Search the archive.

Pillar 18 (Paul, is Non-resident Senior Fellow at the Center for Security Studies of Georgetown University and an Associate Fellow of the Geneva Center for Security Policy. He retired in 2005 from a 28-year career in the U.S. intelligence community. His senior positions included National Intelligence Officer for the Near East and South Asia, Deputy Chief of the DCI Counterterrorist Center, and Executive Assistant to the Director of Central Intelligence. He is a Vietnam War veteran and a retired officer in the U.S. Army Reserve. Dr. Pillar's degrees are from Dartmouth College, Oxford University, and Princeton University. His books include Negotiating Peace (1983), Terrorism and U.S. Foreign Policy (2001), Intelligence and U.S. Foreign Policy (2011), and Why America Misunderstands the World (2016).

To understand what is going on, use as a reference point the National Security Strategy (NSS) that the Trump administration released late last year (while bearing in mind that this type of document, in any administration, is less a true strategy than an expression of attitudes and objectives). The dominant theme of Trump’s NSS, articulated in its opening section, is that the world is overwhelmingly a competitive place in which America’s adversaries are “determined” to do unfriendly things. Russia and China are described as “determined” to “grow their militaries.” Other adversaries are “determined” to threaten Americans through other harmful actions. The NSS explicitly rejects engagement with adversaries, but at least as significant is what the document does not say. It gives not the slightest hint that any of the apparently unfriendly actions by adversaries could in any way be responses to actions or policies of the United States. Instead, any untoward or unwelcome action by an adversary is assumed to have occurred because the adversary was already “determined” to do it. The NSS declares the unipolar objective of “rebuilding our military so that it remains preeminent” and able to “fight and win” anywhere in the world. But it refuses to recognize any way in which even a dominant single pole would influence the thoughts and actions of other players, other than to awe, deter, or defeat them. The administration’s public documents that are subsidiary to the NSS follow the same principal line, although with some wavering. The more recently released National Defense Strategy (NDS) does briefly mention offering “competitors and adversaries an outstretched hand, open to opportunities for cooperation.” But the only way in which the NDS refers to U.S. policies affecting the policies and actions of others is in declaring, “The willingness of rivals to abandon aggression will depend on their perception of U.S. strength and the vitality of our alliances and partnerships.” Its characterization of the thinking and motivations of adversaries is illustrated by its summary comment on China: “As China continues its economic and military ascendance, asserting power through an all-of-nation long-term strategy, it will continue to pursue a military modernization program that seeks Indo-Pacific regional hegemony in the near-term and displacement of the United States to achieve global preeminence in the future.” In other words, it is again all about determination on the part of the other guy. The also recent Nuclear Posture Review (NPR), which gained public attention for its advocacy of the development of new “low yield” U.S. nuclear weapons, sounds downright pacifist in places in comparison with the NSS. It says good things about arms control, accepts as an ultimate goal the abolition of nuclear weapons, and states, “The United States does not wish to regard either Russia or China as an adversary and seeks stable relations with both.” But as with the other Trump administration documents, there is no explicit recognition that what the United States does, including building up its nuclear forces, can affect the decisions of other states, other than to deter them from doing the bad things they supposedly otherwise would want to do. Ingredients of U.S. Policy These documents are products of a process that involves political leadership as well as bureaucracies, and some of this mixture shows through in the results. It is not surprising that the most unexpurgated Trumpism appears in the White House-centric NSS, while the NDS and NPR contain hints of more careful thinking among professionals in the Department of Defense and elsewhere in the national security bureaucracy, some of whom no doubt have thought seriously about the variety of ways in which other states react to U.S. policies. These documents in part reflect the long-established thinking of a Washington national security establishment, a.k.a. The Blob, one of whose tenets is the maintenance of worldwide U.S. hard-power superiority. But a total disregard for how others might respond assertively rather than submissively to what the United States does is more distinctively Trumpian. The same blithe disregard for such responses is glaringly displayed in Trump’s declaration of a trade war. With a new Cold War developing with Russia, and perhaps another one with China, it would once again behoove U.S. policymakers to acquaint themselves with some of the doctrine dating from the earlier Cold War that stemmed from long study by political scientists and strategists. The Trump administration’s posture ignores, for instance, two phenomena in particular, both of which are disadvantageous to U.S. interests when they occur. One is the security dilemma, in which what one state does for defensive reasons, another state perceives as offensive and threatening—and is likely to respond in kind. This situation is increasingly applicable to much of what the United States is doing with its hard power, given that it is the most powerful player in the world and that it plays at least as much in other people’s neighborhoods as it does in its own. Security dilemmas can arise not only vis-à-vis Russia and China but also with other states that the United States has defined as adversaries. The other phenomenon is the arms race: responding to an adversary’s military build-up with a build-up of one’s own in an attempt to restore the previous balance. And then the adversary builds up some more to try to regain its version of a balance—and so on. Putin’s show-and-tell was part of an arms race already under way. Russia explicitly cites the George W. Bush administration’s withdrawal from the Anti-Ballistic Missile (ABM) treaty as a reason for developing the kind of defense-circumventing advanced strategic weapons of which Putin spoke. His government is not making such efforts just because it is “determined” to do so, even when taking into account Putin’s domestic political calculus. The prospects for similar security dilemmas and arms races, at the conventional as well as nuclear level, between the United States and China are considerable. And insofar as domestic sentiment in China may influence the actions of Chinese leaders during an international crisis, those leaders are less, not more, likely to back down in the face of a military threat from the United States. Effects of Multipolarity The multipolarity of today’s world exacerbates security dilemmas and arms races in ways that extend beyond great powers to middle-level ones. The Bush administration explained its withdrawal from the ABM treaty in terms of a need to defend against “rogue” states. The United States explained the later installation of missile defense systems in Eastern Europe in terms of protecting Europe from Iranian missiles, but Russia saw it as a threat to itself and as an effort to negate the capability of Russian strategic weapons. Those who remember the original Cold War may think longingly of how Ronald Reagan declared an arms race with the USSR and “won” in the sense of what happened a few years later to the USSR. But that is a wrong lesson to draw, because similar events are not about to recur. Russians already have lived through their history of the break-up of the union, the disarray of the Yeltsin years, and Putin’s skillful expression and exploitation of the country’s nationalist vibe. Putin isn’t going anywhere soon. Neither, evidently, is Xi Jinping in China, which shows no sign of becoming like the Soviet Union in 1991. By disregarding the rest of the world’s responses to its own policies, the Trump administration is on course to keep adversaries adversarial, if not more so, and to keep the world just as threatening as ever, if not more so. The military spending may become higher and the bluster may become louder, but Americans will be no safer.

