Search the archive.

The U.S. is awash in patents. Over 350,000 U.S. patents were granted in 2019, four times the per capita rate in 1980. From the perspective of research managers at big firms, patents are cheap and easy to get. For example, in the early 2000s Bill Gates decided that Microsoft was patent-poor, and within a few years the company increased annual patent applications by 50%. Patents are easy to get because the standards of patentability are low and because the burden is on the U.S. Patent and Trademark Office to prove an invention is not patentable. Patent examination is slow. It often takes three years or more. Despite increased staffing, the backlog of patent applications has continued to grow, and examiners spend on average only 20 hours reviewing each application. The patent examiner is required to read and understand the invention in an application, determine whether the invention meets the claims of the application, search existing technology to see if the invention already exists and write a response to the application. Helter-skelter examination causes errors – many patents are too broad, or they cover obvious inventions. To draw attention to problems caused by the flood of low-quality patents, billionaire entrepreneur Mark Cuban endowed a chair at the Electronic Frontier Foundation dedicated to elimination of “stupid patents.” Innovative firms that succeed in assembling many pieces of a technology puzzle into a finished product must consult with a patent lawyer to learn whether their new technology is covered by one or more patents owned by others. Ideally an innovator will get permission to use patented technology, usually for a fee, or redesign its technology to steer clear of relevant patents. In practice this patent “clearance” process is difficult, costly and sometimes impossible. For technologies like smartphones, a patent attorney likely would need to review hundreds of patents, including many patents that are not granted until long after the new product is launched. Failure to license relevant patents creates a risk of litigation and the threat the new technology could be forced out of the marketplace. As a result, smartphone patent litigation is far too common. Apple – a smartphone pioneer – has participated in scores of lawsuits around the globe as both a defendant and plaintiff. As a plaintiff, Apple sometimes uses its patents opportunistically to hinder innovation by its rivals. For example, Apple sued Samsung using a patent that claimed the slide-to-unlock feature on a phone as Apple’s invention. Despite strong evidence that inventors before Apple had already accomplished the key steps to implement this feature, Apple convinced the courts that their version of this feature was patentable, and after seven years Samsung agreed to pay license fees to Apple to settle the case. Economic research suggests that these litigation costs and license fees burden innovative firms to such a degree that on balance the patent system discourages innovation. In other words, innovative firms gain a benefit from their patents on their new technology, but that benefit is more than offset by the many patents owned by others that might be asserted against the new technology.

The past several years have seen significant changes for intellectual property law in the United States. The most important change has been the implementation of IPR petitions to challenge the validity of issued patents. While it does not appear that IPR petitions have supplanted traditional patent litigation in district courts, these petitions appear to be filed concurrently with a large number of traditional litigations. The Patent Office and courts have been busy refining the IPR process with rule-making and opinions. Although initially the IPR process appeared favourable only to patent challengers, the current statistics show more patent claims surviving the process. However, the Supreme Court may radically alter the IPR process in two cases currently pending before it.

