing triggering notices,227 how the safe harbor addresses repeat infringers,228 and whether the mechanism needs to shiftfrom "notice and take down" to "notice and stay down"-as is practiced in the European Union following the adoption of the DSM Directive.229

In May 2020, amid the COVID-19 pandemic, the U.S. Copyright Office released a detailed study of Section 512, taking stock of many issues and continuous challenges confronting copyright holders in the digital environment.230 Thus, if a safe harbor for AI developers is to be introduced, one logically wonders what that safe harbor would look like, whether we can draw any lessons from our past efforts of implementing Section 512, and if we will modify the design of this safe harbor231-including whether we could incorporate new developments from abroad, such as the filtering requirement under Article 17 of the DSM Directive.232

Finally, it is worth keeping in mind that secondary liability, like fair use, is a rather distinct feature of U.S. copyright law. Although similar liability exists in other jurisdictions, what works in the United States does not always have direct equivalents.233 Should the United States turn to such liability to address concerns about the unauthorized use of copyrighted works to train AI models that have the capacity to produce competing creative outputs, legislatures and courts in other jurisdictions may have to customize or alter the protection against secondary copyright liability when introducing such protection abroad.234

E. Formalities

Copyright formalities used to separate the United States from the rest of the world. Examples are the affixation of copyright notices,235 the registration of copyright,236 and the requirement to deposit two copies of the copyrighted work into the Library of Congress.237 Before the 1976 Copyright Act entered into effect, the failure to comply with formality requirements, whether inadvertent or not, could result in the forfeiture of copyright.238 Such forfeiture was acceptable in the United States because "[i]t has long been established [in U.S. law] . . . that copyright is not a natural right, but one created by positive law."239

Outside the United States, however, conditioning copyright protection on compliance with formality requirements has always been problematic. In fact, the Berne Convention for the Protection of Literary and Artistic Works (Berne Convention),240 the predominant international copyright treaty, explicitly prohibits such requirements. Article 5(2) of the Convention provides: "The enjoyment and the exercise of these rights shall not be subject to any formality."241 Because of its past formality requirements in copyright law, the United States was unable to join the Berne Convention until March 1989, shortly after it abolished the application of these requirements to non-U.S. copyrighted works.242

Beginning in the early 2000s, commentators, most notably Christopher Sprigman, have begun calling for a reinstatement of formality requirements. Other commentators have also seen benefits for stronger protection of rights management information. Together, they observe that formalities can provide at least two benefits. First, by setting a high bar for copyright protection, formalities help channel works into the public domain.243 Such channeling is particularly important following Eldred v. Ashcroft, which found constitutional the extension of the copyright term and delayed the entry of copyrighted works into the public domain.244 Second, formality requirements help generate the information needed by users and future authors to locate copyright holders, such as their identity and the date of publication. These requirements therefore could be important in the future when computers readily use rights management information to allocate royalties or direct permission requests.245

The need to facilitate such arrangements was part of the reason WIPO members were eager to adopt provisions on rights management information in the WIPO Copyright Treaty246 (WCT) and the WIPO Performances and Phonograms Treaty.247 Article 12(1) of the WCT requires contracting Parties to provide adequate and effective legal remedies against the knowing removal or alteration of electronic rights management information without authorization and against the dissemination of copies of copyrighted works "knowing that [such] information has been removed or altered without authority."248 Article 12(2) defines "rights management information" to include "information which identifies the work, the author of the work, the owner of any right in the work, or information about the terms and conditions of use of the work, and any numbers or codes that represent such information."249 In October 1998, the Digital Millennium Copyright Act added Section 1202 to the 1976 Copyright Act to implement this treaty provision.250

In the AI context, copyright formalities can once again become important for three reasons. First, as noted earlier, they will help allocate royalties or direct permission requests. Thus far, systems that can provide such allocation and direction have not yet been widely deployed. Nevertheless, some online platforms have already provided such allocation. A case in point is the Content ID system deployed by YouTube.251 If the system matches files uploaded by internet users with the reference files provided by copyright holders, it will give the rightsholders the choice to "block a video from being viewed," "monetize the video by running ads against it," and "track the video's viewership statistics."252 When copyright holders choose to monetize the video, the payout will be arranged automatically within the platform.

Second, the introduction of formalities can help the public distinguish between AI-generated works from those created by human authors. To the extent that countries offer different levels of protection to these two types of works-whether provided expressly in the statute253 or through case law interpretations254- having an ability to make a proper distinction between these two types of work will be quite important. Outside the intellectual property context, some jurisdictions have already required the affixation of watermarks to separate human-created and AIgenerated works.255

Third, copyright formalities will help address the problem created by the unauthorized use of copyrighted works to train AI models. There has been ongoing discussion concerning whether copyright holders should have the right to opt out from the use of their works for training these models. In the European Union, for instance, Article 53(1)(c) requires "[p]roviders of general-purpose AI" to "put in place a policy to comply with Union law on copyright and related rights, and in particular to identify and comply with, including through state-of-the-art technologies, a reservation of rights expressed pursuant to Article 4(3) of" the DSM Directive.256 Article 4(3) of the DSM Directive states further that the TDM exception "shall apply on condition that the use of [copyrighted] works and other subject matter . . . has not been expressly reserved by their rightholders in an appropriate manner."257 Should such an opt-out mechanism be developed, formalities (or the inclusion of meta data) will have to be deployed to ensure that those works whose authors have opted out will not be used to train AI models. It therefore remains to be seen whether the arrival of generative AI technology will revive the use of copyright formalities.

F. Summary

Due to the length and scope of this Article, Parts II and III only manage to cover seven areas in which substantial legal and policy reform may emerge at the intersection of copyright and AI. The discussion in these two parts is by no means exhaustive. Instead, it aims to use these seven areas to illustrate how AI technologies may affect different areas of copyright law in the future. This Article also recognizes the fact that jurisdictions around the world may come to very different, if not opposite, conclusions on how copyright law should be reformed in view of the opportunities and challenges provided by AI technologies.

IV. ALIGNMENT OF THE FUTURE PATH

When countries deploy different approaches to address legal and policy issues, including those at the intersection of copyright and AI, legislators and policymakers tend to assume that one jurisdiction has a better approach than the others-or, worse, that their home country has the right or wrong approach while other jurisdictions have the opposite. Oftentimes, legislators and policymakers settle on a certain approach simply because they have insufficient empirical evidence to fully evaluate the different legal and policy options, including those introduced abroad. Some legislators and policymakers are also constrained by factors outside the intellectual property arena, including constitutional requirements, legal tradition, economic system, and national ambition.

To appreciate the possibility of having different approaches that are equally justifiable from both the legal and policy standpoints, section A illustrates this possibility using a hypothetical future scenario in which an AI system that is capable of autonomously generating creative works has been used by three separate owners. This illustration underscores the need for legislators and policymakers to embrace holistic, macro-level, and comparative thinking to facilitate the development of appropriate copyright law in the AI space. Section B then turns to different options available to the United States when working with other members of the international community to chart a future path of copyright law. Although standards adopted through this potentially new path do not have immediate legal effects in a nonself- executing jurisdiction like the United States,258 they will often pave the way for new laws and policies at the domestic level. Cases in point are the impact of the TRIPS Agreement and the WCT259 on U.S. copyright law developments.260 This section further identifies four options for AI-related copyright law reform. While the first three options require international engagement-whether at the multilateral, plurilateral, or bilateral level-the last option shows the possibility of a country harmonizing its laws with those of other jurisdictions without such engagement. This Part takes note of the different approaches taken by past U.S. legislatures and administrations. It further reminds us that the United States could choose to partially harmonize its copyright standards with those of other jurisdictions or not alter these standards at all.

A. An Illustration

To illustrate the different laws and policies available to address AI-related copyright law developments, consider a hypothetical future scenario in which an AI system that is capable of autonomously generating creative works has been used by three separate owners (see fig. 1). In this scenario, the owner for the first five years has designed the AI system and its self-learning algorithms.261 The system is then sold to the second owner, who for the next ten years has selected and inputted data to train the system to generate new creative works. This system is finally sold to the current owner, who continues to select and input new training data but has owned the system for only less than a year. This system benefits from the training conducted by the two prior owners. Assuming we are in a jurisdiction that extends copyright protection to AI-generated works, who should own the rights in the creative outputs autonomously generated by this system: the first, second, or third owner?

Those who answer the first owner recognize the importance of algorithms. The approach is similar to the one found in jurisdictions that have offered copyright protection to computer-generated works. A case in point is the United Kingdom. Section 9(3) of the Copyright, Designs and Patents Act 1988 provides: "In the case of a literary, dramatic, musical or artistic work which is computergenerated, the author shall be taken to be the person by whom the arrangements necessary for the creation of the work are undertaken."262 Because the first owner designed the AI system and its self-learning algorithms, one could certainly argue that this owner is "the person by whom the arrangements necessary for the creation of the work are undertaken."263

Those who answer the second owner are conscious of the high value and impact of training data. Placing a great emphasis on these data makes good sense in a hypothetical scenario involving selflearning algorithms.264 In AI literature, commentators have already widely discussed the insufficiency of scrutinizing AI algorithms to detect machine-based biases and errors.265 As Kartik Hosanagar and Vivian Jair remind us: "[M]achine learning algorithms-and deep learning algorithms in particular-are usually built on just a few hundred lines of code. The algorithms logic is mostly learned from training data and is rarely reflected in its source code."266

Those who answer the third owner analogize the AI system to other computer tools we currently have. When we use MicrosoftWord to create documents, Microsoftdoes not own the copyright in those documents. By analogy, when we use AI systems as tools, AI developers should not own the copyright in the creative outputs. Indeed, these developers may have strong incentives to disclaim ownership in the outputs.267 After all, the more value one could derive from the AI system, the higher the price that individual will be willing to pay. In a scenario where AI systems can be sold to many users (similar to mass-market software) or where those systems have been frequently resold, having a rule that maximizes the market prices of these systems looks very attractive to owners.

In addition to the three choices mentioned above, one may suggest hybrid choices that include more than one owner. In these choices, each owner will obtain some rights in the creative outputs-or at least some portions of the revenue generated from the sales of those outputs. The difficult issue here is not about ownership, but the proper allocation of fractional ownership or sales revenue. In the intellectual property arena, an arrangement that supports fractional ownership already exists in the form of the droit de suite, or resale royalties.268 Found often in continental Europe and other jurisdictions offering strong moral rights, resale royalties provide artists with compensation when their artworks are sold after the initial purchase. Such royalties enable artists who have sold artworks at low prices, often early in their career, to obtain additional compensation that is commensurate with the increased value in their artworks.

In recent years, commentators have also explored the use of blockchain technology to support a droit de suite regime.269 As I noted in a recent book chapter: "By allowing creators to designate in advance the payment of certain portions of future sales back to them, and by hashing this payout arrangement into the chain, blockchain technology successfully institutionalizes resale royalties."270 Indeed, some creators of non-fungible tokens have already received portions of future sales, regardless of whether they reside in a jurisdiction recognizing the droit de suite.271

From both the legal and policy standpoints, all of these answers are justifiable. Without empirical analysis, it is hard to tell whether one answer is superior to the other. Even if we could do so, the answers may change in the future, as AI technology continues to evolve and as more people use AI on a day-to-day basis. Indeed, the stronger self-learning capabilities the AI system has, the less important the algorithms originally installed on that system will be. In addition, whether the current owner should get strong legal rights will depend on whether the AI system is being sold the same way as computer equipment or mass-market software. The wider is the market, the more eager AI developers are to maximize the market value of their systems.

Interestingly, the analysis in this section resembles the question the U.S. Patent and Trademark Office posed the public in its earlier consultation document:

Assuming involvement by a natural person is or should be required, what kind of involvement would or should be sufficient so that the work qualifies for copyright protection? For example, should it be sufficient if a person (i) designed the AI algorithm or process that created the work; (ii) contributed to the design of the algorithm or process; (iii) chose data used by the algorithm for training or otherwise; (iv) caused the AI algorithm or process to be used to yield the work; or (v) engaged in some specific combination of the foregoing activities? Are there other contributions a person could make in a potentially copyrightable AI-generated work in order to be considered an "author"?272

Although this paragraph focuses on the creative contributions of an individual producing an AI-generated work, it shows that U.S. legislators and policymakers are well aware of the different choices available to them for charting the future path of copyright law in the AI space. Thus, the constitutional and incentive-based arguments militating against the extension of copyright protection to AI-generated works do not end the policy inquiry. Instead, legislators and policymakers will need to undertake more holistic, macro-level, and comparative analysis as well as the empirical research needed to back up this analysis.

B. Possible Options

Using a hypothetical future scenario, the previous section shows that many policy choices can be justified from the legal and policy standpoints. This section turns to how the United States could work with other members of the international community to establish new AI-related copyright norms to support its preferred policy choices.

The options available to U.S. policymakers will likely fall in a continuum. On one end is multilateral engagement, and on the other is do-nothing. Between these two poles are plurilateral effort, bilateral initiative, and unilateral action. This section identifies four options that the United States can take to shape the future path of copyright law in the AI space: (1) the negotiation of a new international treaty; (2) the development of softlaw instruments; (3) the creation of a global multi-stakeholder dialogue; and (4) the utilization of choice-of-law principles by domestic courts to adjudicate multijurisdictional and cross-border disputes.273 The first three options can be developed at the multilateral, plurilateral, regional, or bilateral level.