In pursuit of global domination or a relative improvement in their status, major powers including the USA, China and Russia, have been ramping up their investments on AI R&D to arrive at major breakthroughs. Several scholars argue that these nation-states are engaged in a ‘global AI race’ to utilise the technology to boost their economic productivity as well as military effectiveness to get ahead of the rest. For the USA, the prevailing superpower of the international system, AI offers the means to preserve or strengthen their existing position. In other words, AI can arguably compensate for the dent in its international perception due to a perceived decline of power following its misadventures in Afghanistan and the weaning away of economic dominance to China. Meanwhile, the latter seemingly views AI as a shot in the arm that can boost its national power to either supplant the USA or at least match its might to serve its own ambition to emerge as a global superpower in the next few decades. Russia’s perspective partially reflects the Chinese view with the country looking to employ AI in combination with asymmetric cyber tactics and space weapon technologies to enhance its existing military capabilities to gain an edge over the USA. Through AI, Russia hopes to get the USA to the negotiating table through which the country hopes to regain its past glory. The evolving trajectory of the global AI race and its manifestations on relative power capabilities can have a significant impact on the emerging world order. This study has drawn from historical experience to assess whether major powers in the international system today, the USA, China and Russia can harness the opportunities offered by AI to enhance their power status as well as goals envisioned under national interest. Each of these players have variations in their approaches as well as differing abilities as organisations to absorb the technology into the fundamental ways in which they operate. However, possession of advanced capabilities does not by itself guarantee power as an outcome or the elevation and subsequent sustainment of a superior power status. In the oft-quoted words of Russian President Vladimir Putin “AI comes with colossal opportunities, but also threats that are difficult to predict. Whoever becomes the leader in this sphere will become the ruler of the world.” ( Vincent, 2017). As evidenced by the history of technological change, a country can only rise to a superior power status when its institutions operate in synergy to tackle challenges from technological disruption, while utilising the opportunities. Such policy synergy has to be guided clearly from leadership, which also has to show willingness to change existing policies as required, to facilitate the adoption of AI ( Vijayakumar, 2023; Milner and Solstad, 2021). The nature of AI as an unique and novel kind of sociotechnical system including artificial human-like agents poses challenges to projecting its exact impact on the global order. A large part of whether the USA, China or Russia forge ahead and the ways in which the rise or fall these powers might come about is also contingent on whether the technology evolves from ANI to AGI or even ASI. Evolution to both AGI and/or ASI is likely to yield different projections given the geopolitical circumstances of the time. Such a breakthrough shall fundamentally alter the nature of the world order and of power as it is currently understood in IR. If economics is to shape the contours of the competition, China is likely to achieve its superpower ambitions and emerge as an even more powerful player in international affairs. The USA holds a unique advantage in terms of the wealth of technological knowledge it possesses as well as through its private ecosystem which is churning out AI innovations at an ever-increasing rate. The USA also has been consistently attracting the smartest minds from around the world, which can fuel its fast-paced development. The world system will likely continue in its multipolar configuration in this case with commercial AI diffusing fast to enable multiple poles of power to strengthen their position. If military innovation and its impact on war decides the outcome of the AI race, the USA and China are well-positioned to emerge on top. While the USA still has a substantial lead in capabilities and fighting experience compared to the Chinese, AI might slightly change the equation in China’s favour. China’s authoritarian system of governance, access to data, and smart policies such a civil-military fusion may further put the country at an advantage. In this regard, a notable trend within the current multipolar environment points to the evolution of nascent technology-based power blocs, one led by the Sino-Russian alliance and the other represented by the USA and its alliance. Growing strategic competition among the two power blocs could lead to cause the world order to move to a bipolar configuration in the coming decades. Similar to how World War II terminated Germany’s rise trajectory, the power potential of any of the contenders discussed in the article may be constrained by the outbreak of a major war that they involve themselves in. Ultimately, the winner(s) of the AI race may be those who adopt a prudent approach to the technology and make calculated moves towards using this unique AI moment to rise to the top.

According to US Commerce Secretary Gina Raimondo, the United States and European Union (EU) don’t have a moment to wait in setting standards for the development and use of artificial intelligence (AI). “If the US and EU don’t show up,” she warned, “China will, [and] autocracies will.” Raimondo spoke at an Atlantic Council Front Page event on Tuesday alongside European Commission Executive Vice President Margrethe Vestager, who cautioned that the field of standardization in technologies is already being “dominated by nonmarket players or Chinese players.” But “we need to be much more present in standardization for us,” she said. “We need to have a presence.” The leaders spoke shortly after the fifth meeting of the EU-US Trade and Technology Council (TTC) in Washington, where officials touched on everything from AI to climate policy to semiconductors. In the US-EU relationship, “there are irritants for sure,” Raimondo admitted, “but fundamentally what binds us is massively more consequential than the irritants.” Below are more highlights from the conversation, which was moderated by Atlantic Council President and Chief Executive Officer Frederick Kempe. EU+US on AI Raimondo argued that the TTC will “prove to be exceedingly valuable” as AI tools continue to evolve. She said that the “muscle” the TTC has “built up”—in generating trust between private-sector stakeholders and the governments leading the United States and EU—will help in “bringing us together to write the rules of the road of AI.” But while Raimondo said that a “transatlantic approach” to AI could possibly come out of the TTC, she said she doubted whether “joint regulation” is feasible. “It will be some time before the US Congress passes a law that relates to the governing of AI,” the commerce secretary pointed out. “In the absence of that, there’s an awful lot of work to be done.” The commerce secretary explained that normally when governments develop regulations, each country first goes about writing rules separately before gathering with others to harmonize. “With AI, we can harmonize from the get-go because we haven’t yet written these regulations or rules or standards,” she said. In response to concerns that governments will always be regulating too slowly to keep up with tech, Vestager said that she thinks “that is just plain wrong” and that it is the responsibility of governments to ensure that technology respects society’s values. Keep “eyes wide open” on China Vestager outlined the EU’s “very complex relationship” with China, given that it is an important partner in fighting climate change—but it is also a systemic rival and an economic competitor. She explained that the EU is working to “derisk [its] dependencies” on countries such as China by getting more countries to screen foreign direct investments and working to prevent countries from skirting export controls. Regarding China, it is in the “self-interest” of the United States and EU “to work together,” Raimondo argued. “There are real national security concerns for both of us,” she added, “and we have to be eyes wide open about that and work together to protect… our countries.”

“In the very short run, the compromise on the EU AI Act won’t have much direct effect on established AI designers based in the US, because, by its terms, it probably won’t take effect until 2025,” said Paul Barrett back, deputy director of the NYU Stern Center for Business and Human Rights, back in December 2023 when the EU provisionally agreed on the landmark AI regulation. So for now, Barrett says major AI players like OpenAI, Microsoft, Google, and Meta will likely continue to fight for dominance, particularly as they navigate regulatory uncertainty in the US. The AI Act got its start before the explosion in general-purpose AI (GPAI) tools like OpenAI’s GPT-4 large language model, and regulating them became a remarkably complicated sticking point in last-minute discussions. The act divides its rules on the level of risk an AI system has on society, or as the EU said in a statement, “the higher the risk, the stricter the rules.” But some member states grew concerned that this strictness could make the EU an unattractive market for AI. France, Germany, and Italy all lobbied to water down restrictions on GPAI during negotiations. They won compromises, including limiting what can be considered “high-risk” systems, which would then be subject to some of the strictest rules. Instead of classifying all GPAI as high-risk, there will be a two-tier system and law enforcement exceptions for outright prohibited uses of AI like remote biometric identification. That still hasn’t satisfied all critics. French President Emmanuel Macron attacked the rules, saying the AI Act creates a tough regulatory environment that hampers innovation. Barrett said some new European AI companies could find it challenging to raise capital with the current rules, which gives an advantage to American companies. Companies outside of Europe may even choose to avoid setting up shop in the region or block access to platforms so they don’t get fined for breaking the rules — a potential risk Europe has faced in the non-AI tech industry as well, following regulations like the Digital Markets Act and Digital Services Act. AI models trained on publicly available — but sensitive and potentially copyrighted — data have become a big point of contention for organizations, for instance. The approved rules, however, do not create new laws around data collection. While the EU pioneered data protection laws through GDPR, its AI rules do not prohibit companies from gathering information, beyond requiring that it follow GDPR guidelines. “Under the rules, companies may have to provide a transparency summary or data nutrition labels,” Susan Ariel Aaronson, director of the Digital Trade and Data Governance Hub and a research professor of international affairs at George Washington University said when the EU provisionally approved the rules. “But it’s not really going to change the behavior of companies around data.” Aaronson points out that the AI Act still hasn’t clarified how companies should treat copyrighted material that’s part of model training data, beyond stating that developers should follow existing copyright laws (which leave lots of gray areas around AI). So it offers no incentive for AI model developers to avoid using copyrighted data. The AI Act also won’t apply its potentially stiff fines to open-source developers, researchers, and smaller companies working further down the value chain — a decision that’s been lauded by open-source developers in the field. GitHub chief legal officer Shelley McKinley said it is “a positive development for open innovation and developers working to help solve some of society’s most pressing problems.” (GitHub, a popular open-source development hub, is a subsidiary of Microsoft.) Observers think the most concrete impact could be pressuring other political figures, particularly American policymakers, to move faster. It’s not the first major regulatory framework for AI — in July, China passed guidelines for businesses that want to sell AI services to the public. But the EU’s relatively transparent and heavily debated development process has given the AI industry a sense of what to expect. Aaronson said the provisional text (which has since been approved) at least shows that the EU has listened and responded to public concerns around the technology. Lothar Determann, data privacy and information technology partner at law firm Baker McKenzie, says the fact that it builds on existing data rules could also encourage governments to take stock of what regulations they have in place. And Blake Brannon, chief strategy officer at data privacy platform OneTrust, said more mature AI companies set up privacy protection guidelines in compliance with laws like GDPR and in anticipation of stricter policies. He said that depending on the company, the AI Act is “an additional sprinkle” to strategies already in place. The US, by contrast, has largely failed to get AI regulation off the ground — despite being home to major players like Meta, Amazon, Adobe, Google, Nvidia, and OpenAI. Its biggest move so far has been a Biden administration executive order directing government agencies to develop safety standards and build on voluntary, non-binding agreements signed by large AI players. The few bills introduced in the Senate have mostly revolved around deepfakes and watermarking, and the closed-door AI forums held by Sen. Chuck Schumer (D-NY) have offered little clarity on the government’s direction in governing the technology. This doesn’t mean the US will take the same risk-based approach, but it may look to expand data transparency rules or allow GPAI models a little more leniency. Navrina Singh, founder of Credo AI and a national AI advisory committee member, believes that while the AI Act is a huge moment for AI governance, things will not change rapidly, and there’s still a ton of work ahead. “The focus for regulators on both sides of the Atlantic should be on assisting organizations of all sizes in the safe design, development, and deployment of AI that are both transparent and accountable,” Singh told The Verge in December. She adds there’s still a lack of standards and benchmarking processes, particularly around transparency.