IV ENFORCEMENT OF RIGHTS Possible venues for enforcement There are two possible venues for enforcement of patents. First, infringement actions may be brought before the courts of the 94 federal districts that are spread throughout the United States and Puerto Rico. Second, in order to prevent importation of infringing goods, an investigation can be commenced by, and at the discretion of, the International Trade Commission (ITC), an administrative agency that sits in Washington, DC. ii Requirements for jurisdiction and venue An infringement suit may be brought only in a federal district where the defendant resides or, for businesses, has been incorporated. Foreign defendants may be sued in any federal district so long as personal jurisdiction can be found in the United States. The ITC, on the other hand, does not depend on personal jurisdiction over the accused party but has in rem jurisdiction over the accused infringing imported goods. iii Obtaining relevant evidence of infringement and discovery Liberal discovery rules in the United States allow for a significant amount of pretrial discovery under a broad concept of relevance. Discovery is available from the opposing party through a variety of vehicles, including mandatory disclosures, requests for documents and things, interrogatories, depositions of witnesses and expert reports. Discovery is also available from non-parties by a subpoena for documents or deposition. One of the biggest drivers of the cost of discovery is electronic discovery, leading some courts to propose limits on the number of custodians or search terms that need to be searched. For ITC investigations, both party and non-party discovery are also available but must be produced faster given the compressed time frame of such investigations. iv Trial decision-maker Patent infringement actions may be tried before either a federal judge or a jury. In order to be eligible for a jury trial, there must be a claim for monetary damages. Because federal judges do not exclusively handle patent cases, some may have little experience with patent matters. In the case of ITC investigations, an experienced administrative law judge who specialises in patent cases will typically decide the matter. Structure of the trial A patent infringement trial will usually occur in two phases. The first phase determines infringement, validity and unenforceability, and the second phase, if needed, addresses damages. Evidence is presented through fact and expert witnesses, and must be admitted into the record to be considered by the judge or jury. The Federal Rules of Evidence determine what information can be included in the record. Demonstrative evidence may also be used at the trial in the form of summaries and diagrams. At trial, the patentee must prove infringement with evidence meeting the preponderance standard (i.e., infringement is more likely than not). For an accused infringer to prevail on its invalidity or unenforceability defences, it must prove its case with evidence meeting the more rigorous, clear and convincing standard because of the statutory presumption of patent validity. Prior to trial, disputes regarding patent claim interpretation are typically decided by the judge in a process called a Markman hearing. This process can become essentially a miniature trial before the judge and may include witness or expert testimony. Although the outcome of the Markman hearing may determine the eventual result in the case (so that, if the appellate court ultimately decides it was erroneous, the case may have to be retried), the Federal Circuit has consistently refused to review it as an interlocutory matter. vi Infringement Infringement may be direct or indirect. For direct infringement, all of the elements of the claim must be present in the accused product or method. Should a product not literally contain every element, the missing element may be shown using the doctrine of equivalents. The doctrine of equivalents is limited to equivalents that were not foreseeable at the time the patent claim was amended, and cannot include equivalents that are described in the specification and not claimed or were distinguished during patent prosecution. Infringement may also be indirect. Indirect infringement includes inducement and contributory infringement. For indirect infringement to be proven, evidence must show that there is direct infringement and that the indirect infringer is either inducing that infringement or is contributing to that infringement. viii Time to first-level decision For the federal district courts, the median time to trial for patent infringement actions is about two to two-and-a-half years, although this varies widely by federal district. If the case is tried before a jury, the jury will deliberate after all of the evidence is presented and will render its verdict, usually very promptly, thereafter. For a trial before a judge, the time needed to review all of the evidence and prepare the decision can take significantly longer, from months to over a year in some cases. ITC investigations are typically significantly faster, with the time from complaint to a final determination usually taking about a year to a year-and-a-half.

In the most direct mechanism, courts foreclose a particular cause of action, effectively withdrawing all judicial resources from that category of litigation so that they can be reserved for other, "more important" cases. This occurs, for instance, when courts embrace the "floodgates" argument, the idea that

certain sorts of law suits should not be allowed because to do so would 'swamp' the courts with litigation. The court supposes that if it were to allow that type of suit it would lack the time to consider promptly enough other law suits aiming to vindicate rights that are, taken together, more important than the rights it therefore proposes to bar.

For the past several decades, appellate judges and court scholars have struggled with the existence of, and potential responses to, an overwhelming predicament: the growth of the federal appellate caseload has far outpaced the growth of the federal appellate judiciary. 1 The average annual filings per active judgeship stood at 73 in1950, 2 jumped to 137 by 1978, 3 and hovers around 330 today 4 —even though the number of judgeships has nearly tripled in that time. 5 No wonder then that the phrase “crisis in volume” was coined, 6 and is now commonly used, 7 to describe the workload of the federal courts of appeals. To cope with their expanding caseload, appellate courts have adopted a wide array of case management techniques. 8 These include holding oral argument 9 and publishing opinions in a smaller percent- age of cases, 10 and having staff attorneys prepare dispositions in a larger percentage of cases. 11 Although specific practices vary considerably from circuit to circuit, 12 their animating rationale is the same: to keep the courts running, some sets of cases must receive considerably less judicial attention than others. 13

Other things constant, an expansion in intellectual property protection is likely to increase the number of legal claims that allege violations of these expanded rights. Thus, an increase in the legal protection of intellectual property should generate increased litigation in trademarks, copyrights, and patents. It is undisputed that intellectual property protection has grown in the last fifty years or so, particularly since the passage of the 1976 Copyright Act. In a recent paper, Richard Posner and I estimated the expansion in intellectual property rights by the increase in the number of words in the principal intellectual property statutes, because most of the statutes expand intellectual property rights—or create new ones—rather than reduce existing rights. In the 1946-2000 period, Posner and I estimated that the copyright statute increased by nearly eleven-fold (from 11,550 words in 1946 to 124,300 words in 2000); the patent statute increased by more than four-fold (from 24,565 words in 1946 to 110,880 in 2000); and the trademark statute (the Lanham Act) by more than two-fold (from 10, 640 words in 1946 to 24,750 words in 2000). This translates into a 4.4% annual rate of growth for the copyright statute, a 2.9% rate for the patent statute; and a 1.4% rate for the trademark statute. We also showed that these increases were not continuous, but typically coincided with major statutory changes, such as the new Copyright Act in 1976; the Trademark Revision Act in 1988, and amendments to the Copyright Act in 1998 concerning digital copying and the copyright term.