1. International Treaty Negotiations

When confronted with different international norms, policymakers, commentators, and business leaders instinctively call for developing a new international treaty. Although the treaty's binding nature has a strong intuitive appeal, developing such a treaty can be quite difficult and time-consuming. From proposals to draftnegotiation texts to adoption to final ratification, the treaty process tends to be rather lengthy.274

The mid-1990s have seen a barrage of international intellectual property norm-setting activities-including the development of the TRIPS Agreement at the WTO275 and the WCT276 and the WIPO Performances and Phonograms Treaty277 at WIPO. Yet the international community failed to adopt another international intellectual property treaty until more than a decade later. Adopted at WIPO in 2012, 2013, and 2015, respectively, were the Beijing Treaty on Audiovisual Performances,278 the Marrakesh Treaty to Facilitate Access to Published Works for Persons Who Are Blind, Visually Impaired or Otherwise Print Disabled,279 and the Geneva Act of the Lisbon Agreement for the Protection of Appellations of Origin and Their International Registration.280

Most recently, after another decade-long hiatus, WIPO members adopted the WIPO Treaty on Intellectual Property, Genetic Resources and Associated Traditional Knowledge (GRATK Treaty) in May 2024.281 Six months later, WIPO members adopted the Riyadh Design Law Treaty.282 The negotiation of the GRATK Treaty was particularly difficult, and the final instrument represented the cumulative result of more than two decades of explorations and deliberations at the Intergovernmental Committee on Intellectual Property and Genetic Resources, Traditional Knowledge and Folklore at WIPO.283

Moreover, any treaty that is to be developed tends to lag behind technological developments, leading to an inevitable cat-and- mouse chase.284 A case in point is the TRIPS Agreement, which was outdated upon arrival in relation to internet-related developments. As Marci Hamilton observes, "TRIPS reached fruition at the same time that the on-line era became irrevocable. Yet it makes no concession, not even a nod, to the fact that a significant portion of the international intellectual property market will soon be conducted on-line."285 Likewise, Jerome Reichman declares:

[The principal weakness of the TRIPS Agreement] stems from the drafters' technical inability and political reluctance to address the problems facing innovators and investors at work on important new technologies in an Age of Information. The drafters' decision to stuffthese new technologies into the overworked and increasingly obsolete patent and copyright paradigms simply ignores the systemic contradictions and economic disutilities this same approach was already generating in the domestic intellectual property systems.286

Notwithstanding these criticisms, some international treaty provisions have been quite impactful. Even though the TRIPS Agreement was outdated in relation to internet-related developments, it led to extensive intellectual property law reforms in many WTO members, especially those in the developing world.287 The WCT also reshaped the digital copyright regime in many WIPO members, calling for new legislation to support digital rights management tools while extending copyright protection to situations involving the making available of copyrighted works on the internet and in the digital environment.288

2. SoftLaw Instruments

Although the development of international treaties immediately comes to mind when international organizations are discussed, these organizations have also developed softlaw instruments that are arguably more politically palatable. Within WIPO, member states have developed or considered a few softlaw recommendations,289 model laws,290 and other guiding documents.291 The most widely cited example of a softlaw recommendation is the Joint Recommendation Concerning Provisions on the Protection of Well-Known Marks, which provides a set of guidelines for the protection of well-known marks across the world.292 This document recognized, for the first time, WIPO's need to "adapt to the pace of change in the field of industrial property by considering new options for accelerating the development of international harmonized common principles."293 Since its adoption in September 1999, this nonbinding recommendation has been incorporated by reference into the intellectual property chapters of several bilateral, regional, and plurilateral trade agreements.294

At the WTO, members have also taken a similar route. In the early 2000s, these members were actively seeking normative solutions to address the HIV/AIDS, malaria, and tuberculosis epidemics in Sub-Saharan Africa and other parts of the world.295 Unable to reach consensus on the needed changes to the patent provisions in the TRIPS Agreement, WTO members settled on adopting a nonbinding softlaw instrument named the Declaration on the TRIPS Agreement and Public Health.296 Paragraph 6 of this Declaration explicitly "recognize[d] that WTO Members with insufficient or no manufacturing capacities in the pharmaceutical sector could face difficulties in making effective use of compulsory licensing under the TRIPS Agreement."297 To implement this provision, WTO members subsequently engaged in a debate on what follow-up action they should take: another declaration (or an authoritative interpretation), an amendment, a dispute settlement moratorium, or a waiver.298 These members eventually settled on adding a waiver of Article 31(f) to the TRIPS Agreement as the new Article 31bis.299 This amendment was ratified in January 2017 and has since entered into force.300

In the WIPO and WTO examples, the softlaw instruments have been subsequently hardened through the incorporation of international treaty provisions either directly or by reference. These instruments, to some extent, can be viewed as steppingstones for international norm-setting.301 Moreover, AI is one area that can benefit from the creation of new international norms but that has not yet achieved international consensus.302 In view of a lack of such consensus, countries may be more willing to establish a softlaw recommendation or other nonbinding international legal instruments than an international treaty.303

3. A Global Multistakeholder Dialogue

The most widely cited example of a global multistakeholder dialogue is the Internet Governance Forum, which was launched in 2006 shortly after the second phase of the World Summit on the Information Society in Tunis.304 Established as a "multilateral, multi-stakeholder, democratic and transparent" policy dialogue,305 this forum aims to address the concerns raised by the growing digital divide in developing countries and to explore the many unprecedented opportunities generated by the information revolution.306 Thus far, meetings have been convened around the world-from Athens to Addis Ababa and from Guadalajara to Geneva.

Although commentators continue to question this forum's legitimacy, effectiveness, and utility,307 multistakeholder dialogues can provide several benefits. Depending on their ambition, structure, process, and coverage, these dialogues can help kick-start the discussion needed to achieve target goals-whether it is to establish new norms, new institutions, or both. By bringing together a wide variety of stakeholders, facilitating interactions, and building trust, multistakeholderism can help highlight the shared values, interests, and preferences of the different stakeholders while at the same time consolidating their hugely diverse positions. Down the road, multistakeholder dialogues can also help develop the much-needed codes of best practices, softlaw recommendations, or international legal instruments (the first and second option).

It is indeed no surprise that former WIPO Director General Francis Gurry embraced the development of a global multistakeholder dialogue as a potential means to achieve the "seamless global digital marketplace" he proposed in his welcoming address in the 2013 WIPO General Assembly.308 As he explained in a follow-up interview with the Intellectual Property Watch a few months later:

I do not think this is a legislative exercise. This is something that involves a little bit of legislation, for example, the Bruce Willis problem, which is that he has 50,000 songs that he has bought on iTunes, can he give them to his children? If it were 50,000 CDs, he could. So there are some legislative tweaks. But it is mainly about better business models, which is for the private sector to do. It is about improving the culture and understanding, it is about infrastructure, and data standards. That marketplace is a marketplace of data. Metadata constitute creative work and metadata have to talk to each other, so I would like to see us working on developing in a multi-stakeholder dialogue a loose roadmap of things that need to be done to achieve the efficient seamless legal global digital marketplace.309

4. Choice-of-Law Principles

The final option requires the development of choice-of-law principles for use in domestic courts. Although the first three options require international engagement-whether at the multilateral, plurilateral, or bilateral level-this final option can be implemented even without such engagement. This option will be particularly attractive should the U.S. administration reembrace unilateral action or an isolationist approach.

To address multijurisdictional copyright disputes, courts are accustomed to applying choice-of-law principles. The seminal U.S. copyright case that has utilized such principles is Itar-Tass Russian News Agency v Russian Kurier, Inc.,310 in which several Russian journalists sued a New York-based Russian newspaper for allegedly infringing the copyright in their newspaper and magazine articles which were originally published in Russia. After examining the two different interpretations of national treatment, the Second Circuit held that the national treatment provision of the Berne Convention does not govern the choice-of-law question.311 Instead, the appellate court "fill[ed] the interstices of the Act by developing federal common law on the conflicts issue."312

Regarding copyright ownership, the court applied the law of the state that had the most significant relationship to the copyrighted work and to the parties involved-in other words, Russian law.313 With respect to infringement, however, the Second Circuit followed the oft-used choice-of-laws principle of lex loci delicti (law of the place of the wrong).314 Under this principle, the applicable law is the law of the country in which the infringement occurred-namely, the United States.

In the 1990s and early 2000s, the Hague Conference on Private International Law sought to drafta new Convention on Jurisdiction and Foreign Judgments in Civil and Commercial Matters. Although the drafting exercise ultimately failed, due in no small part to the emerging challenges posed by the internet,315 the draftConvention paved the way for the development of two related projects: the American Law Institute Project on Intellectual Property: Principles Governing Jurisdiction, Choice of Law, and Judgments in Transnational Disputes316 and the Max Planck Group on Conflict of Laws in Intellectual Property (CLIP) Principles.317 Although both projects started with a focus on jurisdiction and enforcement of judgments, they go beyond the original foci to cover choice-of-law issues. In December 2020, the International Law Association also adopted the Guidelines on Intellectual Property and Private International Law.318 Named after the Association's biennial conference in Kyoto, Japan, the Kyoto Principles were designed to "apply to civil and commercial matters involving intellectual property rights that are connected to more than one State."319

In addition to these three projects-as well as other similar projects in Japan and South Korea320-commentators have suggested new choice-of-law approaches that may be useful in international copyright contexts. While Paul Berman advocates "[a] cosmopolitan approach to international adjudication [that] allows courts to engage in a dialogue with each other concerning the appropriate definition of community affiliation and the appropriate scope of prescriptive jurisdiction,"321 Graeme Dinwoodie calls on courts to "decide international copyright cases not by choosing an applicable law, but by devising an applicable solution."322 As Professor Dinwoodie reasons:

International copyright disputes implicate interests beyond those at stake in purely domestic copyright cases. National courts should thus be free to decide an issue in an international case using different substantive copyright rules that reflect not only a single national law, but rather the values of all interested systems (national and international) that may have a prescriptive claim on the outcome. This approach to choice of law may unleash the generative power of common law adjudication as a means of developing international copyright norms. And it would accommodate the concerns of dynamic flexibility without compromising the values of national diversity or pluralistic perspective in a way that public law-based copyright lawmaking does not.323

In sum, many efforts have already been undertaken to develop choice-of-law principles for addressing multijurisdictional copyright disputes. To the extent that the United States-or, for that matter, any other jurisdiction-wants to take account of the AI-related copyright law developments from around the world without undertaking international engagement, these principles will enable U.S. courts to bring the alternative approaches practiced abroad into domestic copyright law. These principles are particularly important in the AI space where copyright law developments have begun to diverge324 and where the United States could benefit from the insights provided by AI-related copyright legislation and litigation in other jurisdictions.

5. Summary

This section identified four options that the United States can take to shape the future path of copyright law in the AI space. While these options show how the United States could engage with other countries to establish new AI-related copyright norms or develop these norms domestically through choice-of-law principles, it remains premature to conclude that the country will either transplant copyright standards from abroad or work with other jurisdictions to develop harmonized AI-related international copyright standards.

From interest group politics to legislative inertia to the legitimate interests of countervailing constituencies, one can think of many reasons why Congress will hesitate to take the harmonization route.325 Indeed, such hesitation helps explain the growing divergence of copyright law developments in the AI space across the world.326 One can also think of other reasons why Congress may choose partial harmonization instead, similar to how countries seeking to transplant the U.S. fair use provision have retained part of the status quo while adding new fair use elements that would help the system evolve without undergoing a dramatic paradigm shift.327

To a large extent, the harmonization process can be seen as a spectrum that goes from full harmonization (or verbatim legal transplantation) on one end to no harmonization (or diversification) on the other, with different levels of harmonization in between.328 Even more complicated, harmonization can go upward, downward, or even sideways.329 Just because a country has chosen to harmonize its intellectual property laws in the AI space does not mean that it will necessarily increase or decrease the level of legal protection. From a policy standpoint, how laws are to be harmonized is often as important as whether they will be.

In sum, the key objective of this Part is not to identify an ideal future path of copyright law for the United States, but to call on U.S. legislators and policymakers to start embracing holistic, macrolevel, and comparative thinking to facilitate the development of appropriate copyright law in the AI space. Such thinking will enable these legislators and policymakers to locate an appropriate path of copyright law that will strengthen the United States' economic capabilities, technological developments, and global competitiveness while making its copyright system ready for future technological demands and challenges.

CONCLUSION

The arrival of ChatGPT, Dall-E, Midjourney, Stable Diffusion, and other generative AI tools have sparked many new copyright law questions. Although policymakers and commentators are eager to find out whether AI-related copyright law developments across the world are converging or diverging,330 it is too early to tell whether those developments have brought the world closer together or caused them to driftfurther apart.331 Our difficulty in providing an assessment is understandable. Not only is the development of generative AI nascent, but members of the international community have the ability to engage with each other to chart a new path of copyright law in the AI space-for instance, by harmonizing extant norms, developing new norms, or creating alternative institutions and mechanisms.

To the extent that the United States is willing to undertake greater international engagement in the AI space, there will be many options to shape the future path of copyright law. Nevertheless, if the country chooses not to engage with other members of the international community, U.S. courts could still rely on choice-of-law principles to bring the alternative legal approaches practiced abroad into domestic copyright law. The existence of these options does not mean that the United States will change its copyright law in the AI space in the near future. It does reveal the different pathways for AI-related copyright law reforms.

In the next few decades, AI is likely to have a substantial and longstanding impact on the United States' economic and technological developments, global competitiveness, and readiness for future technological demands and challenges. Policymakers should therefore seize the opportunities provided by AI, including generative AI technology, while proactively adjusting the copyright system to alleviate challenges posed by this new technology. After all, the path of copyright law chosen by the United States will have a major impact on the country's economic and technological future. That path deserves our careful, urgent, and utmost attention.

Footnote

1. See, e.g., ChatGPT-Release Notes, https://help.openai.com/en/articles/6825453- chatgpt-release-notes (last visited Oct. 6, 2024) (documenting the releases of OpenAI's products).

2. In chronological order, these public hearings are as follows: Hearing on Oversight of A.I.: Rules for Artificial Intelligence Before the Subcomm. on Privacy, Tech. & the L. of the U.S. Senate Comm. on the Judiciary, 118th Cong. (2023); Hearing on Artificial Intelligence and Intellectual Property-Part I: Patents, Innovation, and Competition Before the Subcomm. on Intell. Prop. of the U.S. Senate Comm. on the Judiciary, 118th Cong. (2023); Hearing on Artificial Intelligence and Human Rights Before the Subcomm. on Hum. Rts. & the L. of the U.S. Senate Comm. on the Judiciary, 118th Cong. (2023); Hearing on Artificial Intelligence and Intellectual Property- Part II: Copyright Before the Subcomm. on Intell. Prop. of the U.S. Senate Comm. on the Judiciary, 118th Cong. (2023); Hearing on Oversight of A.I.: Principles for Regulation Before the Subcomm. on Privacy, Tech. & the L. of the U.S. Senate Comm. on the Judiciary, 118th Cong. (2023); Hearing on Oversight of A.I.: Legislating on Artificial Intelligence Before the Subcomm. on Privacy, Tech. & the L. of the U.S. Senate Comm. on the Judiciary, 118th Cong. (2023); Hearing on Oversight of A.I.: The Future of Journalism Before the Subcomm. on Privacy, Tech. & the L. of the U.S. Senate Comm. on the Judiciary, 118th Cong. (2024); Hearing on AI in Criminal Investigations and Prosecutions Before the Subcomm. on Crim. Just. & Counterterrorism of the U.S. Senate Comm. on the Judiciary, 118th Cong. (2024); Hearing on Oversight of AI: Election Deepfakes Before the Subcomm. on Privacy, Tech. & the L. of the U.S. Senate Comm. on the Judiciary, 118th Cong. (2024); Hearing on Oversight of AI: Insiders' Perspectives Before the Subcomm. on Privacy, Tech. & the L. of the U.S. Senate Comm. on the Judiciary, 118th Cong. (2024).