PREFACE Transnational threats interact in a complex system along with threats from state-actors, often reinforcing each other and creating compounding and cascading risks to U.S. national security. Increasing interconnections among countries also have created new opportunities for transnational interference and conflict. Several clear and direct challenges are the rapid development of technologies, the spread of repression beyond physical borders, the threats posed by transnational organized crime and terrorism, and the societal effects of international migration. CONTESTED SPACES Disruptive Technology New technologies—particularly in the fields of AI and biotechnology—are being developed and are proliferating at a rate that makes it challenging for companies and governments to shape norms regarding civil liberties, privacy, and ethics. The convergence of these emerging technologies is likely to create breakthroughs, which could lead to the rapid development of asymmetric threats—such as advanced UAVs—to U.S. interests and probably will help shape U.S. economic prosperity. • For example, stealth technology has significantly impacted conventional defense systems and has driven the efforts of varying countries to start a new round of research on detection systems and guided weapons. A key trend is the development of advanced materials with enhanced stealth properties with reduced reflection and absorption properties. Advances in AI and new machine learning models are moving AI into its industrial age, with potentially huge economic impacts for both winners and followers and unintended consequences— from rampant deepfakes and misinformation to the development of AI-generated computer viruses or new chemical weapons. Generative AI is a means for discovering and designing novel technologies and advanced system-level processes that could strengthen a country’s technological, economic, and broader strategic competitiveness. • China is pursuing AI for smart cities, mass surveillance, healthcare, drug discovery, and intelligent weapons platforms. Chinese AI firms are already world leaders in voice and image recognition, video analytics, and mass surveillance technologies. • PRC researchers have described the application of generative AI to drug discovery as “revolutionary.” On average, it takes more than 10 years and billions of dollars to develop a new drug. AI can make drug discovery faster and cheaper by using machine-learning models to predict how potential drugs might behave in the body and cut down on the need for painstaking lab work on dead-end compounds. [ 31 ] • Russia is using AI to create deepfakes and is developing the capability to fool experts. Individuals in warzones and unstable political environments may serve as some of the highest value targets for such deepfake malign influence. Innovators in synthetic biology probably will control new military and commercial applications and hold trillions of dollars in production capacity, including supply chains for products that vary from disease-resistant crop seeds to metals to pharmaceuticals. • Countries, such as China and the United States, that lead biotechnological breakthroughs in fields such as precision medicine, synthetic biology, big data, and biomimetic materials, will not only drive industry growth, but also international competition and will exert substantial influence over the global economy for generations.

On the question of whether ingesting copyrighted works to train LLMs is fair use, LCA points to the history of courts applying the US Copyright Act to AI. For instance, under the precedent established in Authors Guild v. HathiTrust and upheld in Authors Guild v. Google, the US Court of Appeals for the Second Circuit held that mass digitization of a large volume of in-copyright books in order to distill and reveal new information about the books was a fair use. While these cases did not concern generative AI, they did involve machine learning. The courts now hearing the pending challenges to ingestion for training generative AI models are perfectly capable of applying these precedents to the cases before them. Why are scholars and librarians so invested in protecting the precedent that training AI LLMs on copyright-protected works is a transformative fair use? Rachael G. Samberg, Timothy Vollmer, and Samantha Teremi (of UC Berkeley Library) recently wrote that maintaining the continued treatment of training AI models as fair use is “essential to protecting research,” including non-generative, nonprofit educational research methodologies like text and data mining (TDM). If fair use rights were overridden and licenses restricted researchers to training AI on public domain works, scholars would be limited in the scope of inquiries that can be made using AI tools. Works in the public domain are not representative of the full scope of culture, and training AI on public domain works would omit studies of contemporary history, culture, and society from the scholarly record, as Authors Alliance and LCA described in a recent petition to the US Copyright Office. Hampering researchers’ ability to interrogate modern in-copyright materials through a licensing regime would mean that research is less relevant and useful to the concerns of the day. As the lawsuits illustrate, the availability of generative AI trained on datasets that include copyrightable material has raised questions about the intersection of copyright law and AI. But as discussed above, many of the questions raised have already been litigated. Nick Garcia, policy counsel at Public Knowledge, pointed out during a recent Chamber of Progress panel on AI, art, and copyright that concerns about web crawling to collect data—a practice that the Times takes issue with in its lawsuit—have been around for decades, and courts have found web crawling to be a fair use. New York Times v. Microsoft et al. is, of course, just one legal battle through which the courts will interpret copyright law in the US, and it may be years before these cases are settled. Copyright law as it applies to AI will also be informed by the US Copyright Office Study, which will culminate in a report this year. LCA will monitor these lawsuits and pursue opportunities to advance the interests of scholars, educators, students, and the public via selected amicus briefs and discussions of the issues and the range of library concerns with legislators and regulators.