A recent paper by two Chinese academics suggests that China is now widely recognized as a formidable contender in the realm of emerging and disruptive technologies (EDTs), particularly in the field of artificial intelligence (AI). The communist nation, according to the researchers, has established itself as a global frontrunner in the practical implementation of AI across various sectors, including facial recognition, speech technology, financial technology (fintech), research publications, patents, and deep learning. Notably, China's advancements in AI hold significant implications for military applications, including nuclear warfare. As other researchers have warned, the Chinese Communist Party’s (CCP) refusal “to keep autonomy out of its nuclear command and control” should concern nations across the world. In an extensive essay, published on January 16 of this year, Paul Scharre and Michael Depp, two leading analysts at the Centre for New American Security, note the many ways in which the marrying of AI and nuclear weapons could prove to be catastrophic. “Improperly used,” they caution, “AI in nuclear operations could have world-ending effects.” Although “the United States has been extremely transparent and forward-leaning in establishing and communicating its policies on military AI and autonomous systems,” they write, other countries - like the aforementioned China and Russia - have not. The increasing adoption of AI and automation in militaries, add the analysts, has raised concerns about their use in nuclear operations. Without formal agreements, and more transparency from the likes of China and Russia, there is a risk of a gradual shift towards an avenue that could undermine nuclear stability. Scharee and Depp discuss Russia's Perimeter “dead-hand” system, which was developed by the Soviet Union in the 1980s as a failsafe in the event of a decapitation strike. It consists of a network of sensors that detect nuclear attacks. If these sensors are triggered while Perimeter is active, the system waits for a signal from senior military commanders. If no signal is received, indicating a successful decapitation attempt by the enemy, Perimeter bypasses the normal chain of command and grants nuclear launch authority to a junior officer on duty. Russian officials have confirmed that the system is still operational and has undergone significant improvements in recent years. The purpose of Perimeter was to alleviate the pressure on Soviet and Russian leaders to make hasty nuclear decisions with limited information. However, with the addition of AI it’s possible the system could become even more terrifying. The inclusion of autonomous systems in nuclear launch decisions poses a threat to stability, contend the authors. Moreover, AI increases the potential for nuclear accidents. The absence of human judgment during critical stages of nuclear operations could lead to errors or an increase in false alarms, thereby increasing the risk of nuclear incidents.