3. See, e.g., U.S. PAT. & TRADEMARK OFF., PUBLIC VIEWS ON ARTIFICIAL INTELLIGENCE AND INTELLECTUAL PROPERTY POLICY (2020) (collecting public views at the intersection of intellectual property and AI).

4. See Copyright and Artificial Intelligence, U.S. COPYRIGHT OFF., https://www. copyright.gov/ai (last visited Oct. 6, 2024) (listing the public listening sessions relating to literary works, visual arts, audiovisual works, and music and sound recordings); Latest AI News and Reports, U.S. PAT. & TRADEMARK OFF., https://www.uspto.gov/initiatives/ artificial-intelligence/artificial-intelligence-reports (last visited Oct. 6, 2024) (collecting Federal Register notices on its request for comments and public listening sessions).

5. See Regulation 2024/1689, of the European Parliament and of the Council Laying Down Harmonised Rules on Artificial Intelligence and Amending Regulations (EC) No 300/2008, (EU) No 167/2013, (EU) No 168/2013, (EU) 2018/858, (EU) 2018/1139 and (EU) 2019/2144 and Directives 2014/90/EU, (EU) 2016/797 and (EU) 2020/1828, 2024 O.J. (L 144) 1 [hereinafter EU AI Act].

6. Shengcheng Shi Rengong Zhineng Fuwu Guanli Zhanxing Banfa (...) [Interim Measures for the Management of Generative Artificial Intelligence Services] (promulgated by the Cyberspace Admin. of China, July 10, 2024, effective Aug. 15, 2024), https://www.chinalawtranslate.com/en/generative-ai-interim [hereinafter Interim Measures].

7. See Artificial Intelligence and Intellectual Property, WORLD INTELL. PROP. ORG., https://www.wipo.int/about-ip/en/frontier_technologies/ai_and_ip.html (last visited Oct. 7, 2024) (discussing this effort).

8. See discussion infra notes 28-29.

9. See Thaler v. Perlmutter, 687 F. Supp. 3d 140 (D.D.C. 2023) (upholding the Copyright Office's decision not to reconsider its refusal to register the artwork "A Recent Entrance to Paradise"); Letter from the Copyright Rev. Bd. to Ryan Abbott, Esq., Brown, Neri, Smith & Khan, LLP (Feb. 14, 2022), https://www.copyright.gov/rulings-filings/ review-board/docs/a-recent-entrance-to-paradise.pdf (providing the Copyright Review Board's decision).

10. See Letter from the U.S. Copyright Off. to Van Lindberg, Taylor English Duma LLP (Feb. 21, 2023), https://www.copyright.gov/docs/zarya-of-the-dawn.pdf (providing the Copyright Review Board's decision).

11. See Letter from the Copyright Rev. Bd. to Tamara Pester, Esq., Tamara S. Pester, LLC (Sept. 5, 2023), https://www.copyright.gov/rulings-filings/review-board/docs/ Theatre-Dopera-Spatial.pdf [hereinafter Copyright Review Board's Decision on "Théâtre D'opéra Spatial"] (providing the Copyright Review Board's decision).

12. See Letter from the Copyright Rev. Bd. to Alex P. Garens, Esq., Day Pitney, LLP (Dec. 11, 2023), https://www.copyright.gov/rulings-filings/review-board/docs/SURYAST.pdf [hereinafter Copyright Review Board's Decision on "Suryast"] (providing the Copyright Review Board's decision).

13. U.S. CONST. art. I, § 8, cl. 8.

14. 17 U.S.C. §§ 101, 201, 203, 304; see also Naruto v. Slater, 888 F.3d 418, 426 (9th Cir. 2018) ("The terms 'children,' 'grandchildren,' 'legitimate,' 'widow,' and 'widower' all imply humanity and necessarily exclude animals that do not marry and do not have heirs entitled to property by law.").

15. See discussion infra Section I.B.

16. Pamela Samuelson, Allocating Ownership Rights in Computer-Generated Works, 47 U. PITT. L. REV. 1185, 1199 (1986) (footnote omitted).

17. See, e.g., Reto M. Hilty, Jörg Hoffmann & Stefan Scheuerer, Intellectual Property Justification for Artificial Intelligence, in ARTIFICIAL INTELLIGENCE AND INTELLECTUAL PROPERTY 50, 62-64 (Lee Jyh-An, Reto M. Hilty & Liu Kung-Chung eds., 2021) [hereinafter AI AND IP] (discussing why the incentive theory is unconvincing and loses relevancy in the AI context); Benjamin Mitra-Kahn, Economic Reasons to Recognise AI Inventors, in RESEARCH HANDBOOK ON INTELLECTUAL PROPERTY AND ARTIFICIAL INTELLIGENCE 376, 384-87 (Ryan Abbott ed., 2022) [hereinafter RESEARCH HANDBOOK] (explaining why the incentive argument for AI inventors falls short); Shlomit Yanisky-Ravid, Generating Rembrandt: Artificial Intelligence, Copyright, and Accountability in the 3A Era-The Human-Like Authors Are Already Here-A New Model, 2017 MICH. ST. L. REV. 659, 700 ("Unlike humans, AI systems do not need incentives to create artworks.").

18. See generally H.L.A. HART, THE CONCEPT OF LAW (1961) (providing a seminal work on legal positivism).

19. See International Copyright Issues and Artificial Intelligence, U.S. COPYRIGHT OFF. (July 26, 2023), https://www.copyright.gov/events/international-ai-copyright-webinar (exploring this question).

20. See Matthew Sag & Peter K. Yu, The Globalization of Copyright Exceptions for AI Training, 74 EMORY L.J. (forthcoming 2025) (discussing the ongoing copyright litigation in the generative AI context in the United States). Some industry groups, policymakers, and commentators have referred to such training as "ingestion." See, e.g., THE AUTHORS GUILD, COMMENTS OF THE AUTHORS GUILD: ARTIFICIAL INTELLIGENCE AND COPYRIGHT 15 (2023), https://authorsguild.org/app/uploads/2023/10/Authors-Guild-Comments-AI-and- Copyright-October-30-2023.pdf ("[T]raining [large language models], at this stage, requires ingestions of complete works."). However, that term is misleading. See Sag & Yu, supra ("In most cases, training data influences the model without becoming part of the model.").

21. For discussions at the intersection of intellectual property and AI, see generally AI AND IP, supra note 17; RESEARCH HANDBOOK, supra note 17.

22. See Ryan Abbott, Intellectual Property and Artificial Intelligence: An Introduction, in RESEARCH HANDBOOK, supra note 17, at 1, 16 (discussing the "Artificial Inventor Project").

23. See Sun Haochen, Redesigning Copyright Protection in the Era of Artificial Intelligence, 107 IOWA L. REV. 1213, 1215-16 (2022) (collecting these decisions).

24. See Rebecca Currey & Jane Owen, In the Courts: Australian Court Finds AI Systems Can Be "Inventors," WIPO MAG. (Sept. 2021), https://www.wipo.int/wipo_magazine/en/ 2021/03/article_0006.html; DABUS Gets Its First Patent in South Africa Under Formalities Examination, IP WATCHDOG (July 29, 2021), https://ipwatchdog.com/2021/07/29/dabusgets- first-patent-south-africa-formalities-examination/id=136116 [hereinafter DABUS Gets Its First Patent].

25. See Currey & Owen, supra note 24.

26. See DABUS Gets Its First Patent, supra note 24.

27. Commissioner of Patents v Thaler [2022] FCAFC 62, rev'g en banc [2021] FCA 879, cert. denied, [2022] HCA 199 (Austl.).

28. See Sukanya Sarkar, Exclusive: India Recognises AI as Co-Author of Copyrighted Artwork, MANAGING IP (Aug. 5, 2021), https://www.managingip.com/article/2a5czmpwixyj23wyqct1c/ exclusive-india-recognises-ai-as-co-author-of-copyrighted-artwork ("For the first time ever in India, the copyright office has recognised an artificial intelligence tool-RAGHAV Artificial Intelligence Painting App-as the co-author of a copyright-protected artistic work.").

29. See Sukanya Sarkar, Exclusive: Indian Copyright Office Issues Withdrawal Notice to AI Co-Author, MANAGING IP (Dec. 13, 2021), https://www.managingip.com/article/ 2a5d0jj2zjo7fajsjwwlc/exclusive-indian-copyright-office-issues-withdrawal-notice-to-ai-coauthor ("The Indian Copyright Office has issued a notice of withdrawal to Ankit Sahni, the man who secured India's first-ever copyright registration recognising an artificial intelligence tool as the co-author of an artwork . . . ."); see also Copyright Review Board's Decision on "Suryast," supra note 12 (providing the Copyright Review Board's decision); Vedika Chawla, Ankit Sahni's AI "Co-Authored" Artwork Denied Registration by US, Continues to Be Registered in India, SPICYIP (Dec. 15, 2023), https://spicyip.com/2023/12/ ankit-sahnis-ai-co-authored-artwork-denied-registration-by-us-continues-to-be-registeredin- india.html (discussing the denial of copyright protection for this AI-generated artwork in the United States).

30. See discussion infra Section II.A.

31. See, e.g., Manya Koetse, In the Race for AI Supremacy, China and the US Are Travelling on Entirely Different Tracks, THE GUARDIAN (Jan. 8. 2024), https://www.theguardian.com/ world/2024/jan/09/in-the-race-for-ai-supremacy-china-and-the-us-are-travelling-onentirely- different-tracks (discussing the AI race between China and the United States).

32. Directive 1996/9, of the European Parliament and of the Council on the Legal Protection of Databases, 1996 O.J. (L 77) 20 (EC) [hereinafter EU Database Directive].

33. Agreement on Trade-Related Aspects of Intellectual Property Rights, Apr. 15, 1994, Marrakesh Agreement Establishing the World Trade Organization, Annex 1C, 1869 U.N.T.S. 299 [hereinafter TRIPS Agreement].

34. See generally INTELLECTUAL PROPERTY & FREE TRADE AGREEMENTS (Christopher Heath & Anselm Kamperman Sanders eds., 2007) (collecting articles that discuss free trade agreements in the intellectual property context); Robert Burrell & Kimberlee Weatherall, Exporting Controversy? Reactions to the Copyright Provisions of the U.S.-Australia Free Trade Agreement: Lessons for U.S. Trade Policy, 2008 U. ILL. J.L. TECH. & POL'Y 259 (criticizing the United States-Australia Free Trade Agreement); Peter K. Yu, Currents and Crosscurrents in the International Intellectual Property Regime, 38 LOY. L.A. L. REV. 323, 392-400 (2004) [hereinafter Yu, Currents and Crosscurrents] (discussing the growing use of bilateral and regional trade agreements to push for higher intellectual property standards).

35. U.S. CONST. art. I, § 8, cl. 8.

36. See JAMES MADISON, NOTES OF DEBATES IN THE FEDERAL CONVENTION OF 1787, at 477-78 (Adrienne Koch ed., 1966) (identifying these proposals).

37. See id. at 580-81.

38. James Madison offered the following brief commentary in The Federalist: The utility of [the copyright] power will scarcely be questioned. The copyright of authors has been solemnly adjudged, in Great Britain, to be a right of the common law. The right to useful inventions seems with equal reason to belong to the inventors. The public good fully coincides in both cases with the claims of individuals. The States cannot separately make effectual provision for either of the cases, and most of them have anticipated the decision of this point, by laws passed at the instance of Congress.

THE FEDERALIST NO. 43, at 271-72 (James Madison) (Clinton Rossiter ed., 1961).

39. Howard B. Abrams, Copyright, Misappropriation, and Preemption: Constitutional and Statutory Limits of State Law Protection, 1983 SUP. CT. REV. 509, 516 n.38.

40. Peter K. Yu, How Copyright Law Could Affect Pop Music Without Our Knowing It, 83 UMKC L. REV. 363, 366 n.15 (2014).

41. Burrow-Giles Lithographic Co. v. Sarony, 111 U.S. 53, 57-58 (1884).

42. 17 U.S.C. §§ 101, 201, 203, 304.

43. Naruto v. Slater, 888 F.3d 418, 426 (9th Cir. 2018).

44. See Daniel J. Gervais, The Machine as Author, 105 IOWA L. REV. 2053, 2073-85 (2020) (discussing the humanness for authorship); see also Sam Ricketson, The 1992 Horace S. Manges Lecture-People or Machines: The Bern Convention and the Changing Concept of Authorship, 16 COLUM.-VLA J.L. & ARTS 1, 10 (1991) (characterizing "the need for the author to be a human being" as an attribute of authorship for the purposes of protection under the Berne Convention for the Protection of Literary and Artistic Works (Berne Convention)).

45. U.S. COPYRIGHT OFF., COMPENDIUM OF U.S. COPYRIGHT OFFICE PRACTICES § 313.2 (3d ed. 2021).

46. See supra text accompanying notes 9-12.

47. Burrow-Giles, 111 U.S. at 58 (emphasis added).

48. Arthur R. Miller, Copyright Protection for Computer Programs, Databases, and Computer-Generated Works: Is Anything New Since CONTU?, 106 HARV. L. REV. 977, 1062 (1993).

49. Burrow-Giles, 111 U.S. at 58 (emphasis added).

50. Id.

51. U.S. COPYRIGHT OFF., supra note 45, § 313.2.

52. See Peter K. Yu, The Future Path of Artificial Intelligence and Copyright Law in the Asian Pacific, 96 COMPUTS. & L. (forthcoming 2025) [hereinafter Yu, Future Path]; cf. Lindsay v. Wrecked & Abandoned Vessel R.M.S. Titanic, No. 97 Civ. 9248(HB), 1999 WL 816163, at ·5 (S.D.N.Y. Oct. 13, 1999) ("All else being equal, where a plaintiffalleges that he exercised such a high degree of control over a film operation . . . such that the final product duplicates his conceptions and visions of what the film should look like, the plaintiffmay be said to be an 'author' within the meaning of the Copyright Act." (emphasis added)).

53. Limoumou Su Liumoumou Qinhai Zuopin Shumingquan, Xinxi Wangluo Chuanboquan Jiufen An (...) [Li v. Liu], (2023) Jing 0491 Min Chu No. 11279 ((2023)...0491...11279...) (Beijing Internet Ct. Nov. 27, 2023), translated at https://patentlyo.com/media/2023/12/Li-v-Liu-Beijing-Internet- Court-20231127-with-English-Translation.pdf. For discussions of this case, see generally He Tianxiang, AI Originality Revisited: Can We Prompt Copyright over AI-Generated Pictures?, 73 GRUR INT'L 299 (2024); Lu Tian, Chinese Court Deems AI-Generated Image Has Copyright- Assessing the Possibly Over-Hasty "Spring Breeze" Case, IPKAT (Dec. 27, 2023), https://ipkitten.blogspot.com/2023/12/chinese-court-deems-ai-generated-image.html. This Article features the Author's translation.