More recently, in the United States, the Senate Subcommittee on Intellectual Property held a hearing in June 2023 on AI and patent law and AI and copyright law as part of a series on AI and IP.208 The USPTO issued a Request For Comments on AI and patent inventorship in February 2023.209As part of an initiative launched in 2023,210 the USCO has been holding listening sessions on AI and copyright with stakeholders and conducting webinars to inform the public and clarify its position and disclosure requirements on AI-generated output and copyright law.211 In addition, the USCO issued a notice seeking comments on a number of issues related to AI-generated works, including the legal status of AI generated output.212

Functionally, the duties of the Copyright Office have expanded and diversified over time. The Copyright Act dictates that the Office perform “all administrative functions and duties” required to implement the Act.76 This includes: (1) advising Congress, (2) providing information/assistance to federal departments, agencies, and the judiciary, (3) participating in meetings of international organizations, (4) conducting studies and programs, and (5) performing “such other functions as Congress may direct, or as may be appropriate in furtherance of the functions and duties [above].”77 These duties appear wholly legislative in nature; however, they have been supplemented by grants of specific rulemaking authority, which have given the Office a larger role in administration of a “complex federal statute.”78 Thus, much like the Library of Congress, the Copyright Office now appears to serve dual functions: the Office’s research and policy input seem to suggest a legislative role, while its rule- making and administrative activities “are more consistent with an executive agency.”79 It appears to be these new rulemaking and administrative functions that led President Clinton to declare, in a signing statement accompanying the Digital Millennium Copyright Act (DMCA), that the Office is an executive branch entity for constitutional purposes.80 At least one court has echoed that it is an agency’s function, rather than its structure, that dictates the branch of government to which it belongs and that it is not necessary to reconcile these two sets of characteristics.81 The Fourth Circuit in Eltra notoriously stated: “it would appear indisputable that the operations of the Office of Copyright are executive.”82 The registration of copyrights, the court said, “cannot be likened to the gathering of information ‘relevant to the legislative process’ nor does the Register perform a function ‘which Congress might delegate to one of its own committees.’”83 This point is itself arguable, since copyright registration is not re- quired to bring copyrights into existence. The intake and filing of copyright holders’ submissions of existing copyrights for registration seem to be merely an extension of the Library of Congress’s manifestly legislative recordkeeping function.84 Regardless, the Eltra court concluded the Copyright Office was emphatically executive, claiming Supreme Court agreement with this conclusion since 1909—despite not actually citing any alleged Supreme Court precedent.85 Importantly, Eltra was decided prior to three important Supreme Court cases on the separation of powers.86 These cases “embrace a formal rather than functional approach to situating entities like the [Copyright] Office within the legislative or executive branches, and potentially undermine Eltra’s analysis.”87 Most notably, in United States v. Brooks,88 the Eastern District of Pennsylvania articulated this syllogistic reasoning: the Library of Congress is clearly a part of Congress; the Library of Congress is therefore part of the legislative branch; the Copyright Office is a division of the Library of Congress; the Copyright Office is therefore part of the legislative branch.89 Any other conclusion, the court said, would be “untenable.”90 The Brooks court’s presiding judge responded specifically to the reasoning in Eltra, noting, “The status of the Copyright Office is an open question in the Third Circuit, and I do not find Eltra persuasive.”91 Although the court acknowledged that the Copyright Office is required to perform some administrative functions and duties under the Copyright Act, it disregarded Eltra’s functional approach.92 “Acting similar to an executive agency,” the court said, “is not the same as being part of the executive branch.”93 Of course, if the status of the Library of Congress is considered an open question, as it is in at least one circuit, the Brooks court’s otherwise compelling logic would seem inconclusive.94

Other authorities also have reached the same conclusion. A national commission empaneled by Congress and the President to study the impact of technology on copyright law issued a report in 1978 which concluded that “there is no reasonable basis for considering that a computer in any way contributes authorship to a work produced through its use.”[7] The Copyright Office has refused to register machine-created works for years,[8] stating that “[t]o qualify as a work of ‘authorship’ a work must be created by a human being…[T]he 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.”[9] Although not binding, Courts give deference to the Copyright Office’s reasonable interpretation of the Copyright Act.[10] Finally, a district court recently ruled, consistent with Copyright Office guidance, that an AI-generated work was not registrable because it lacked human authorship.[11] Can people claim copyright protection in works they create with the assistance of AI? Yes, but only if, and only to the extent that, the human contribution is independently copyrightable. Non-human contributions remain unprotectable in such “joint” works. That is consistent with longstanding requirements that an author make independently copyrightable contributions to a work to qualify as an author.[12] Consistent with this authority, in March 2023 the U.S. Copyright Office has given lengthy guidance about registering AI-assisted works which makes clear that only human, independently protectable human contributions to a work can qualify for copyright protection: “If a work’s traditional elements of authorship were produced by a machine, the work lacks human authorship and the Office will not register it. For example, when an AI technology receives solely a prompt from a human and produces complex written, visual, or musical works in response, the ‘traditional elements of authorship’ are determined and executed by the technology—not the human user. Based on the Office’s understanding of the generative AI technologies currently available, users do not exercise ultimate creative control over how such systems interpret prompts and generate material. Instead, these prompts function more like instructions to a commissioned artist—they identify what the prompter wishes to have depicted, but the machine determines how those instructions are implemented in its output. “For example, if a user instructs a text-generating technology to ‘write a poem about copyright law in the style of William Shakespeare,’ she can expect the system to generate text that is recognizable as a poem, mentions copyright, and resembles Shakespeare’s style. But the technology will decide the rhyming pattern, the words in each line, and the structure of the text. When an AI technology determines the expressive elements of its output, the generated material is not the product of human authorship. As a result, that material is not protected by copyright and must be disclaimed in a registration application.” “In other cases, however, a work containing AI-generated material will also contain sufficient human authorship to support a copyright claim. For example, a human may select or arrange AI-generated material in a sufficiently creative way that ‘the resulting work as a whole constitutes an original work of authorship.’ Or an artist may modify material originally generated by AI technology to such a degree that the modifications meet the standard for copyright protection. In these cases, copyright will only protect the human-authored aspects of the work, which are ‘independent of’ and do ‘not affect’ the copyright status of the AI-generated material itself.” “This policy does not mean that technological tools cannot be part of the creative process. Authors have long used such tools to create their works or to recast, transform, or adapt their expressive authorship. For example, a visual artist who uses Adobe Photoshop to edit an image remains the author of the modified image, and a musical artist may use effects such as guitar pedals when creating a sound recording. In each case, what matters is the extent to which the human had creative control over the work’s expression and ‘actually formed’ the traditional elements of authorship.”[13] Do the creators of AI infringe copyright when for training purposes they copy into the AI’s “database” third parties’ copyrighted works that the AI uses to generate new works in response to user prompts? That is unclear at the moment, though there may be judicial guidance in the next few months. A potentially infringing “copy” is made to the extent it is uploaded without permission, but the success of an obvious fair use defense is unpredictable. “Fair use” is a complicated copyright defense that allows certain kinds of copying in certain circumstances for what usually are deemed to be socially desirable purposes. Explaining it in detail is beyond the scope of this paper. However, at least one appellate court has already held that Google’s efforts to copy every book in the English language to create a searchable database, and to provide digital copies to libraries, was fair use.[14] If Google can do that for its search engine, can it do it for AI? Why or why not? Do the works the AI creates with the uploaded third party works on which it was trained those third parties’ copyrights? Most of the time, probably not under current copyright law. Protectable elements in the underlying works input into the AI for training purposes that the AI uses to create the AI-generated work will likely be so diluted that the output images or text will not be substantially similar reproductions or qualify as derivatives of the underlying works. Several recently-filed pending cases may decide this issue, and a court in at least one has tentatively advised that it intends to so rule, on either substantial similarity or fair use grounds.[15] Issuance or written decisions may provide greater clarity on this issue.

For decades, the Copyright Office has been a small and sleepy office within the Library of Congress. Each year, the agency’s 450 employees register roughly half a million copyrights, the ownership rights for creative works, based on a two-centuries-old law. In recent months, however, the office has suddenly found itself in the spotlight. Lobbyists for Microsoft, Google, and the music and news industries have asked to meet with Shira Perlmutter, the register of copyrights, and her staff. Thousands of artists, musicians and tech executives have written to the agency, and hundreds have asked to speak at listening sessions hosted by the office. The attention stems from a first-of-its-kind review of copyright law that the Copyright Office is conducting in the age of artificial intelligence. The technology — which feeds off creative content — has upended traditional norms around copyright, which gives owners of books, movies and music the exclusive ability to distribute and copy their works. The agency plans to put out three reports this year revealing its position on copyright law in relation to A.I. The reports are set to be hugely consequential, weighing heavily in courts as well as with lawmakers and regulators. “We are now finding ourselves the subject of a lot of attention from the broader general public, so it is a very exciting and challenging time,” Ms. Perlmutter said. The Copyright Office’s review has thrust it into the middle of a high-stakes clash between the tech and media industries over the value of intellectual property to train new A.I. models that are likely to ingest copyrighted books, news articles, songs, art and essays to generate writing or images. Since the 1790s, copyright law has protected works so an author or artist “may reap the fruits of his or her intellectual creativity,” the Copyright Office declares on its website. That law is now a topic of hot debate. Authors, artists, media companies and others say the A.I. models are infringing on their copyrights. Tech companies say that they aren’t replicating the materials and that they consume data that is publicly available on the internet, practices that are fair use and within the bounds of the law. The fight has led to lawsuits, including one by The New York Times against the ChatGPT creator OpenAI and Microsoft. And copyright owners are pushing for officials to rein in the tech companies. “What the Copyright Office is doing is a big deal because there are important principles of law and lots and lots of money involved,” said Rebecca Tushnet, a professor of copyright and intellectual property law at Harvard Law School. “At the end of the day, the issue is not whether these models will exist. It’s who will get paid.” Congress created the Copyright Office in 1870 to register licenses for books, maps, essays and other creative works and store those works for the use of lawmakers at the Library of Congress. The first registration was given to the “Philadelphia Spelling Book,” a children’s language book. When Ms. Perlmutter, a veteran copyright official and former intellectual property lawyer for Time Warner, was appointed to lead the Copyright Office in late 2020, she promised to bring the office into the modern era by focusing on big tech trends. She took inspiration from previous leaders, who dealt with technological innovations including the camera, records, Xerox machines, the internet and streaming music, all of which required the office to weigh in on how copyright would apply and advise Congress on proposed updates to the law. Right away, A.I. became a hot topic. Stephen Thaler, a computer scientist, tried to register an A.I.-generated art piece for a copyright by submitting an application on the Copyright Office’s website. In 2019, the office rejected his first attempt to register the piece, a pixelated scene of train tracks running through a tunnel overgrown with brush and flowers called “A Recent Entrance to Paradise.” In February 2022, Ms. Perlmutter declined his second attempt to register the piece on the same grounds: Copyrights were given only to original works created by humans. The decision — a first on an A.I.-produced work — set an important precedent. Artists and lawmakers flooded Ms. Perlmutter’s office with emails and phone calls asking her to also intervene in the way A.I. companies were using copyrighted material to train their systems. In August, she opened the formal review of A.I. and copyright law. The office said it would examine whether the use of intellectual property to train A.I. models violated the law and would look more deeply into whether machine-generated works could be eligible for copyright protections. The office said it would also review how A.I. tools were creating content that used the names, images and likenesses of individuals without their consent or compensation. “The attention on A.I. is intense,” Ms. Perlmutter said in an interview. “The current generative A.I. systems raise a lot of complicated copyright issues — some have called them existential — that really require us to start grappling with fundamental questions about the nature and value of human creativity.” The interest in the office’s review was overwhelming. The office solicited public comments on the topic and received more than 10,000 responses on a form on its website. A typical policy review gets no more than 20 comments, the office said. Tech companies argued in comments on the website that the way their models ingested creative content was innovative and legal. The venture capital firm Andreessen Horowitz, which has several investments in A.I. start-ups, warned in its comments that any slowdown for A.I. companies in consuming content “would upset at least a decade’s worth of investment-backed expectations that were premised on the current understanding of the scope of copyright protection in this country.” OpenAI, Microsoft, Meta (Facebook’s parent) and Google are currently relying on a 2015 court decision in a case filed by the Authors Guild. The guild sued Google in 2005 for scanning books to use in excerpts in its search engine results and to share with libraries. A court ruled that Google had not violated copyright law. It said that the scanning of entire books was permissible because Google didn’t make the full book available and that it was “transformative” use of copyrighted material. Google relied on an exemption to copyright law known as “fair use” that allows limited replication of copyrighted material for things like criticism, parody or other transformational uses. Google, Meta and the A.I. start-up Anthropic all echoed arguments from that case in their comments to the Copyright Office, including that A.I. copies the information to analyze data, not repurpose it for creative works. Authors, musicians and the media industry argued that by taking their content without permission or licensing payments, the A.I. companies were robbing them of their livelihoods. “The absence of consent and compensation in this process is theft,” Justine Bateman, the “Family Ties” actress and author, wrote in comments to the Copyright Office. News Corp, which publishes The Wall Street Journal and The New York Post, implored the office to “not lose sight of this simple truth: Protecting content creators is one of copyright law’s core missions.” (The Times also submitted a comment.) Ms. Perlmutter said she and a staff of about two dozen copyright lawyers were going through each comment filed to the office. Still, the office may not offer clear-cut views that will satisfy either the tech companies or creative people. “As technology gets more and more sophisticated, the challenges are exponentially more difficult and the risks and rewards are exponentially greater,” Ms. Perlmutter said.

However, the carbon emissions of AI training and inference today are minuscule compared to the computing sector as a whole, let alone the more carbon-intensive activities of our industrial civilization, such as construction and transportation, which together account for more than a third of our total carbon emissions.

While the use of AI may continue to grow dramatically over the coming years, its energy efficiency will also likely continue to improve dramatically, making it unclear whether AI will ever be among the most significant carbon emitters on a global scale.