54. See discussion infra Section I.B.

55. See generally Edward Lee, Prompting Progress: Authorship in the Age of AI, 76 FLA. L. REV. 1445 (2024) [hereinafter Lee, Prompting Progress]; Edward Lee, The Code Red for Copyright Law, 76 FLA. L. REV. F. 1 (2024).

56. Lee, Prompting Progress, supra note 55, at 1581.

57. Id.

58. Id.

59. See supra notes 9-14 and accompanying text.

60. See infra text accompanying notes 97-102.

61. See generally William M. Landes & Richard A. Posner, An Economic Analysis of Copyright Law, 18 J. LEGAL STUD. 325, 326-33 (1989) (discussing the basic economics of copyright).

62. Twentieth Century Music Corp. v. Aiken, 422 U.S. 151, 156 (1975).

63. Peter K. Yu, Anticircumvention and Anti-anticircumvention, 84 DENV. U. L. REV. 13, 17 (2006).

64. See supra text accompanying notes 16-17.

65. One could debate whether investment protection theory is part of incentive theory, broadly construed. See Hilty et al., supra note 17, at 61-70 (discussing different incentive theories that may be implicated in the AI context). This section does not intend to weigh in on this debate. Instead, it aims to show that arguments exist on both sides of the debate concerning whether copyright protection is needed to provide AI developers with economic incentives. Thanks to Shani Shisha for asking a question in this direction.

66. Such investment, however, is often speculative. Cf. Peter K. Yu, The Comparative Law and Economics of Counterfeits and Post-Sale Confusion, in RESEARCH HANDBOOK ON THE LAW AND ECONOMICS OF TRADEMARK LAW 363, 366 (Glynn S. Lunney, Jr. ed., 2023) (questioning whether stronger trademark protection would necessarily result in more investment in quality). AI developers can always choose not to reinvest profits to further technological development. Even if they opt to reinvest portions of those profits, it is frequently debatable whether such reinvestment will be substantial enough to justify the high costs of protection in the first place.

67. EU Database Directive, supra note 32, art. 7(1).

68. Id.

69. See Communication from the Commission on "Building a European Data Economy," at 13, COM(2017) 9 final (Oct. 1, 2017) (advancing the proposal); Commission StaffWorking Document on the Free Flow of Data and Emerging Issues of the European Data Economy, at 33-36, SWD (2017) 2 final (examining the proposal); see also Peter K. Yu, Data Producer's Right and the Protection of Machine-Generated Data, 93 TUL. L. REV. 859, 884-96 (2019) [hereinafter Yu, Data Producer's Right] (discussing whether the proposed EU data producer's right would meet the needs of our present technological environment and of the U.S. business and scientific communities and society at large).

70. See generally Dev S. Gangjee, The Data Producer's Right-An Instructive Obituary, in THE CAMBRIDGE HANDBOOK OF PRIVATE LAW AND ARTIFICIAL INTELLIGENCE 332 (Ernest Lim & Phillip Morgan eds., 2024) (discussing the end of the debate on the data producer's right).

71. Regulation 2023/2854, of the European Parliament and of the Council on Harmonised Rules on Fair Access to and Use of Data and Amending Regulation (EU) 2017/2394 and Directive (EU) 2020/1828 (Data Act), 2023 O.J. (L 71) 1.

72. See discussion infra Section II.A.

73. See generally Peter K. Yu, Intellectual Property, Foreign Investment and Sustainable Development, in THE ELGAR COMPANION TO INTELLECTUAL PROPERTY AND THE SUSTAINABLE DEVELOPMENT GOALS 537, 543-51 (Bita Amani, Caroline Ncube & Matthew Rimmer eds., 2024) (discussing investment protection regimes).

74. See Yu, Data Producer's Right, supra note 69, at 923-25 (discussing the complications the data producer's right would create in the investment protection regime).

75. See Rochelle Cooper Dreyfuss, ISDS and Intellectual Property in 2020: Protecting Public Health in the Age of Pandemics, in YEARBOOK ON INTERNATIONAL INVESTMENT LAW & POLICY 2020, at 206, 213 (Lisa E. Sachs, Lise J. Johnson & Jesse Coleman eds., 2022) (noting that the pandemic-related dispute filed in 2020 did not involve intellectual property assets); Peter K. Yu, The Changing Chemistry Between Intellectual Property and Investment Law, in IMPROVING INTELLECTUAL PROPERTY: A GLOBAL PROJECT 405, 411 (Susy Frankel, Margaret Chon, Graeme B. Dinwoodie, Barbara Lauriat & Jens Schovsbo eds., 2023) ("Since the COVID-19 pandemic, the use of [investor-state dispute settlement] in the intellectual property area has grounded to a halt."). Investor-state dispute settlement allows multinational corporations to use international arbitration to resolve cross-border intellectual property disputes with host states. For the Author's discussion of the use of investor-state dispute settlement in the intellectual property context, see generally Peter K. Yu, The Investment-Related Aspects of Intellectual Property Rights, 66 AM. U. L. REV. 829 (2017); Peter K. Yu, The Pathways of Multinational Intellectual Property Dispute Settlement, in INTELLECTUAL PROPERTY AND INTERNATIONAL DISPUTE RESOLUTION 123, 132-36 (Christopher Heath & Anselm Kamperman Sanders eds., 2019).

76. See Peter K. Yu, The International Enclosure Movement, 82 IND. L.J. 827, 892-901 (2007) [hereinafter Yu, International Enclosure Movement] (discussing the incentiveinvestment divide).

77. See Rochelle Cooper Dreyfuss, Expressive Genericity: Trademarks as Language in the Pepsi Generation, 65 NOTRE DAME L. REV. 397, 405-07 (1990) [hereinafter Dreyfuss, Expressive Genericity] (criticizing the "if value, then right" logic); see also Wendy J. Gordon, On Owning Information: Intellectual Property and the Restitutionary Impulse, 78 VA. L. REV. 149, 178-80 (1992) (criticizing the "value is property" model). See generally Alfred C. Yen, Brief Thoughts About If Value/Then Right, 99 B.U. L. REV. 2479 (2019) (discussing the "if value/then right" principle and its consequences for intellectual property law).

78. Felix S. Cohen, Transcendental Nonsense and the Functional Approach, 35 COLUM. L. REV. 809, 811-12 (1935).

79. Id. at 815; see also Dreyfuss, Expressive Genericity, supra note 77, at 409 ("By equating 'value' with 'right,' the decisions fail to create an internal reference point against which to measure the need for exclusivity.").

80. See Daryl Lim & Peter K. Yu, The Antitrust-Copyright Interface in the Age of Generative Artificial Intelligence, 74 EMORY L.J. (forthcoming 2025).

81. ANU BRADFORD, DIGITAL EMPIRES: THE GLOBAL BATTLE TO REGULATE TECHNOLOGY (2023).

82. See Peter K. Yu, Toward a Nonzero-Sum Approach to Resolving Global Intellectual Property Disputes: What We Can Learn from Mediators, Business Strategists, and International Relations Theorists, 70 U. CIN. L. REV. 569, 606-08 (2002).

83. See John F. Duffy, Harmony and Diversity in Global Patent Law, 17 BERKELEY TECH. L.J. 685, 702-03 (2002). But see Ruth Gana Okediji, Copyright and Public Welfare in Global Perspective, 7 IND. J. GLOBAL LEGAL STUD. 117, 125-44 (1999) (exploring the tension between free trade and intellectual property protection).

84. See Duffy, supra note 83, at 699-701.

85. See Claudio R. Frischtak, Harmonization Versus Differentiation in Intellectual Property Rights Regime, in GLOBAL DIMENSIONS OF INTELLECTUAL PROPERTY RIGHTS IN SCIENCE AND TECHNOLOGY 89, 93-97 (Mitchel B. Wallerstein, Mary Ellen Mogee & Roberta A. Schoen eds., 1993) (arguing that countries should tailor their intellectual property system by taking into account their economic needs, productive and research capabilities, and institutional and budgetary constraints); Yu, International Enclosure Movement, supra note 76, at 828 (criticizing the adoption of "one-size-fits-all legal standards that ignore their local needs, national interests, technological capabilities, institutional capacities, and public health conditions").

86. See Duffy, supra note 83, at 707-08.

87. See id. at 703-06.

88. See id. at 706-07.

89. See EU Database Directive, supra note 32, art. 11(3), at 27 (extending sui generis database protection to third countries based on agreements); see also J.H. Reichman & Pamela Samuelson, Intellectual Property Rights in Data?, 50 VAND. L. REV. 51, 96-97 (1997) (discussing the European Commission's "strict criterion of material reciprocity"); Yu, Currents and Crosscurrents, supra note 34, at 379 (discussing the reciprocal provisions in the EU Database Directive).

90. See Yu, Data Producer's Right, supra note 69, at 875-79 (identifying five reasons why Congress declined to introduce sui generis database protection in the United States).

91. Feist Publ'ns, Inc. v. Rural Tel. Serv. Co., 499 U.S. 340, 346 (1991).

92. See Yu, Data Producer's Right, supra note 69, at 876.

93. Peter K. Yu, Bridging the Digital Divide: Equality in the Information Age, 20 CARDOZO ARTS & ENT. L.J. 1, 46 (2002).

94. See Reichman & Samuelson, supra note 89, at 113-24 (discussing the adverse impact of sui generis database protection on scientific research and education); J.H. Reichman & Paul F. Uhlir, Database Protection at the Crossroads: Recent Developments and Their Impact on Science and Technology, 14 BERKELEY TECH. L.J. 793, 796-821 (1999) (discussing the adverse impact of database protection laws on scientific, technical, and educational users of factual data and information).

95. See Yochai Benkler, Constitutional Bounds of Database Protection: The Role of Judicial Review in the Creation and Definition of Private Rights in Information, 15 BERKELEY TECH. L.J. 535, 562-65 (2000) (discussing the anticompetitive nature of database protection laws); Reichman & Samuelson, supra note 89, at 124-30 (discussing how sui generis database protection would impede competition in the market for value-adding products and services).

96. Comm'n of the Eur. Cmtys., First Evaluation of Directive 96/9/EC on the Legal Protection of Databases 22, 24 (Dec. 12, 2005), cited in Yu, Data Producer's Right, supra note 69, at 866 n.30; see also James Boyle, James Boyle: Two Database Cheers for the EU, FIN. TIMES (Jan. 2, 2006), https://www.ft.com/content/99610a50-7bb2-11da-ab8e-0000779e2340 (discussing the European Commission's evaluation report).

97. See, e.g., Dominican Republic-Central America Free Trade Agreement, Aug. 5, 2004, https://ustr.gov/sites/default/files/uploads/agreements/cafta/asset_upload_file934_ 3935.pdf; TRIPS Agreement, supra note 33; United States-Australia Free Trade Agreement, U.S.-Austl., May 18, 2004, https://ustr.gov /sites/default/files/uploads/agreements/fta/ australia/asset_upload_file148_5168.pdf; United States-Singapore Free Trade Agreement, U.S.-Sing., May 6, 2003, https://ustr.gov/sites/default/files/uploads/agreements/fta/ singapore/asset_upload_file708_4036.pdf [hereinafter United States-Singapore FTA].

98. See Pub. L. No. 103-465 § 514, 108 Stat. 4809, 4976-81 (1994) (codified as amended at 17 U.S.C. 104A).

99. 35 U.S.C. § 154(c)(1).

100. 15 U.S.C. § 1051(b) (allowing for trademark application based on "bona fide intention to use [the] trademark").

101. See, e.g., 17 U.S.C. § 302(a) (providing a copyright term of the life of the author plus seventy years); Council Directive 93/98, Harmonizing the Term of Protection of Copyright and Certain Related Rights, art. 1(1), 1993 O.J. (L 290) 9, 11 (EEC) (providing a copyright term of the life of the author plus seventy years).

102. Leahy-Smith America Invents Act, Pub. L. No. 112-29, 125 Stat. 284 (2011).

103. See supra text accompanying notes 97-102.

104. See Peter K. Yu, Clusters and Links in Asian Intellectual Property Law and Policy, in ROUTLEDGE HANDBOOK OF ASIAN LAW 147, 150-51 (Christoph Antons ed., 2017) [hereinafter Yu, Clusters and Links] (noting the role of the WTO and its TRIPS Agreement in driving convergence of intellectual property laws in Asia).

105. See discussion infra Section II.A.

106. See, e.g., Authors Guild v. Google, Inc., 804 F.3d 202 (2d Cir. 2015); Authors Guild, Inc. v. HathiTrust, 755 F.3d 87 (2d Cir. 2014); A.V. ex rel. Vanderhye v. iParadigms, LLC, 562 F.3d 630 (4th Cir. 2009). For the discussions of the non-expressive use of copyrighted works and related defenses, see generally Edward Lee, Technological Fair Use, 83 S. CAL. L. REV. 797 (2010); Matthew Sag, Copyright and Copy-Reliant Technology, 103 NW. U. L. REV. 1607 (2009) [hereinafter Sag, Copy-Reliant Technology]; Sag & Yu, supra note 20; Pamela Samuelson, Fair Use Defenses in Disruptive Technology Cases, 71 UCLA L. REV. (forthcoming 2024).

107. See Sag & Yu, supra note 20 (discussing the ongoing copyright litigation in the generative AI context in the United States).

108. See discussion infra Section II.B.

109. For discussions of fair use in the technological context, see generally Lee, Technological Fair Use, supra note 106; Sag, Copy-Reliant Technology, supra note 106.

110. See Sag & Yu, supra note 20.

111. The discussion of AI developments in Asia in this Part draws on Yu, Future Path, supra note 52.

112. See Copyright and Related Rights Act 2000 (Act. No. 28/2000) § 21(f) (Ir.) (stipulating that "in the case of a work which is computer-generated, [the author means the person who creates a work and includes] the person by whom the arrangements necessary for the creation of the work are undertaken"); Copyright Act 1994, s 5(2)(a) (N.Z.) (stating that "in the case of a literary, dramatic, musical, or artistic work that is computer-generated, [the person who creates a work 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, c. 48, § 9(3) (UK) ("In the case of 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.").