Rapidly improving artificial intelligence (AI) capabilities will help accelerate the energy transition. Both established and emergent AI capabilities—such as large language models (LLMs)—can be applied to an array of strategic, technical, financial, and policy challenges posed by decarbonization. It is critical for energy transition stakeholders to monitor, understand, and carefully apply these capabilities to their unique decarbonization challenges, while also addressing the risks involved. The most consequential new class of AI, generative AI, is able to analyze and create text, audio, code, and even molecular design—doing so faster and often with higher quality than human-created counterparts. Generative AI uses extraordinary volumes of training data and novel data-processing mechanisms which require unprecedented computational power. Data center load growth, driven by a range of factors, is forcing utilities across the United States and Europe to revisit system planning needs. Indeed, this added demand is—in some regions—delaying the retirement of coal-fired power plants. To ensure that climate targets are met, data center growth must coincide with transmission upgrades, energy efficiency improvements, and new low-carbon generation capacity. More broadly, policymakers must also consider how to harness the potential from generative AI while managing complex uncertainties, from inaccurate outputs and data leakage to AI-enabled cyberattacks on critical infrastructure. The deployment of generative AI will require rigorous human oversight, particularly in the early stages. Given the capabilities of generative AI, integration into organizational workflows can help energy stakeholders in multiple ways—for example, lower regulatory compliance costs, consider strategic planning options, and evaluate the financial risk around their low-carbon investments, among others. 1. Strategic planning Recent demonstrations of generative AI capabilities are impressive. Generative AI can already outline, summarize, and draft documents cheaper and faster than many humans. It can also help humans conduct strategic tasks more effectively. A study by Harvard Business School examined the effects of GPT-4—the model behind ChatGPT—on knowledge workers’ productivity, finding that GPT-4 significantly improved workers’ abilities to generate effective ideas and develop implementation plans. Another study from University College London found that a collection of LLMs could give strategic recommendations at a comparable level to human experts. As strategic planning use cases are systemic and across industries, improvements in productivity would apply across the decarbonization value chain. 2. Regulatory compliance Some generative AI use cases will directly enhance clean energy project developers’ ability to manage cumbersome regulatory processes. As generative AI capabilities are integrated into institutional workflows, they will assist on tasks ranging from simple emails to complex, costly, and time-consuming regulatory processes. The Pacific Northwest National Laboratory, as part of its PolicyAI, initiative, recently found that LLMs could streamline the public comment-review process under the National Environmental Policy Act (NEPA), which is burdensome for many renewables firms. Importantly, generative AI may aid regulators by accelerating reviews of a variety of environmental impact studies. For instance, after New York State attempted to ease traffic and pollution by passing traffic congestion pricing, an exhaustive environmental review took five years and more than 4,000 pages of analysis. By streamlining portions of these document-intensive regulatory tasks, generative AI can speed up environmental reviews, giving infrastructure projects a quicker go/no-go decision. 3. Decarbonization investment analytics A range of AI tools, using both existing techniques and generative AI, are being developed to assist with financial and economic modeling, a critical but resource-intensive task for renewable energy projects. While still at the early stages, generative AI tools may be able to partially or even fully build financial models or propose complex scenario plans. In addition, AI is already being used to enhance corporate due diligence by detecting anomalies in financial statements, summarizing earnings call transcripts, or rapidly analyzing trade press. These capabilities will continue to assist both investors and corporate mergers-and-acquisitions teams in their decarbonization investments. 4. Energy asset management Financial and economic modeling tools overlap with another essential aspect of decarbonization: advanced energy asset management. Currently, communications with energy asset field operators are typically executed via middle management and dashboards with both planned and ad hoc analytics. Generative AI may enable more simplified analytics and communication with the workers physically assessing and repairing assets. At the energy asset management level, generative AI tools could deliver improvements in compiling, summarizing, and communicating asset performance in a customized manner for financial managers. 5. Wildfire risk assessment In parallel to generative AI, another area of quiet yet significant advancement has been machine-learning (ML) models for weather forecasting, which have produced some extraordinary results. Further advances in weather forecasting could help mitigate the climate change-driven fire season. Wildfires themselves exacerbate the climate crisis—global fires produce emissions of about 2 gigatons of carbon dioxide equivalent per year, equal to 4 percent of total global emissions. These fires can also force large populations indoors for weeks due to health risks and poor air quality. Further investment in AI/ML-based modeling could help manage these risks by predicting the probable location and magnitude of potential wildfires and improving real-time surveillance of smoke, enabling firefighters to combat the over 80,000 wildfires that occur in the United States alone every year. Despite the current AI hype cycle and the early-stage risks around generative AI, improving the broad range of AI models will be integral to developing a low-carbon economy. The magnitude and pace will be difficult to predict, as models are integrated into institutional workflows. Human oversight, particularly around critical infrastructure, must remain comprehensive. If managed appropriately, these emergent capabilities will yield important advances in regulatory analysis, environmental management, strategic planning, and an array of challenges essential to achieving net-zero emissions.

This paper has examined whether a socially-minded entity might be able to control AGI if the technology underlying it is non-excludable. It is shown that either entity will be preempted by a for-profit firm, limiting both its ability and incentive to mitigate unrestricted AGI products that emerge in equilibrium. In other words, it demonstrates that a form of ‘self-regulation’ is unlikely to be effective. This leads to a natural conclusion that regulating the harmful effects of AGI, or any other technology for that matter, is likely to be conducted by government policy. Indeed, if such policies can create conditions by which externalities are internalised by for-profit firms (for instance, through product liability laws), it will likely bolster conditions by which socially-minded activities could prove more effective. That, however, is a subject left to future research. Also left for future research is whether the government intervention required in this case could remain under the control of national governments or whether it requires a global effort. The results in this paper point, for the same reasons as given in this paper, to the latter.14

If 2023 was the year of AI hype, 2024 is the year that trust in AI becomes mission-critical. A string of AI fails has shown that a self-regulatory, pro-innovation approach to AI can leave us less safe, without actually making AI better. Without effective guardrails, we may be speeding towards another AI winter. The EU AI Act and Biden’s AI Executive Order are raising the bar with comprehensive rules. Organizations that cut corners in the race to be first-to-market may find themselves locked out of lucrative government contracts and R&D opportunities. AI isn’t new. Neither is AI risk. AI experts have long warned of the real-world harms poorly-designed AI tools can cause. The AI Incident Database chronicles incidents dating back to 1983, where a human operator averted nuclear disaster by identifying a computer-generated false alarm. The database now contains 629 incidents. Best sellers like Weapons of Math Destruction and documentaries like Coded Bias exposed the public to a range of AI failures, making AI harms a mainstream topic well before ChatGPT entered the hype cycle. Throughout this time, concrete measures to mitigate AI risk were proposed. The EU’s GDPR introduced new rules in 2016 to address AI bias, biometrics, profiling and algorithmic decision-making. The ACM Conference on Fairness, Accountability, and Transparency has been running since 2018. Scores of AI frameworks and standards have been published. Yet too many organizations continue to move fast and break things. They develop and release AI products without assessing or mitigating risks or even confirming their products work as advertised. When they rely on ‘black box’ models, they’re often hard pressed to explain how their AI works, or if it works at all. Yet without a clear legal requirement that is likely to be enforced, they can reap profits with an MVP and deal with regulatory or consumer blowback after someone has been harmed and if they feel compelled by regulators, public pressure or a genuine desire to fix the problem. This lawful-but-awful approach turns us into human guinea pigs. And as stories of peoples harm by AI continue to make headlines, it erodes trust in AI and fuels AI fears.

The second class of AI models are “large” models, which are the type of models that motivated the research in this paper. These models are trained on such a large volume of data that it is not possible to establish provenance ahead of time because, for the general operation of these models, no individual piece of content or even collection of content is significant enough to impact the vast majority of output from these AI models. This is not to say that the provision and availability of such AI models may not end up having a commercially harmful impact on a content provider. It is just that it is not clear in advance whether that might occur, nor is it clear what precise value specific content has in reducing AI training costs, even if the content used in training as a whole has an identifiable benefit. Once again, how content is used in training large models has been discussed considering human analogues. This is from OpenAI’s response to the New York Times’ lawsuit: Just as humans obtain a broad education to learn how to solve new problems, we want our AI models to observe the range of the world’s information, including from every language, culture, and industry. Because models learn from the enormous aggregate of human knowledge, any one sector—including news—is a tiny slice of overall training data, and any single data source—including The New York Times—is not significant for the model’s intended learning.5 Here is a scenario that captures the connection: Scenario 3a Teaching (AI Version): A developer builds a chatbot that teaches students economics based on a corpus of textbooks. Scenario 3b Teaching (Human Version): A person teaches students economics based on their reading of a corpus of textbooks. Note that in teaching, it is generally agreed upon that human teachers are not subject to copyright prohibitions, but it is precisely the issue with AI that remains an open copyright regulatory question. Once again, it is important to use the model to understand precisely why the human version involves a particular conclusion to see if it can assist in understanding the approach for the AI version. Analysing large AI models requires some adjustment to the baseline model of Section 2. First, the lack of provenance means that there is no opportunity for negotiation for the use of particular content, nor is it possible to know what the precise commercial harm (if any) to a content provider might be. Thus, the baseline model is amended to include uncertainty over potential commercial harm prior to setting prices for original content and AI use and to remove the ability to negotiate. Nonetheless, a copyright regime could prohibit (say, subject to large statutory infringement damages) the use of such content or permit its use. The model of Section 4 examines the expected social welfare outcomes from each of these traditional rights regimes and demonstrates that having no copyright protection is superior to copyright protection if the value of the content as a whole in lowering AI training costs is proportionately high relative to the expected commercial harm across all content providers. The large AI model context also creates an opportunity for a different type of rights regime. Section 4 proposes and analyses a mechanism that is an ex post ’fair use’ regime. Rather than intended to operate prior to potential infringement taking place, this mechanism operates ex post. This mechanism allows an AI model to be trained on all available content, and then, once original content owners observe leakage and realise commercial harm, exercising an option to sue the AI provider for lost profits. The regime sets a threshold on the magnitude of such harm, and those content providers who experience harm above that threshold exercise the option for lost profits while others do not. It is demonstrated that this mechanism offers stronger incentives for content providers than no copyright protection but also stronger incentives for AI training than traditional rights regimes. Moreover, subject to the financial feasibility of the AI provider, the threshold can be set low, implying that the mechanism will generate the highest expected social welfare amongst rights regimes.

As artificial intelligence (AI) moves beyond performing simple tasks to creating original content, it blurs the lines between humans and machines. In doing so, it challenges a core tenet of many traditional intellectual property (IP) frameworks: only works created by humans are protected by copyright laws. The development and training of AI models also raise another important copyright issue: whether the use of third-party content in that exercise might infringe on the copyright of that content. Emerging use cases around generative AI are disrupting traditional views of creativity, authorship and ownership and pushing the boundaries of copyright law. As the world catches up with innovation, the resulting legal ambiguity impacts all sides of the AI equation – developers, content creators and copyright owners. To create content, generative AI learns from processing vast amounts of information. This training process can involve using copyright-protected content, raising a critical question: does training generative AI models constitute copyright infringement, exposing developers to related claims? If so, is new legislation needed to address this issue in a manner that does not unreasonably impede innovation but still respects the rights of creators? These questions have led some jurisdictions to consider amendments to existing text and data mining exemptions and/or fair use/fair dealing exceptions to cover the training of generative AI. Such exemptions permit certain activities that would otherwise constitute copyright infringement. Other jurisdictions are still awaiting guidance, creating a drastically uneven regulatory landscape that will continue for some time. But, no matter where they stand, legislators need to consider this issue to incentivise innovation while protecting content creators' rights. Looking past the legal uncertainty at traditional means of creativity, humans constantly learn as we receive and process information. Typically, things we create are inextricably linked to what we have learned from others and our experiences. One could argue that AI is no different in that it learns from the underlying datasets to produce something new. Copyright - ownership of content Copyright law is a key means by which innovation and creation are rewarded – and thus enabled – raising the stakes for courts and intellectual property authorities to define copyright ownership in the context of generative AI. Depending on where you are in the world, the definition of copyright varies, a significant legal consideration in its own right. But virtually everywhere, copyright requires meaningful human contribution. Generative AI, which often involves little human input, shatters the mould of this traditional framework. It calls into question whether AI-generated content can be protected by copyright in the first place and, if so, then who owns this copyright? Is it the developer of the AI model who arguably enabled the 'creative' process? Is it the owner of the AI model? Or is the person who considered and input the prompts that allowed the AI tool to generate the content? Of course, this depends on whether there was sufficient human contribution to justify copyright protection in the first place. Currently, most generative AI tools provide that users own the content they generate in their terms of use. However, if the content is not copyright-protected in the first place, then the terms of use can't change that position. There is less clarity around who is responsible if AI-generated content infringes someone else's copyright in a protected work. To answer these questions requires courts and intellectual property authorities to challenge the scope of traditional copyright law in relation to generative AI, which could lead to changes to existing legislation covering ownership of computer-generated content. Eventually, it is possible that copyright law will need to be further reimagined and redefined to account for the possibility of recognizing a machine as its own entity with independent authority and the ability to own and protect content, though this does not appear to be on the horizon in the short term. In this more progressive scenario, it is unclear where liability would rest for negative consequences (including copyright infringement) caused by generative AI outputs. Looking forward (and backward) Technology moves quickly, while the winds of legal change blow slowly. Eventually, the world will see laws introduced and pressure-tested in the courts to address copyright issues relating to generative AI, as well as the myriad other issues that surround AI-related developments. Legal uncertainty associated with new technologies always creates risk, requiring new approaches to emerging issues that strike a balance between existing rights and innovation. By way of a somewhat recent example, the internet also broke the norms and bounds of copyright law, requiring new legislation and judicial guidance on issues such as the protection of online content, legal responsibility for uploading and downloading content, ownership of user-generated content and the liability of online intermediaries like hosts and ISPs. With AI too, the world will soon see legal developments that look to consider the interests of all stakeholders, including AI developers, owners and users and those who own copyright in the content needed to train these tools. These developments will also ideally provide clarity on who might own the intellectual property rights in AI-generated content.

In thinking about the requirements of “human origination” under the copyright and patent laws, the core questions are: (1) did a human being originate any or all of the subject matter? and (2) if a human being did not originate every aspect of the subject matter, what portion comprises the “human contribution?” The use of AI to achieve or assist in generating the result(s) for which copyright or patent protection are sought, raises a number of considerations that invoke or seek to circumvent those questions. Moreover, there are a number of stages in AI processes or AI- assisted processes in which the “spark” of human creativity may be evident. For example, a non- exhaustive list of points of potential authorship or inventorship can include: (1) the corporation or other entity (if any) that commissioned the development of the AI technology, such as the algorithm(s); (2) the creator(s) of the AI technology, such as the algorithm(s); (3) the individual(s) who coded the system; (4) the individual(s) who defined the goals, biases, or training materials employed in the AI system; (5) the individual(s) who intervened in the particular AI process or task to affect the result; or (6) the individual(s) who took a product of the AI process and modified it or added content to it that expresses at least a modicum of post-AI originality.13 Of course, this latter scenario – the addition of a “human element” to the AI end product – fits very comfortably in traditional notions of authorship and inventorship.14 Authors frequently “borrow” from, or find inspiration in, pre-existing materials and inventors often improve upon or find new uses for items or processes that already exist. In the realm of copyright the process is demonstrated by Andy Warhol’s spectacularly successful modifications of prior third-party works to create new ones.15 This is a staple technique in numerous artforms, although very few practitioners would adopt Warhol’s explanation – “I want to be a machine” – for the use of depersonalizing commonality and repetition in the artistic process.”16 The key, however, is that there are a number of competing interests that underlie the creation of AI systems, their employment, and their products. No matter the result, however, it is clear that intellectual property laws (and contract law applicable to, e.g., transfers and licenses) will play robust parts in the development of the industry and the commercialization of its products. II. U.S. Copyright and AI As the use of AI systems in the creation of expressive works continues to grow in terms of applications and economic consequences, it was inevitable that the core issues of AI and human creativity would surface in the context of the U.S. copyright registration process.17 This section of the discussion will focus on “the frontline” territory of the copyright registration process – a process that, in most cases, is the prerequisite to enforcement of a copyright claim.18 Also a selection of relevant U.S. judicial decisions and other considerations will be discussed.