113. Beijing Feilin Lushi Shiwusuo Su Beijing Baidu Wangxun Keji Youxian Gongsi Zhuzuoquan Qinquan Jiufen Yi An (...) [Beijing Film Law Firm v. Beijing Baidu Netcom Sci. Tech. Co.], (2018) Jing 0491 Min Chu No. 239 ((2018)...0491...239...) (Beijing Internet Ct. Apr. 25, 2019), https:// www.bjinternetcourt.gov.cn/cac/zw/1556272978673.html; see also Beijing Internet Court Ruling in First Case of Copyright Infringement of AI-Generated Article, BEIJING INTERNET CT. (May 30, 2019), https://english.bjinternetcourt.gov.cn/2019-05/30/c_170.htm.

114. Shenzhen Tengxun Su Shanghai Yingxun Zhuzuoquan Qinquan An (...) [Shenzhen Tencent Comput. Sys. Co. v. Shanghai Yingxun Tech. Co.] (2019) Yue 0305 Min Chu No. 14010 ((2019)...0305...14010...) (Shenzhen Nanshan Dist. Ct. Dec. 24, 2019), https://www.chinajusticeobserver.com/law/x/2019-yue-0305-min-chu- 14010/chn [hereinafter Tencent Case]; see also Zhou Bo, Artificial Intelligence and Copyright Protection-Judicial Practice in Chinese Courts, WORLD INTELL. PROP. ORG., https://www. wipo.int/export/sites/www/about-ip/en/artificial_intelligence/conversation_ip_ai/pdf/ ms_china_1_en.pdf (last visited Oct. 4, 2024).

115. Tencent Case, supra note 114.

116. Li v. Liu, supra note 53.

117. Id.

118. See discussion supra Part I.

119. Kurt B. Gerstner, Copyright and AI-The Korean View, DENTONS (Apr. 11, 2022), https://www.dentons.com/en/insights/articles/2022/april/11/copyright-and-ai-thekorean- view.

120. See, e.g., Sun, supra note 23, at 1236-48 (advocating the provision of ten years of sui generis protection to AI works generated with human contributions).

121. See generally LEE KAI-FU, AI SUPERPOWERS: CHINA, SILICON VALLEY, AND THE NEW WORLD ORDER (2018) (discussing China's AI developments and related technological ambitions).

122. See Guowuyuan Guanyu Yinfa Xinyidai Rengong Zhineng Fazhan Guihua De Tongzhi, Guofa [2017] Sanshiwu Hao (... 2017...35...) [Notice of the Next-Generation Artificial Intelligence Development Plan, Notice No. 35 [2017]] (issued by the State Council, July 20, 2017).

123. See Sag & Yu, supra note 20 (discussing this race).

124. See Yu, International Enclosure Movement, supra note 76, at 892-901 (discussing the incentive-investment divide).

125. See generally Yu, Clusters and Links, supra note 104, at 148-50 (discussing these origins).

126. Peter K. Yu, The Rise and Decline of the Intellectual Property Powers, 34 CAMPBELL L. REV. 525, 554 (2012).

127. See generally CHATGPT IS EATING THE WORLD, https://chatgptiseatingtheworld.com (last visited Oct. 4, 2024) (collecting and discussing these cases); DAIL-THE DATABASE OF AI LITIGATION, https://blogs.gwu.edu/law-eti/ai-litigation-database (last visited Oct. 12, 2024) (providing a database about ongoing and completed AI litigation).

128. See Michael M. Grynbaum & Ryan Mac, The Times Sues OpenAI and MicrosoftOver A.I. Use of Copyrighted Work, N.Y. TIMES (Dec. 27, 2023), https://www.nytimes.com/2023/ 12/27/business/media/new-york-times-open-ai-microsoft-lawsuit.html; Aimee Picchi, George R.R. Martin, John Grisham and Other Major Authors Sue OpenAI, Alleging "Systematic Theft," MONEY WATCH (Sept. 20, 2023), https://www.cbsnews.com/news/openai-lawsuitgeorge- rr-martin-john-grisham-copyright-infringement; Zachary Small, Sarah Silverman Sues OpenAI and Meta over Copyright Infringement, N.Y. TIMES (July 10, 2023), https://www. nytimes.com/2023/07/10/arts/sarah-silverman-lawsuit-openai-meta.html.

129. 17 U.S.C. § 107.

130. Peter K. Yu, Fair Use and Its Global Paradigm Evolution, 2019 U. ILL. L. REV. 111, 115- 16 [hereinafter Yu, Paradigm Evolution]. It is worth adding that "there are remarkable similarities between the fair dealing regime in Canada and the fair use regime in the United States." Id.

131. For discussions of the importance of TDM exceptions to the copyright systems, see generally Michael W. Carroll, Copyright and the Progress of Science: Why Text and Data Mining Is Lawful, 53 U.C. DAVIS L. REV. 893 (2019); Christophe Geiger, Giancarlo Frosio & Oleksandr Bulayenko, Crafting a Text and Data Mining Exception for Machine Learning and Big Data in the Digital Single Market, in INTELLECTUAL PROPERTY AND DIGITAL TRADE IN THE AGE OF ARTIFICIAL INTELLIGENCE AND BIG DATA 95 (Xavier Seuba, Christophe Geiger & Julien Pénin eds., 2018); Matthew Sag, The New Legal Landscape for Text Mining and Machine Learning, 66 J. COPYRIGHT SOC'Y U.S.A. 291 (2019) [hereinafter Sag, New Legal Landscape].

132. Copyright, Designs and Patents Act 1988, c. 48, § 29A(1).

133. Directive 2019/790, of the European Parliament and of the Council on Copyright and Related Rights in the Digital Single Market and Amending Directives 96/9/EC and 2001/29/EC, arts. 3-4, 2019 O.J. (L 130) 92, 113-14 [hereinafter DSM Directive]. Article 3 nonetheless limits the exception's applicability to "research organisations and cultural heritage institutions," while Article 4 conditions the exception on the lack of express reservation by the relevant rights holders. Id.

134. Id. art. 2(2), at 112.

135. TRIPS Agreement, supra note 33, art. 9.2.

136. See Sag & Yu, supra note 20.

137. CHOSAKUKENHŌ [Copyright Act], Law No. 48 of 1970, art. 47-7 (Japan) (amended 2009).

138. See He Tianxiang, Copyright Exceptions Reform and AI Data Analysis in China: A Modest Proposal, in AI AND IP, supra note 17, at 196, 209-11 (comparing the old Articles 30-4 and 47-7 with the amended Article 30-4).

139. CHOSAKUKENHŌ [Copyright Act], Law No. 48 of 1970, art. 30-4 (Japan) [hereinafter Japanese Copyright Act]. For discussions of this provision, see generally Sag & Yu, supra note 20; Tatsuhiro Ueno, The Flexible Copyright Exception for "Non-Enjoyment" Purposes- Recent Amendment in Japan and Its Implication, 70 GRUR INT'L 145 (2021); Yu, Future Path, supra note 52.

140. Japanese Copyright Act, supra note 139, art. 30-4.

141. Id. arts. 47-4, 47-5.

142. Id. art. 30-4.

143. JAPAN COPYRIGHT OFF., GENERAL UNDERSTANDING ON AI AND COPYRIGHT IN JAPAN: OVERVIEW 8 (2024), https://www.bunka.go.jp/english/policy/copyright/pdf/ 94055801_01.pdf.

144. See id. at 11. The document elaborates on what it would mean to "unreasonably prejudice the interests of the copyright holder" under Article 30-4 in the context of reproducing a copyrighted database for AI training when licenses are available, especially if the rightsholder has restricted its use via technological protection measures. See id. at 10.

145. See Ueno, supra note 139, at 149 ("Japan could be said to be a 'paradise' for machine learning and TDM.").

146. See generally ALAN WATSON, LEGAL TRANSPLANTS: AN APPROACH TO COMPARATIVE LAW (2d ed. 1993) (providing the seminal work on legal transplants). Paul Edward Geller defines "legal transplant" as "any legal notion or rule which, after being developed in a 'source' body of law, is . . . introduced into another, 'host' body of law." Paul Edward Geller, Legal Transplants in International Copyright: Some Problems of Method, 13 UCLA PAC. BASIN L.J. 199, 199 (1994). For the Author's discussions of legal transplants, see generally Peter K. Yu, Can the Canadian UGC Exception Be Transplanted Abroad?, 26 INTELL. PROP. J. 175 (2014); Peter K. Yu, Digital Copyright Reform and Legal Transplants in Hong Kong, 48 U. LOUISVILLE L. REV. 693, 709-13 (2010) [hereinafter Yu, Digital Copyright Reform]; Peter K. Yu, The Transplant and Transformation of Intellectual Property Laws in China, in GOVERNANCE OF INTELLECTUAL PROPERTY RIGHTS IN CHINA AND EUROPE 20 (Nari Lee, Niklas Bruun & Li Mingde eds., 2016).

147. Some commentators have studied these reverse transplants in the non-intellectual property context. See, e.g., Sital Kalantry, Reverse Legal Transplants, 99 N.C. L. REV. 49 (2021) (examining the development of "reverse legal transplants" in the context of abortion restrictions).

148. Copyright Act 2021 (2020 Rev Ed) § 244 (Sing.) [hereinafter Singapore Copyright Act].

149. See INTELL. PROP. OFF. OF SING., COPYRIGHT FACTSHEET ON COPYRIGHT ACT 2021, at 18 (2022), https://www.ipos.gov.sg/docs/default-source/resources-library/copyright/ copyright-act-factsheet.pdf (noting that this exception "will always be available, regardless of any contract term purporting to prevent or restrict them").

150. Singapore Copyright Act, supra note 148, § 244(2)(a)-(b).

151. Id. § 244(2)(c).

152. See id. § 244(2)(d).

153. See id. § 244(2)(e).

154. See id.

155. Id. § 243.

156. INTELL. PROP. OFF. OF SING., supra note 149, at 13.

157. Singapore Copyright Act, supra note 148, § 243.

158. Id. § 244.

159. Despite this possibility, it is worth pointing out that fair use dates back to the 1841 case of Folsom v. Marsh and remains an essential tool to facilitate AI development in the United States. Folsom v. Marsh, 9 F. Cas. 342 (C.C.D. Mass. 1841) (No. 4901). For discussions of Folsom, see generally L. Ray Patterson, Folsom v. Marsh and Its Legacy, 5 J. INTELL. PROP. L. 431 (1998); R. Anthony Reese, The Story of Folsom v. Marsh: Distinguishing Between Infringing and Legitimate Uses, in INTELLECTUAL PROPERTY STORIES 259 (Jane C. Ginsburg & Rochelle C. Dreyfuss eds., 2006); Peter K. Yu, Tales of the Unintended in Copyright Law, 67 STUD. L. POL. & SOC'Y 1, 2-6 (2015).

160. A common form of pushback is the action taken by the Office of the United States Trade Representative under Section 301 of the 1974 Trade Act. The provision permits the U.S. President to investigate and impose sanctions on countries engaging in unfair trade practices that threaten the United States' economic interests, including the inadequate protection and enforcement of intellectual property rights. See 19 U.S.C. §§ 2411-2420 (2018). For discussions of the operation of the Section 301 process, see generally Joe Karaganis & Sean Flynn, Networked Governance and the USTR, in MEDIA PIRACY IN EMERGING ECONOMIES 75 (Joe Karaganis ed., 2011); Paul C.B. Liu, U.S. Industry's Influence on Intellectual Property Negotiations and Special 301 Actions, 13 UCLA PAC. BASIN L.J. 87 (1994).

161. Copyright Act (Cap 63, 2006 Rev Ed) § 35 (Sing.).

162. United States-Singapore Free Trade Agreement, May 6, 2003, U.S.-Sing., https://ustr. gov/sites/default/files/uploads/agreements/fta/singapore/asset_upload_file708_4036.pdf.

163. Peter K. Yu, Customizing Fair Use Transplants, LAWS, Mar. 2018, no. 9, at 7.

164. See Singapore Copyright Act, supra note 148, §§ 190-194.

165. Compare 17 U.S.C. § 107, with Singapore Copyright Act, supra note 148, § 243.

166. Zhonghua Renmin Gongheguo Zhuzuoquan Fa ( ...) [Copyright Law of the People's Republic of China] [Chinese Copyright Law] (promulgated by the Standing Comm. Nat'l People's Cong., Sept. 7, 1990, amended Nov. 11, 2020, effective June 1, 2021), art. 24 http://www.npc.gov.cn/englishnpc/c23934/202109/ ae0f0804894b4f71949016957eec45a3.shtml (China).

167. Id. art. 24(13). For discussions of this provision, see generally He Tianxiang, The Copyright Limitations of the 2020 Copyright Law of China: A Satisfactory Compromise?, 69 J. COPYRIGHT SOC'Y U.S.A. 107 (2022); Hua (Jerry) Jie, Copyright Exceptions for Text and Data Mining in China: Inspiration from Transformative Use, 69 J. COPYRIGHT SOC'Y U.S.A. 123 (2022). For discussions of the Third Amendment, see generally Peter K. Yu, The Long and Winding Road to Effective Copyright Protection in China, 49 PEPP. L. REV. 681 (2022) [hereinafter Yu, Long and Winding Road]; Peter K. Yu, Third Amendment to the Chinese Copyright Law, 69 J. COPYRIGHT SOC'Y U.S.A. 1 (2022). See generally Symposium, Third Amendment to the Chinese Copyright Law, 69 J. COPYRIGHT SOC'Y U.S.A. 5 (2022) (collecting essays that closely examine this amendment).

168. Chinese Copyright Law, supra note 166, art. 24(13).

169. See Yu, Long and Winding Road, supra note 167, at 721 (noting "the drafting of the pending implementing regulations").

170. Interim Measures, supra note 6.

171. EU AI Act, supra note 5.

172. Interim Measures, supra note 6, art. 4(3).

173. Id. art. 7(2).

174. Sag, Copy-Reliant Technology, supra note 106, at 1630.

175. See Sag, New Legal Landscape, supra note 131, at 299, 302 ("TDM is . . . an inherent part of Artificial Intelligence research using machine learning. . . . Th[e] distinction between expressive and non-expressive uses of copyrighted works is essential to understanding how copyright should apply to TDM."); Oren Bracha, The Work of Copyright in the Age of Machine Production, 38 HARV. J.L. & TECH. (forthcoming 2024) (calling for greater attention in the AI context to the enjoyment of the expressive value of copyrighted works).

176. Lim & Yu, supra note 80.

177. Japanese Copyright Act, supra note 139, art. 30-4.

178. See Copyright, Designs and Patents Act, 1988, c. 48, § 29A(1) (UK) (permitting "[t]he making of a copy of a work by a person who has lawful access to the work" in order to "carry out a computational analysis of anything recorded in the work for the sole purpose of research for a non-commercial purpose"); DSM Directive, supra note 133, arts. 3-4, at 113- 14 (providing TDM exceptions for "research organisations and cultural heritage institutions" and for those using copyrighted works whose use has not been expressly reserved by the relevant copyright holders); Singapore Copyright Act, supra note 148, § 244(2)(a)-(b) (providing a copyright exception for computational data analysis).