Intellectual Property (IP) rights are fundamental to how economies organize innovation and steer the diffusion of knowledge. Yet, while research in economics and management of innovation has largely focused on patents (Awaworyi Churchill et al., 2022; Hall & Harhoff, 2012), other IP rights, and copyright, in particular, have received considerably less attention. Copyright law has continuously developed in response to emerging technologies and their challenges and opportunities not only for creators but also for intermediaries and consumers (Bently et al., 2010). For most of copyright history, storage and distribution technology for written works, sound recordings, photographs, and films were analog. With a broad scope and long protection terms, copyrights have played an essential but relatively passive role in the development of content industries, creating significant value for rightsholders but also a deadweight loss for society. Digital technologies, perhaps starting with the introduction of the microchip and personal computers and much accelerated by the internet, have immensely challenged the copyright system (Greenstein et al., 2013). First, this led some jurisdictions to extend copyright law to new subject matters, for example, by adding protection for software and databases. Second, technology exacerbated the enforcement of demand-side copyright infringement through peer-to-peer file-sharing and unlicensed content platforms. Third, by substantially lowering the cost of creating, distributing, and promoting works, we have seen market entry at an unprecedented scale (Aguiar et al., 2023; Waldfogel, 2017; Wu & Zhu, 2022). However, not all new works are original, but some represent the re-use of existing works which may be potentially infringing—in the following marked as supply-side copyright infringement. In this paper, we provide a synthesis of the empirical literature on the law and economics of copyright in the digital age. First, we review important institutional cornerstones of copyright, focusing on the European Union (EU) and the United States, and also introduce the key developments in copyright law in the digital age. Second, we review the available empirical evidence, with particular attention to the economic impact of digital technologies and the varied interests of stakeholders in the digital economy. We structure our review around the issues of copyright infringement, enforcement, and digital business models and licensing. Third, having established what is known in the field, we continue to develop a research agenda. In particular, we highlight open questions about the heterogeneity and diversity of works and actors in the current copyright system. For example, evidence is limited on economically significant application areas of copyright, such as software code, and with respect to the diverse socio-demographic backgrounds of creators. Further, as licensing technology continues to evolve, we argue that it becomes ever more critical to understand its effects on the involved stakeholders and overall welfare. We argue that copyright law plays a pivotal role in the next wave of technological innovation, acting as a gatekeeper concerning data and Artificial Intelligence (AI). For example, it will be important to study how (the absence of) copyright exceptions for training data affect AI innovation. As governments continue to regulate the digital world, copyright increasingly intersects with competition policy and privacy protection. Finally, we describe how digitization has brought researchers new tools and datasets that can help to answer important questions

Copyright Protection of AI-Generated Contents Current law makes it very difficult, if not impossible, to copyright AI-generated content.[3] Yet, paradoxically, the law might be turning things upside down by considering copyright for prompts while denying it for the content AI produces. [4] The underlying principle of the law is that the work must have “human authorship.”[5] However, arguably, all man-made creations inherently involve human creativity. For instance, since LLMs cannot generate output without prompts written by humans, doesn’t this satisfy the legal requirement for human authorship? This leads to the next question: what exactly constitutes “human authorship”? The Copyright Registration Guidance’s answer is: “users do not exercise ultimate creative control over how such systems interpret prompts and generate material. Instead, these prompts function more like instructions to a commissioned artist—they identify what the prompter wishes to have depicted, but the machine determines how those instructions are implemented in its output.”[6] However, even in the case of a commissioned artist, the law regarding works made for hire would transfer authorship to the entity that commissioned the work. See 17 U.S.C. § 201(b). In other words, drawing an analogy between AI and a commissioned artist suggests that authorship would more naturally belong to the entity that commissioned the work: here, the prompt writers. Alternatively, if one views AI as merely a tool, then there is even less justification for denying copyright on AI-generated content. Indeed, when someone uses a pen to write a poem, no one questions the authorship of the person holding the pen. Why, then, do some argue that when a prompt is inputted and the machine produces an output, the prompt writer should be denied authorship? Many explanations exist, but one might be our unfamiliarity with AI’s capabilities. We see a significant transformation from a simple input to a complex output and feel disconnected from the authorship because the process is beyond our understanding. However, by denying our authorship in this transformation, we are essentially dismissing the influence of our own desires in shaping the AI’s output. Our desires are the very power that can guide the AI to fulfill our deepest needs. We can cultivate this power by recognizing and rewarding the content generated by prompt writers, who articulate our desires at a conscious level. Granting copyright protection to AI-generated content in exchange for the disclosure of prompts, while establishing that these prompts are not protected by copyright when used with LLMs, could resolve many issues. First, it offers incentives for prompt writers to share their prompts with the public, enhancing our awareness of our own thoughts and desires. Second, users would be free to use the prompts however they wish without restriction. Third, it simplifies oversight for the LLM platforms. Instead of filtering prompts, platforms need only ensure that the outputs are not identical to the copyrighted content initially generated by the copyright owner. Indeed, this approach mirrors the patent system, where a limited monopoly is granted for a useful machine or method in exchange for disclosing how to make them. This has significantly advanced technology over the past 200 years. Similarly, we could advance human understanding by rewarding individuals’ ability to articulate the intricate details of human desires, thus revealing what we truly want, and helping us either achieve these desires or let them go when we understand their true nature. Whether this is worth pursuing is a decision for Congress. While concerns might arise that using prompts to generate content is so easy that it could overwhelm the Copyright Office, this issue isn’t new; people are already free to register copyrights for countless photographs, yet only do so when it makes economic sense due to application fees. To prevent abuse, the Copyright Office could set an appropriate fee for registering AI-generated content and might require the applicant to disclaim all training materials used to develop the LLM. These measures would help balance economic effects and achieve equilibrium. Therefore, it seems we might be overly cautious about awarding copyright to AI-generated content for reasons we don’t fully understand. Upon considering practical factors, it might make sense to rethink our approach. Are we indeed doing things upside down?

At the crux of this issue is the question of data use. What data do companies like OpenAI have a right to use, and what do concepts like “transform” really mean in these contexts? Questions like this, surrounding the data we train AI systems, or models, like ChatGPT on, remain a fierce academic battleground. The law often lags behind the behaviour of industry. If you’ve used AI to answer emails or summarise work for you, you might see ChatGPT as an end justifying the means. However, it perhaps should worry us if the only way to achieve that is by exempting specific corporate entities from laws that apply to everyone else. Not only could that change the nature of debate around copyright lawsuits like this one, but it has the potential to change the way societies structure their legal system. Fundamental questions Cases like this can throw up thorny questions about the future of legal systems, but they can also question the future of AI models themselves. The New York Times believes that ChatGPT threatens the long-term existence of the newspaper. On this point, OpenAI says in its statement that it is collaborating with news organisations to provide novel opportunities in journalism. It says the company’s goals are to “support a healthy news ecosystem” and to “be a good partner”. Even if we believe that AI systems are a necessary part of the future for our society, it seems like a bad idea to destroy the sources of data that they were originally trained on. This is a concern shared by creative endeavours like the New York Times, authors like George R.R. Martin, and also the online encyclopedia Wikipedia. Advocates of large-scale data collection – like that used to power Large Language Models (LLMs), the technology underlying AI chatbots such as ChatGPT – argue that AI systems “transform” the data they train on by “learning” from their datasets and then creating something new. Effectively, what they mean is that researchers provide data written by people and ask these systems to guess the next words in the sentence, as they would when dealing with a real question from a user. By hiding and then revealing these answers, researchers can provide a binary “yes” or “no” answer that helps push AI systems towards accurate predictions. It’s for this reason that LLMs need vast reams of written texts. If we were to copy the articles from the New York Times’ website and charge people for access, most people would agree this would be “systematic theft on a mass scale” (as the newspaper’s lawsuit puts it). But improving the accuracy of an AI by using data to guide it, as shown above, is more complicated than this. Firms like OpenAI do not store their training data and so argue that the articles from the New York Times fed into the dataset are not actually being reused. A counter-argument to this defence of AI, though, is that there is evidence that systems such as ChatGPT can “leak” verbatim excerpts from their training data. OpenAI says this is a “rare bug”. However, it suggests that these systems do store and memorise some of the data they are trained on – unintentionally – and can regurgitate it verbatim when prompted in specific ways. This would bypass any paywalls a for-profit publication may put in place to protect its intellectual property. Language use But what is likely to have a longer term impact on the way we approach legislation in cases such as these is our use of language. Most AI researchers will tell you that the word “learning” is a very weighty and inaccurate word to use to describe what AI is actually doing. The question must be asked whether the law in its current form is sufficient to protect and support people as society experiences a massive shift into the AI age. Whether something builds on an existing copyrighted piece of work in a manner different from the original is referred to as “transformative use” and is a defence used by OpenAI. However, these laws were designed to encourage people to remix, recombine and experiment with work already released into the outside world. The same laws were not really designed to protect multi-billion-dollar technology products that work at a speed and scale many orders of magnitude greater than any human writer could aspire to.