179. Mark A. Lemley, How Generative AI Turns Copyright Upside Down, 25 COLUM. SCI. & TECH. L. REV. 190, 190 (2024) [hereinafter Lemley, Generative AI].

180. 17 U.S.C. § 102(a).

181. Id. § 102(b).

182. Johnson Controls, Inc. v. Phoenix Control Sys., Inc., 886 F.2d 1173, 1175 (1989).

183. See JOINT INST. FOR INNOVATION POL'Y & INST. FOR INFO. L., UNIV. OF AMSTERDAM, TRENDS AND DEVELOPMENTS IN ARTIFICIAL INTELLIGENCE: CHALLENGES TO THE INTELLECTUAL PROPERTY RIGHTS FRAMEWORK 82 (2020) [hereinafter IVIR AI STUDY] ("Copyright doctrine and case law lend support to our conclusion that the production of an artefact executed by a largely autonomous AI system could qualify as a work protected under EU copyright law on condition that a human being initiated and conceived the work and subsequently redacted the AI-assisted output in a creative manner."); PAMELA SAMUELSON, CHRISTOPHER JON SPRIGMAN & MATTHEW SAG, COMMENTS IN RESPONSE TO THE COPYRIGHT OFFICE'S NOTICE OF INQUIRY ON ARTIFICIAL INTELLIGENCE AND COPYRIGHT 3-4 (2023), https://www.regulations.gov/comment/COLC-2023-0006-8854 ("[T]here is no reason in principle why prompts couldn't be detailed enough to meet the traditional threshold of authorship in some cases. . . . Furthermore, refining a series of text prompts and choosing among different outputs should also be recognized as a way in which a human using Generative AI could meet the authorship standard."); Lemley, Generative AI, supra note 179, at 200 ("[A]s people become more accustomed to using generative AI, perhaps they will write more and more detailed prompts to tailor the output to what they want. And a sufficiently detailed direction to a computer may embody creativity, just as a sufficiently detailed instruction to a human camera operator would.").

184. U.S. COPYRIGHT OFF., supra note 45, § 313.2.

185. See Copyright Review Board's Decision on "Théâtre D'opéra Spatial," supra note 11, at 6 ("Mr. Allen describes 'input[ting] numerous revisions and text prompts at least 624 times' before producing the Midjourney Image . . . .").

186. See Li v. Liu, supra note 53.

187. See id.

188. See Copyright Registration of Computer Programs, U.S. COPYRIGHT OFF., https://www.copyright.gov/circs/circ61.pdf.

189. See Lemley, Generative AI, supra note 179, at 206 ("Access will have to play a much larger role and similarity a much smaller role in a prompt-based copyright infringement system."). See generally id. at 202-08 (discussing how a prompt-based model of copyright would upend the traditional test for copyright infringement, which includes analyses based on access and substantial similarities).

190. See, e.g., Ty, Inc. v. GMA Accessories, Inc., 132 F.3d 1167, 1169 (7th Cir. 1997) ("The Copyright Act forbids only copying; if independent creation results in an identical work, the creator of that work is free to sell it."); Selle v. Gibb, 741 F.2d 896, 901 (7th Cir. 1984) ("[N]o matter how similar the two works may be (even to the point of identity), if the defendant did not copy the accused work, there is no infringement.").

191. See WORLD INTELL. PROP. ORG. [WIPO], REVISED ISSUES PAPER ON INTELLECTUAL PROPERTY POLICY AND ARTIFICIAL INTELLIGENCE 4 (2020) ("'AI-generated' and 'generated autonomously by AI' are terms that are used interchangeably and refer to the generation of an output by AI without human intervention. . . . This is to be distinguished from 'AIassisted' outputs that are generated with material human intervention and/or direction.").

192. See Frank Pasquale, A Rule of Persons, Not Machines: The Limits of Legal Automation, 87 GEO. WASH. L. REV. 1, 54 (2019) (calling on the legal profession to focus more on intelligence augmentation); Liza Vertinsky & Todd M. Rice, Thinking About Thinking Machines: Implications of Machine Inventors for Patent Law, 8 B.U. J. SCI. & TECH. L. 574, 612 (2002) ("'[I]ntelligence augmentation' allows the effects of automatization to creep up the skill chain, providing for the substitution of white collar jobs by machines and allowing people with less formal training and education to perform more sophisticated tasks."); Albert H. Yoon, The Post-Modern Lawyer: Technology and the Democratization of Legal Representation, 66 U. TORONTO L.J. 456, 466 (2016) ("Intelligence augmentation . . . reflects a symbiotic relationship between humans and technology. Humans continue to perform the task at hand, but they do so interactively with technology in order to do it better.").

193. See Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence, 37 C.F.R. Part 202 (2023) ("[A]pplicants have a duty to disclose the inclusion of AI-generated content in a work submitted for registration and to provide a brief explanation of the human author's contributions to the work.").

194. EU AI Act, supra note 5, art. 53(1)(d); see also Sag & Yu, supra note 20 (discussing Article 53 of the EU AI Act).

195. For discussions of the complexities in the generative AI supply chain, see generally Katherine Lee, A. Feder Cooper & James Grimmelmann, Talkin' 'bout AI Generation: Copyright and the Generative-AI Supply Chain, 70 J. COPYRIGHT SOC'Y U.S.A. (forthcoming 2025); David Gray Widder & Dawn Nafus, Dislocated Accountabilities in the "AI Supply Chain": Modularity and Developers' Notions of Responsibility, 10 BIG DATA & SOC'Y 1 (2023).

196. See Ty, Inc. v. GMA Accessories, Inc., 132 F.3d 1167, 1170 (7th Cir. 1997) ("The parties' bean-bag pigs bear little resemblance to real pigs . . . . Real pigs are not the only pigs in the public domain. But [the defendant] has not pointed to any fictional pig in the public domain that [its bean-bag pig] resembles.").

197. See, e.g., Benjamin L.W. Sobel, Artificial Intelligence's Fair Use Crisis, 41 COLUM. J.L. & ARTS 45 (2017); supra Section II.B.

198. See U.S. DEP'T OF COM., INTERNET POL'Y TASK FORCE, COPYRIGHT POLICY, CREATIVITY, AND INNOVATION IN THE DIGITAL ECONOMY 21 (2013) ("[Fair use] is a fundamental linchpin of the U.S. copyright system."); see also David Nimmer, A Modest Proposal to Streamline Fair Use Determinations, 24 CARDOZO ARTS & ENT. L.J. 11, 11 (2006) ("[T]he safeguard of fair use constitutes a vital and indispensable part of [U.S.] copyright laws . . . ."); Pamela Samuelson, Unbundling Fair Uses, 77 FORDHAM L. REV. 2537, 2618 (2009) ("Fair use is an essential doctrine in U.S. copyright law that counterbalances what would otherwise be an unreasonably broad grant of rights to authors and an unduly narrow set of negotiated exceptions and limitations.").

199. See supra note 130 and accompanying text.

200. See, e.g., Hong Kong Copyright Ordinance, (1997) Cap. 528, §§ 38, 41A, 54A (H.K.) (incorporating the fairness factors); Hubbard v. Vosper, [1972] 2 Q.B. 84 (Eng.) (defining fair dealing by identifying factors that resemble those found in the U.S. fair use provision); see also Giuseppina D'Agostino, Healing Fair Dealing? A Comparative Copyright Analysis of Canada's Fair Dealing to U.K. Fair Dealing and U.S. Fair Use, 53 MCGILL L.J. 309, 342-43 (2008) (extracting from English copyright law the following fairness factors: nature of the work, how the work was obtained, amount taken, uses made, commercial benefit, motives for the dealing, consequences of the dealing, and purpose achieved by different means); Peter K. Yu, The Quest for a User-Friendly Copyright Regime in Hong Kong, 32 AM. U. INT'L L. REV. 283, 323 (2016) ("[B]ecause of the common law tradition in those Commonwealth jurisdictions embracing the fair dealing model, the use of fairness factors often emerge through case law even when those factors have not been written into the statutory provisions.").

201. 17 U.S.C. § 107.

202. Barton Beebe, An Empirical Study of U.S. Copyright Fair Use Opinions, 1978-2005, 156 U. PA. L. REV. 549, 586 (2008); see also Harper & Row, Publishers, Inc. v. Nation Enters., 471 U.S. 539, 566 (1985) (noting that the fourth factor "is undoubtedly the single most important element of fair use"); Tang Xiyin, Can Copyright Holders Do Harm to Their Own Works? A Reverse Theory of Fair Use Market Harm, 54 U.C. DAVIS L. REV. 1245, 1251 (2021) ("While it is often said that th[e] market harm factor is one of the most important in any fair use analysis, in the past decade, courts and commentators have dedicated much more energy to interpreting the first fair use factor, which looks to whether the infringing use is 'transformative.'" (footnotes omitted)).

203. See Campbell v. Acuff-Rose Music, Inc., 510 U.S. 569 (1994) (introducing the transformative use doctrine).

204. See Jane C. Ginsburg, Fair Use in the United States: Transformed, Deformed, Reformed?, 2020 SING. J. LEGAL STUD. 265, 266 ("[T]he 'transformative use' analysis has engulfed all of fair use, becoming transformed, and perhaps deformed, in the process."); Matthew Sag, Predicting Fair Use, 73 OHIO ST. L.J. 47, 84 (2012) ("[T]ransformative use by the defendant is a robust predictor of a finding of fair use."). Clark D. Asay, Arielle Sloan & Dean Sobczak, Is Transformative Use Eating the World?, 61 B.C. L. REV. 905 (2020) (providing an empirical analysis of the dominance of the transformative use doctrine).

205. See Google LLC v. Oracle Am., Inc., 593 U.S. 1, 40 (2021) ("[W]here Google reimplemented a user interface, taking only what was needed to allow users to put their accrued talents to work in a new and transformative program, Google's copying of the Sun Java [application programming interface] was a fair use of that material as a matter of law.").

206. See Andy Warhol Found. for the Visual Arts, Inc. v. Goldsmith, 598 U.S. 508, 532 (2023) ("[T]he first fair use factor considers whether the use of a copyrighted work has a further purpose or different character, which is a matter of degree, and the degree of difference must be balanced against the commercial nature of the use.").

207. Authors Guild, Inc. v. HathiTrust, 755 F.3d 87, 98 (2d Cir. 2014) (quoting Cariou v. Prince, 714 F.3d 694, 710 (2d Cir. 2013)) (internal quotation marks omitted).

208. Authors Guild v. Google, Inc., 804 F.3d 202, 220 (2d Cir. 2015); see also Asay et al., supra note 204, at 948-49 ("[F]actor two appears to play a quite limited role in influencing resolution of the other fair use factors. . . . [T]his result is not because factor two considerations are irrelevant, but rather because other factors within the fair use test already address those considerations more effectively.").

209. Sega Enters. Ltd. v. Accolade, Inc., 977 F.2d 1510, 1524 (9th Cir. 1992).

210. Id. at 1527.

211. MARSHALL A. LEAFFER, UNDERSTANDING COPYRIGHT LAW 585 (8th ed. 2024).

212. See Clark D. Asay, Independent Creation in a World of AI, 14 FIU L. REV. 201, 218-19 (2020) (noting the potentially differing second factor analysis in the fair use determination of machine-generated works).

213. Id. at 218.

214. See Sega, 977 F.2d at 1524, 1527 (finding the third factor favoring the plaintiffdue to the defendant's disassembly of the entirety of the protected computer programs).

215. See Lukas Selin, Demystifying Tokens in LLMs, TOKES COMPARE BLOG, https:// tokescompare.io/demystifying-tokens-in-llms (discussing the conversion to and use of tokens in large language models). For discussions of the AI training process, see generally Lee et al., supra note 195; Matthew Sag, Copyright Safety for Generative AI, 61 HOUS. L. REV. 295, 313-21 (2023).

216. See, e.g., Metro-Goldwyn-Mayer Studios Inc. v. Grokster, Ltd., 545 U.S. 913 (2005); In re Aimster Copyright Litig., 334 F.3d 643 (7th Cir. 2003); A&M Records, Inc. v. Napster, Inc., 239 F.3d 1004 (9th Cir. 2001).

217. See Michael D. Murray, Generative AI Art: Copyright Infringement and Fair Use, 26 SMU SCI. & TECH. L. REV. 259, 263 (2023) (arguing that the focus of the debate on copyright infringement in the AI context should be shifted "from the persons compiling the training dataset used to train the AI system and the designers and creators of the AI system itself to the end users of the AI system who conceive of and cause the creation of images").

218. ABC, Inc. v. Aereo, Inc., 573 U.S. 431 (2014) (citation omitted).

219. Perfect 10, Inc. v. Visa Int'l Serv. Ass'n (Perfect 10 v. Visa), 494 F.3d 788, 797 (9th Cir. 2007).

220. Perfect 10, Inc. v. Amazon.com, Inc., 487 F.3d 701, 729 (9th Cir. 2007).

221. Perfect 10 v. Visa, 494 F.3d at 797.

222. Sony Corp. of Am. v. Universal City Studios, Inc., 464 U.S. 417 (1984).

223. Id. at 442.

224. See Metro-Goldwyn-Mayer Studios Inc. v. Grokster, Ltd., 545 U.S. 913, 919 (2005) ("[O]ne who distributes a device with the object of promoting its use to infringe copyright, as shown by clear expression or other affirmative steps taken to foster infringement, is liable for the resulting acts of infringement by third parties.").

225. 17 U.S.C. § 512; see also Yu, Digital Copyright Reform, supra note 146, at 709-13 (discussing the safe harbors for online service providers).

226. See Viacom Int'l, Inc. v. YouTube, Inc., 676 F.3d 19, 30-35 (2d Cir. 2012) (exploring whether Section 512 of the 1976 Copyright Act "requires 'actual knowledge' or 'aware[ness]' of facts or circumstances indicating 'specific and identifiable infringements'").

227. See Lenz v. Universal Music Corp., 815 F.3d 1145, 1153 (9th Cir. 2016) ("[B]ecause 17 U.S.C. § 107 created a type of non-infringing use, fair use is 'authorized by the law' and a copyright holder must consider the existence of fair use before sending a takedown notification under § 512(c).").

228. See Ventura Content, Ltd. v. Motherless, Inc., 885 F.3d 597, 617 (9th Cir. 2018) (noting that "[v]arious factors may bear on whether a service provider has 'adopted and reasonably implemented' its policy for terminating, 'in appropriate circumstances,' repeat infringers" as required by Section 512(i) of the 1976 Copyright Act); see also Annemarie Bridy, Graduated Response American Style: "Six Strikes" Measured Against Five Norms, 23 FORDHAM INTELL. PROP. MEDIA & ENT. L.J. 1 (2012) (discussing the copyright alert system); Peter K. Yu, The Graduated Response, 62 FLA. L. REV. 1373 (2010) (discussing the graduated response system).

229. See DSM Directive, supra note 133, art. 17 (providing the "notice-and-stay-down" mechanism through the introduction of filtering requirements).

230. U.S. COPYRIGHT OFF., SECTION 512 OF TITLE 17: A REPORT OF THE REGISTER OF COPYRIGHTS (2020).

231. Lee et al., supra note 195, also explores the creation of a notice-and-takedown regime to address challenges AI technology has posed to the copyright system.

232. DSM Directive, supra note 133, art. 17.

233. See Graeme W. Austin, Importing Kazaa-Exporting Grokster, 22 SANTA CLARA HIGH TECH. L.J. 577, 592-608 (2006) (discussing the difficulty of exporting U.S. laws concerning indirect liability theories to other parts of the world).

234. See Nari Lee, Intellectual Property Law in China-From Legal Transplant to Governance, in GOVERNANCE OF INTELLECTUAL PROPERTY RIGHTS IN CHINA AND EUROPE 5, 9 (Nari Lee, Niklas Bruun & Li Mingde eds., 2016) ("[I]f the law is there to recognize a preexisting normative order, without localization, the laws that are introduced to a foreign culture may only be implemented successfully as a matter of an 'unusual and accidental coincidence' as noted by Montesquieu."); Yu, Customizing Fair Use Transplants, supra note 163, at 11 ("[R]egardless of whether a legal transplant is widely supported by the local populace or forced upon them from abroad, the transplanted law needs to be customized to local conditions if it is to be effective and if it is to receive wide public support."); Yu, Digital Copyright Reform, supra note 146, at 755 ("[L]ike the transplant of plants or human organs, the [legal transplantation] process requires a careful process of evaluation, selection, adaptation, and assimilation.").

235. See 17 U.S.C. §§ 401-406 (stipulating the notice requirements).

236. See id. §§ 408-412 (stipulating the registration requirements).

237. See id. § 407 (stipulating the deposit requirements).

238. See Estate of Martin Luther King, Jr., Inc. v. CBS, Inc. 194 F.3d 1211, 1214 (11th Cir. 1999) ("When a general publication occurred, the author either forfeited his work to the public domain . . . or, if he had therebefore complied with federal statutory requirements, converted his common law copyright into a federal statutory copyright.").

239. Christopher Sprigman, Reform(aliz)ing Copyright, 57 STAN. L. REV. 485, 531 (2004).

240. Berne Convention for the Protection of Literary and Artistic Works art. 5(2), Sept. 9, 1886, 828 U.N.T.S. 221 (last revised at Paris July 24, 1971).

241. Id.

242. The Berne Convention does not affect how the United States protects domestic copyrighted works. See id. art. 5(3) ("Protection in the country of origin is governed by domestic law.").

243. See Sprigman, supra note 239, at 487 ("Formalities . . . facilitated licensing by lowering the cost of identifying rightsholders, moved works for which copyright was not desired into the public domain, and encouraged the use of public domain works by lowering the cost of confirming that a work was available for use.").

244. Eldred v. Ashcroft, 537 U.S. 186 (2003).

245. See Jane C. Ginsburg, Copyright Legislation for the "Digital Millennium," 23 COLUM.- VLA J.L. & ARTS 137, 157 (1999) (noting that Section 1202 of the 1976 Copyright Act, which protects rights management information, "is designed to promote the dissemination of copyrighted works by facilitating the grant or license of rights under copyright (particularly through electronic contracting)").

246. WIPO Copyright Treaty, Dec. 20, 1996, 2186 U.N.T.S. 121 [hereinafter WCT].

247. WIPO Performances and Phonograms Treaty, Dec. 20, 1996, 2186 U.N.T.S. 203 [hereinafter WPPT].

248. WCT, supra note 246, art. 12(1).

249. Id. art. 12(2).

250. 17 U.S.C. § 1202.

251. See How Content ID Works, YOUTUBE HELP, https://support.google.com/youtube/ answer/2797370?hl=en (last visited Oct. 7, 2024) (providing an overview of YouTube's Content ID system). For discussions of the Content ID system, see generally Maayan Perel & Niva Elkin-Koren, Accountability in Algorithmic Copyright Enforcement, 19 STAN. TECH. L. REV. 473, 509-16 (2016); Matthew Sag, Internet Safe Harbors and the Transformation of Copyright Law, 93 NOTRE DAME L. REV. 499, 543-60 (2017).

252. How Content ID Works, supra note 251.

253. See Sun, supra note 23, at 1236-48.

254. See supra text accompanying notes 212-213.

255. See Hulianwang Xinxi Fuwu Shendu Hecheng Guanli Guiding (...) [Provisions on the Management of Deep Synthesis in Internet-Based Information Service] (promulgated by Cyberspace Admin. of China, Nov. 3, 2022, effective Jan. 10, 2023), arts. 16-18 (requiring the affixation of watermarks); see also Peter Henderson, Should the United States or the European Union Follow China's Lead and Require Watermarks for Generative AI?, GEO. J. INT'L AFFS. (May 24, 2023), https://gjia.georgetown.edu/2023/ 05/24/should-the-united-states-or-the-european-union-follow-chinas-lead-and-requirewatermarks- for-generative-ai (discussing this regulation). For discussions of Chinese technology regulation in the area of generative AI, see generally ANGELA HUYUE ZHANG, HIGH WIRE: HOW CHINA REGULATES BIG TECH AND GOVERNS ITS ECONOMY 277-91 (2024); ZENG JINGHAN, ARTIFICIAL INTELLIGENCE WITH CHINESE CHARACTERISTICS: NATIONAL STRATEGY, SECURITY AND AUTHORITARIAN GOVERNANCE (2022); Cheng Jing & Zeng Jinghan, Shaping AI's Future? China in Global AI Governance, 32 J. CONTEMP. CHINA 794 (2023); Angela Huyue Zhang, The Promise and Perils of China's Regulation of Artificial Intelligence, 63 COLUM. J. TRANSN'L L. (forthcoming 2025); Matt Sheehan, China's AI Regulations and How They Get Made, CARNEGIE ENDOWMENT FOR INT'L PEACE (July 20, 2023), https://carnegieendowment. org/research/2023/07/chinas-ai-regulations-and-how-they-get-made?lang=en; Paul Triolo & Kendra Schaefer, China's Generative AI Ecosystem in 2024: Rising Investment and Expectations, NAT'L BUREAU OF ASIAN RSCH. (June 27, 2024), https://www.nbr.org/publication/chinasgenerative- ai-ecosystem-in-2024-rising-investment-and-expectations.

256. EU AI Act, supra note 5, art. 53(1)(c).

257. DSM Directive, supra note 133, art. 4(3).

258. See Itar-Tass Russian News Agency v. Russian Kurier, Inc., 153 F.3d 82, 90 (2d Cir. 1998) (emphasizing the importance of the Berne Convention Implementation Act of 1988 when reviewing "the source of law for selecting a conflicts rule"); see also Peter K. Yu, The Comparative Lessons of Itar-Tass Russian News Agency v. Russian Kurier, in THE CAMBRIDGE HANDBOOK OF INTELLECTUAL PROPERTY IN CENTRAL AND EASTERN EUROPE 110, 116-17 (Mira T. Sundara Rajan ed., 2019) (discussing the legal effects of international treaties on U.S. soil).

259. WCT, supra note 246.

260. See, e.g., Digital Millennium Copyright Act of 1998, Pub. L. No. 105-304, 112 Stat. 2860 (codified as amended in scattered sections of 17 U.S.C.); Uruguay Round Agreements Act, Pub. L. No. 103-465 § 514, 108 Stat. 4809, 4976-81 (1994).

261. See PEDRO DOMINGOS, THE MASTER ALGORITHM: HOW THE QUEST FOR THE ULTIMATE LEARNING MACHINE WILL REMAKE OUR WORLD 6 (2015) ("Learning algorithms- also known as learners-are algorithms that make other algorithm

262. Copyright, Designs and Patents Act 1988, c. 48, § 9(3) (UK).

263. Id.

264. For discussions of deep learning, see generally ETHEM ALPAYDIN, MACHINE LEARNING: THE NEW AI 104-09 (2016); JOHN D. KELLEHER, DEEP LEARNING (2019); JOHN D. KELLEHER & BRENDAN TIERNEY, DATA SCIENCE 121-30 (2018); THIERRY POIBEAU, MACHINE TRANSLATION 181-95 (2017).

265. See Joshua A. Kroll, Joanna Huey, Solon Barocas, Edward W. Felten, Joel R. Reidenberg, David G. Robinson & Harlan Yu, Accountable Algorithms, 165 U. PA. L. REV. 633, 641 (2017) ("[W]ithout full transparency-including source code, input data, and the full operating environment of the software-even the disclosure of audit logs showing what a program did while it was running provides no guarantee that the disclosed information actually reflects a computer system's behavior."); see also id. at 657-60 (discussing the limits to transparency in the algorithmic context).

266. Kartik Hosanagar & Vivian Jair, We Need Transparency in Algorithms, but Too Much Can Backfire, HARV. BUS. REV. (July 23, 2018), https://hbr.org/2018/07/we-needtransparency- in-algorithms-but-too-much-can-backfire; see also Daniel Gervais, Exploring the Interfaces Between Big Data and Intellectual Property Law, 10 J. INTELL. PROP. INFO. TECH. & ELEC. COM. L. 3, 5 (2019) ("[A]ny human contribution to the output of deep learning systems is 'second degree.'").

267. Cf. IVIR AI STUDY, supra note 183, at 117 ("If 'off-the-shelf' AI systems are used to create content, co-authorship claims by AI developers will . . . be unlikely for commercial reasons, since AI developers will normally not want to burden customers with downstream copyright claims.").

268. See generally U.S. COPYRIGHT OFF., RESALE ROYALTIES: AN UPDATED ANALYSIS: A REPORT OF THE REGISTER OF COPYRIGHTS (2013) (discussing resale royalties).

269. See EDWARD LEE, CREATORS TAKE CONTROL: HOW NFTS REVOLUTIONIZE ART, BUSINESS, AND ENTERTAINMENT 137-44 (2023); Megan E. Noh, Sarah C. Odenkirk & Yayoi Shionoiri, GM! Time to Wake Up and Address Copyright and Other Legal Issues Impacting Visual Art NFTs, 45 COLUM. J.L. & ARTS 315, 327-29 (2022).

270. Peter K. Yu, Deploying Blockchain Technology in the Copyright Office, in NFTS, CREATIVITY AND THE LAW: WITHIN AND BEYOND COPYRIGHT 64, 78 (Enrico Bonadio & Caterina Sganga eds., 2024).

271. See Camille Brown, Coded Copyright?: How Copyright Enforcement, Remuneration, and Verification Terms in Blockchain-Enhanced Contract Models for Online Art Sales Compare to Their Traditional Counterparts, 31 S. CAL. INTERDISCIPLINARY L.J. 617, 638-47 (2022) (providing case studies that show the deployment by Christie's, DADA.nyc, and OpenSea of blockchain-enhanced contract models).

272. U.S. PAT. & TRADEMARK OFF., supra note 3, at 22.

273. In an earlier book chapter, the Author advocated the use of these four reform pathways to address the global digital distribution of media and entertainment content. See generally Peter K. Yu, A Seamless Global Digital Marketplace of Media and Entertainment Content, in RESEARCH HANDBOOK ON INTELLECTUAL PROPERTY IN MEDIA AND ENTERTAINMENT 265, 278-89 (Megan Richardson & Sam Ricketson eds., 2017) [hereinafter Yu, Seamless Global Digital Marketplace] (discussing these reform pathways). This section updates this earlier discussion and adapts it to the AI context.

274. See Colin B. Picker, A View from 40,000 Feet: International Law and the Invisible Hand of Technology, 23 CARDOZO L. REV. 149, 184 (2001) ("[D]elay is the rule in the formation of international law. Usually, international law is created over long periods, by the gradual acceptance of customary state practice or after long treaty negotiations."); Yu, Seamless Global Digital Marketplace, supra note 273, at 281-82 (discussing the lengthy treaty process and how the resulting treaties tend to lag behind technological developments); Peter K. Yu, Trade Agreement Cats and the Digital Technology Mouse, in SCIENCE AND TECHNOLOGY IN INTERNATIONAL ECONOMIC LAW: BALANCING COMPETING INTERESTS 185, 202 (Bryan Mercurio & Ni Kuei-Jung eds., 2014) [hereinafter Yu, Trade Agreement Cats] ("[F]rom initial negotiation to final ratification to full implementation, it takes a considerable amount of time, effort, energy, and resources to complete a trade agreement. The rate at which such an agreement is developed can hardly keep pace with the rate of technological change.").

275. TRIPS Agreement, supra note 33.

276. WCT, supra note 246.

277. WPPT, supra note 247.

278. Beijing Treaty on Audiovisual Performances, June 23, 2012, 51 I.L.M. 1214.

279. Marrakesh Treaty to Facilitate Access to Published Works for Persons Who Are Blind, Visually Impaired or Otherwise Print, June 27, 2013, 52 I.L.M. 1312.

280. Geneva Act of the Lisbon Agreement for the Protection of Appellations of Origin and Their International Registration, Oct. 31, 1958 (last revised at Geneva May 20, 2015), https://www.wipo.int/wipolex/en/text/370297.

281. WIPO Treaty on Intellectual Property, Genetic Resources and Associated Traditional Knowledge, WIPO Doc. GRATK/DC/7 (May 24, 2024).

282. Press Release, WIPO, WIPO Member States Adopt Riyadh Design Law Treaty (Nov. 22, 2024), https://www.wipo.int/pressroom/en/articles/2024/article_0017.html.

283. The Intergovernmental Committee was established at the 2000 WIPO General Assembly. Press Release, WIPO, WIPO General Assemblies Wrap Up (Oct. 3, 2000), https://www.wipo.int/pressroom/en/prdocs/2000/wipo_pr_2000_243.html; see also Peter K. Yu, WIPO Negotiations on Intellectual Property, Genetic Resources and Associated Traditional Knowledge, 57 AKRON L. REV. (forthcoming 2024) (discussing the origin of the negotiations surrounding the GRATK Treaty). See generally PROTECTING TRADITIONAL KNOWLEDGE: THE WIPO INTERGOVERNMENTAL COMMITTEE ON INTELLECTUAL PROPERTY AND GENETIC RESOURCES, TRADITIONAL KNOWLEDGE AND FOLKLORE (Daniel F. Robinson, Ahmed Abdel-Latif & Pedro Roffe eds., 2017) [hereinafter PROTECTING TRADITIONAL KNOWLEDGE] (collecting articles that offer detailed analyses of the Intergovernmental Committee's effort); Symposium, Traditional Knowledge, Intellectual Property, and Indigenous Culture, 11 CARDOZO J. INT'L & COMPAR. L. 239 (2003) (collecting articles from the first academic symposium on traditional knowledge and traditional cultural expressions in a U.S. law school).

284. See generally Yu, Trade Agreement Cats, supra note 274 (discussing this cat-andmouse chase).

285. Marci A. Hamilton, The TRIPS Agreement: Imperialistic, Outdated, and Overprotective, 29 VAND. J. TRANSNAT'L L. 613, 614-15 (1996).

286. J.H. Reichman, The Know-How Gap in the TRIPS Agreement: Why Software Fared Badly, and What Are the Solutions, 17 HASTINGS COMM. & ENT. L.J. 763, 766 (1995) (footnote omitted). But see Patricia L. Judd, The TRIPS Balloon Effect, 46 N.Y.U. J. INT'L L. & POL. 471, 527 (2014) ("No treaty, large or small, bilateral or multilateral, regional or multinational, can hope to keep up with recent and ongoing technological changes. What TRIPS does have is a malleability that can aid it in keeping up with the times. It does not need specific Internetoriented provisions to be relevant in an Internet age." (footnote omitted)).

287. See Yu, International Enclosure Movement, supra note 76, at 858-62.

288. WCT, supra note 246.

289. See, e.g., WIPO, Joint Recommendation Concerning Provisions on the Protection of Marks, and Other Industrial Property Rights in Signs, on the Internet, WIPO Doc. 845(E) (Oct. 2001); WIPO, Joint Recommendation Concerning Provisions on the Protection of Well-Known Marks, WIPO Doc. 833(E) (Sept. 1999) [hereinafter Joint Recommendation on Well-Known Marks]; WIPO, Intergovernmental Comm. on Intell. Prop. & Genetic Resources, Traditional Knowledge and Folklore [IGC], Joint Recommendation on Genetic Resources and Associated Traditional Knowledge, U.N. Doc. WIPO/GRTKF/IC/47/21 (May 26, 2023) (providing the proposal from Japan, South Korea, and the United States); IGC, Joint Recommendation on the Use of Databases for the Defensive Protection of Genetic Resources and Traditional Knowledge Associated with Genetic Resources, U.N. Doc. WIPO/GRTKF/IC/47/17 (May 17, 2023) (providing the proposal). See generally Gary E. Marchant & Carlos Ignacio Gutierrez, SoftLaw 2.0: An Agile and Effective Governance Approach for Artificial Intelligence, 24 MINN. J.L. SCI. & TECH. 375 (2023) (discussing the use of softlaw in the AI context). For discussions of softlaw, see generally COMMITMENT AND COMPLIANCE: THE ROLE OF NON-BINDING NORMS IN THE INTERNATIONAL LEGAL SYSTEM (Dinah Shelton ed., 2000); RESEARCH HANDBOOK ON SOFT LAW (Mariolina Eliantonio, Emilia Korkea-Aho & Ulrika Mörth eds., 2023); Kenneth W. Abbott & Duncan Snidal, Hard and SoftLaw in International Governance, 54 INT'L ORG. 421 (2000); Margot E. Kaminski & Shlomit Yanisky-Ravid, The Marrakesh Treaty for Visually Impaired Persons: Why a Treaty Was Preferable to SoftLaw, 75 U. PITT. L. REV. 255 (2014); Gregory C. Shaffer & Mark A. Pollack, Hard vs. SoftLaw: Alternatives, Complements, and Antagonists in International Governance, 94 MINN. L. REV. 706 (2010).

290. As a WIPO document explains:

Model laws or guidelines are a standardized set of legal provisions that can serve as a reference for national legislation. They are meant to provide guidance for states in developing their own laws or regulations. They can help ensure a consistent level of protection across countries and promote harmonization of [intellectual property] laws at the international level. They are not legally binding and each country remains free to adopt or adjust provisions in a manner that suits them.

IGC, Legal Principles Related to an International Instrument, at 3, U.N. Doc. WIPO/GRTKF/ IC/47/12 (Apr. 28, 2023); see also IGC, The Protection of Traditional Knowledge: Updated DraftGap Analysis, U.N. Doc. WIPO/GRTKF/IC/47/8 (Mar. 21, 2023) (discussing the use of softlaw recommendations, model laws, and other guiding documents in the TK context).

291. See, e.g., WIPO, DOCUMENTING TRADITIONAL KNOWLEDGE-A TOOLKIT (2017) (providing a toolkit for documenting TK).

292. Joint Recommendation on Well-Known Marks, supra note 289.

293. Id. at 3.

294. See, e.g., Regional Comprehensive Economic Partnership Agreement art. 11.26.2, Nov. 15, 2020, https://asean.org/wp-content/uploads/2021/04/All-Chapters.pdf; Trans- Pacific Partnership Agreement art. 18.22.3, Feb. 4, 2016, https://ustr.gov/trade-agreements/ free-trade-agreements/trans-pacific-partnership/tpp-full-text; United States-Singapore FTA, supra note 97, art. 16.1.2(b).

295. For discussions of these public health crises, see generally NEGOTIATING HEALTH: INTELLECTUAL PROPERTY AND ACCESS TO MEDICINES (Pedro Roffe, GeoffTansey & David Vivas-Eugui eds., 2006); SUSAN K. SELL, PRIVATE POWER, PUBLIC LAW: THE GLOBALIZATION OF INTELLECTUAL PROPERTY RIGHTS 146-62 (2003); Debora Halbert, Moralized Discourses: South Africa's Intellectual Property Fight for Access to AIDS Drugs, 1 SEATTLE J. SOC. JUST. 257 (2002); Ellen 't Hoen, TRIPS, Pharmaceutical Patents, and Access to Essential Medicines: A Long Way from Seattle to Doha, 3 CHI. J. INT'L L. 27 (2002).

296. World Trade Organization, Declaration on the TRIPS Agreement and Public Health of 14 November 2001, WTO Doc. WT/MIN(01)/DEC/2, 41 I.L.M. 746 (2002).

297. Id. ¶ 6.

298. See Yu, International Enclosure Movement, supra note 76, at 872-86 (discussing the negotiation of Article 31bis of the TRIPS Agreement).

299. TRIPS Agreement, supra note 33, arts. 31(f), 31bis.

300. Press Release, World Trade Org., WTO IP Rules Amended to Ease Poor Countries' Access to Affordable Medicines (Jan. 23, 2017), https://www.wto.org/english/news_e/ news17_e/trip_23jan17_e.htm.

301. See Dominic Keating, The WIPO IGC: A U.S. Perspective, in PROTECTING TRADITIONAL KNOWLEDGE, supra note 283, at 265, 273 ("Softlaw is not legally binding. However, it can be a valuable tool for solving international problems. It can also be a platform for future work.").

302. See Yu, Future Path, supra note 52.

303. Peter K. Yu, Remarks at Online Webinar on "International Copyright Issues and Artificial Intelligence," U.S. COPYRIGHT OFF. (July 26, 2023), https://streammedia. loc.gov/copyright/International-Copyright-Issues-and-Artificial-Intelligence.mp4.

304. World Summit on the Information Society, Tunis Agenda for the Information Society 2005, ¶ 73, U.N. Doc. WSIS-05/TUNIS/DOC/6(Rev 1)-E (Nov. 18, 2005).

305. Id.

306. See Peter K. Yu, A Tale of Two Development Agendas, 35 OHIO N.U. L. REV. 465, 537- 38 (2009) (discussing the World Summit on the Information Society and the creation of the Internet Governance Forum); see also JEREMY MALCOLM, MULTI-STAKEHOLDER GOVERNANCE AND THE INTERNET GOVERNANCE FORUM 415-521 (2008) (providing an early assessment of this forum).

307. See Monika Ermert, Internet Governance Forum: Ten Years After, INTELL. PROP. WATCH (Nov. 16, 2015), https://web.archive.org/web/20170225162628/http://www.ipwatch. org/2015/11/16/internet-governance-forum-ten-years-after.

308. Francis Gurry, 2013 Address by the Director General, WORLD INTELL. PROP. ORG. (Sept. 23, 2013), https://www.wipo.int/about-wipo/en/dg_gurry/speeches/a_51_dg_ speech.html.

309. Catherine Saez, WIPO Director Gurry Speaks on Naming New Cabinet, Future of WIPO, INTELL. PROP. WATCH (May 8, 2014, 11:48 PM), https://web.archive.org/web/ 20140802160306/www.ip-watch.org/2014/05/08/wipo-director-gurry-speaks-on-namingnew- cabinet-future-of-wipo.

310. Itar-Tass Russian News Agency v. Russian Kurier, Inc., 153 F.3d 82 (2d Cir. 1998).

311. See id. at 90 (discussing the source of applicable conflict-of-law rules).

312. Id.

313. See id. at 90-91.

314. See id.

315. Avril D. Haines, The Impact of the Internet on the Judgments Project: Thoughts for the Future, HAGUE CONF. ON PRIV. INT'L L. (Feb. 2002), www.hcch.net/upload/wop/gen_ pd17e.pdf.

316. INTELLECTUAL PROPERTY: PRINCIPLES GOVERNING JURISDICTION, CHOICE OF LAW, AND JUDGMENTS IN TRANSNATIONAL DISPUTES (AM. L. INST. 2008); see also Symposium on Constructing International Intellectual Property Law: The Role of National Courts, 77 CHI.-KENT L. REV. 991 (2002) (discussing the draftConvention and the proposal advanced by Rochelle Dreyfuss and Jane Ginsburg that eventually became the draftAmerican Law Institute Principles).

317. See generally EUR. MAX PLANCK GP. ON CONFLICT OF L. IN INTELL. PROP., CONFLICT OF LAWS IN INTELLECTUAL PROPERTY: THE CLIP PRINCIPLES AND COMMENTARY (2013) (discussing the CLIP Principles).

318. INT'L L. ASS'N COMM. ON INTELL. PROP. & PRIV. INT'L L., GUIDELINES ON INTELLECTUAL PROPERTY AND PRIVATE INTERNATIONAL LAW (2020), reprinted in 12 J. INTELL. PROP. INFO. TECH. & ELEC. COM. L. 84 (2021). The Author was a member of the International Law Association Committee on Intellectual Property and Private International Law.

319. Id. Guideline 1(1).

320. See Marketa Trimble, Advancing National Intellectual Property Policies in a Transnational Context, 74 MD. L. REV. 203, 207 n.10 (2015) (listing the Transparency of Japanese Law Project's Transparency Proposal on Jurisdiction, Choice of Law, Recognition and Enforcement of Foreign Judgments in Intellectual Property in 2009 and the Korean Private International Law Association's Principles on International Intellectual Property Litigation in 2010).

321. Paul SchiffBerman, Towards a Cosmopolitan Vision of Conflict of Laws: Redefining Governmental Interests in a Global Era, 153 U. PA. L. REV. 1819, 1865 (2005).

322. Graeme B. Dinwoodie, A New Copyright Order: Why National Courts Should Create Global Norms, 149 U. PA. L. REV. 469, 476 (2000).

323. Id.

324. See discussion supra Section II.A.

325. See Sag & Yu, supra note 20 (noting that interest group politics, legislative inertia, and the legitimate interests of countervailing constituencies may induce countries to strike compromises by taking the middle path); Yu, Paradigm Evolution, supra note 130, at 148-55 (discussing the impact of interest group politics and legislative inertia on fair use transplants).

326. See discussion supra Part II.

327. See Yu, Paradigm Evolution, supra note 130, at 128-41 (noting that countries have engaged more in "paradigm evolution" than "paradigm shifts" in their efforts to transplant the U.S. fair use provision).

328. See Yu, Currents and Crosscurrents, supra note 34, at 429-435 (discussing the "spectrum of options" in the harmonization process).

329. See, e.g., Rochelle Cooper Dreyfuss, Harmonization: Top Down, Bottom Up-and Now Sideways? The Impact of the IP Provisions of Megaregional Agreements on Third Party States, in MEGAREGULATION CONTESTED: GLOBAL ECONOMIC ORDERING AFTER TPP 345, 345 (Benedict Kingsbury, David M. Malone, Paul Mertenskötter, Richard B. Stewart, Thomas Streinz & Atsushi Sunami eds., 2019) (advancing the concept of "sideways" harmonization at the intersection of international trade and intellectual property); Laurence R. Helfer, Regime Shifting: The TRIPS Agreement and New Dynamics of International Intellectual Property Lawmaking, 29 YALE J. INT'L L. 1, 60 (2004) (noting "either upward or downward" crossborder harmonization of intellectual property protection standards); Sarah R. Wasserman Rajec, The Harmonization Myth in International Intellectual Property Law, 62 ARIZ. L. REV. 735, 782 (2020) (noting "an upward . . . or downward departure from the 'standard' level of [intellectual property] protection"); see also Sag & Yu, supra note 20 (noting the difficulty in determining in the intellectual property context whether a regulatory race is to the top or the bottom).

330. See International Copyright Issues and Artificial Intelligence, supra note 19 (exploring this question).

331. See Yu, Future Path, supra note 52. The term "crossvergence" refers to "a simultaneous convergence and divergence of regulatory standards, which is the result of the continuous and dynamic interactions between convergence and divergence forces." Peter K. Yu, TPP, RCEP, and the Crossvergence of Asian Intellectual Property Standards, in GOVERNING SCIENCE AND TECHNOLOGY UNDER THE INTERNATIONAL ECONOMIC ORDER: REGULATORY DIVERGENCE AND CONVERGENCE IN THE AGE OF MEGAREGIONALS 277, 292 (Peng Shin-Yi, Liu Han-Wei & Lin Ching-Fu eds., 2018). The term is derived from a concept advanced in international management and other disciplines. See, e.g., David A. Ralston, The Crossvergence Perspective: Reflections and Projections, 39 J. INT'L BUS. STUD. 27 (2008). Other commentators have also used a similar formulation. See, e.g., Wang Heng, Divergence, Convergence or Crossvergence of Chinese and U.S. Approaches to Regional Integration: Evolving Trajectories and Their Implications, 10 TSINGHUA CHINA L. REV. 149 (2018).