The United States has yet to show up to this information fight. U.S. cyberstrategists prioritize defending physical infrastructure—routers, servers, and endpoint devices such as laptops and smartphones—but consistently underestimate the economic and political significance of the information carried on that infrastructure. The U.S. private sector, which rarely acts in the American national interest, is primarily responsible for protecting data and information platforms. It can be tough to craft regulations and national security policies for data and technology that do not fall afoul of democratic and capitalist values. A national information strategy can sound, or indeed become, Orwellian without the right political leadership. But to thrive in the twenty-first century, democracies must now put information at the center of domestic, security, and foreign policy. MARSHALING DATA Russia’s interference in U.S. elections may be the most well-known case of a state using sensitive political and social information to attack another, but it is by no means the only example. In 2014, Chinese hackers stole the personal information of more than 22 million people connected to U.S. security clearance processes. That breach not only demonstrated the vulnerability of U.S. government systems to hostile actors but also provided China with a tremendous resource. China could use the stolen data to profile and target U.S. officials and their families. It could also use the “clean” and highly structured data to train the algorithms that power military-related artificial intelligence projects, sharpening the weapons in its cyberwar arsenal. For similar reasons, China sponsors the theft of commercial data. Security officials believe that the data breach that rocked credit reporting giant Equifax in 2017—in which hackers stole the financial and personal information of approximately 147 million Americans—originated in China. That stolen U.S. information can make up for gaps in China’s own datasets. Although consumers provide plenty of data to Chinese e-commerce giants, China’s financial and health-care industries are threadbare compared to their U.S. counterparts. The Equifax operation provided the Chinese with detailed, organized information on nearly half the American population. Neat, easily searchable data, however acquired, is vital for companies to train artificial intelligence applications for both military and commercial use. No advanced economy boasts a national artificial intelligence strategy more mercantilist than China’s. Beijing seeks to lead the world in this area by 2030. Xi’s doctrine of civil-military fusion yokes together the research efforts of private corporations, China’s military industrial base, and the intelligence agencies of the People’s Liberation Army. China subscribes to the zero-sum ethos of data competition; states want to acquire data from others, while limiting access to their own data. To that end, China, India, Russia, and other countries have introduced tough data localization laws to prevent companies from taking certain types of information across national borders. In the private sector, access to data tends to create a virtuous cycle: more data lets companies build better applications, which makes them more profitable and allows them to harvest and monetize even more data. The dynamic explains why information giants, like Alibaba, Amazon, Facebook, Google, and Tencent are monopolies. But this commercial logic also prompts states to think in zero-sum terms. The United States should prepare for the prospect of authoritarian regimes ramping up cyberattacks against the databases of its major companies. If China, for instance, wanted to boost its own national champion as the dominant player in a particular market, even a minor cyberattack against a rival firm could be very damaging. Stealthily disrupting or “poisoning” inputs through a data-integrity attack or by manipulating the quality-control algorithms that govern industrial processes and the delivery of products could result in inferior, potentially harmful outcomes. Before an affected company even detects this kind of attack, it could face numerous economic, legal, and reputational problems, losing consumers and becoming dislodged from supply chains. Data-driven machine learning systems now assist decision-making in many areas of the private and public sectors. A data integrity attack does not even need to be successful to be catastrophic. The perception that an organization’s data is poisoned and its decision-making processes corrupted could reduce public trust in a company, in a government agency, and even in democratic process. Since the Internet was introduced in China in the 1990s, China has developed the world’s most sophisticated system of network control, censorship, and propaganda, and has maintained heavy restrictions on market access for foreign media and technology companies. Russia is following China’s lead. During recent protests in Moscow, authorities jammed mobile Internet access in the city to keep protestors from communicating with one another. In a more ambitious gambit, Russia plans to test a procedure that would altogether disconnect Russian Internet users from the global Internet. Even the United Kingdom’s Ministry of Defense has warned that democracies may need to step back from a commitment to the open Internet and consider “national or regional cyber borders.” Such moves would unwind decades of economic liberalism, fragment the relative openness of current information flows, and give rise to suspicion and animosity. Unless the United States can chart a different course, the world might stumble down this path.

The Seventh Circuit may have said the quiet part out loud: "The standing rule reduces the workload of the federal judiciary...." (205) While literally true, that is neither an accepted nor acceptable justification for standing doctrine. (206) The existence of a "Case" or "Controversy" does not depend on how pleased the court will be to hear it. Article III does not give the judiciary the ability to set its caseload. And a dispute will be no less adversarial between the parties if it is the tenth hearing on a court's calendar rather than the first. Indeed, around the time that the Supreme Court threw standing jurisprudence for a loop in Clapper, it admonished that federal courts are obligated to hear cases falling within their jurisdiction. (207) Yet the appeal of the standing dismissal to the federal judge is impossible to ignore. The massive and increasing caseload in the federal courts is well known. In the data-protection context, cases typically promise to be long and involved. Most arise from large-scale data breaches or inappropriate, indiscriminate data collection. As such, they are usually brought as class actions and implicate complex technical issues. One can certainly understand why overworked judges might be eager to get such cases off their dockets. With that in mind, dismissal for lack of standing has a special appeal over other means of dispensing with a case: it is a non-merits dismissal. Even if erroneous or improperly motivated, a dismissal for lack of standing works relatively little injustice, as its effect is only to banish the suit to state court. Moreover, as the data show, state-law and common-law claims predominate in data-protection litigation--the majority of cases analyzed involved such claims, while fewer than half of cases asserted claims under federal law. Allowing state courts to deal with state-law claims is a hallmark feature of the federal system, CAFA notwithstanding. The Northern District of Illinois' BIPA series of cases may proceed from this impulse. Most were brought initially in state court, then removed to federal court by defendants who seemed intent on challenging plaintiffs' right to bring suit at all. Confronted with the prospect of overseeing a lengthy, complicated class-action lawsuit brought by plaintiffs who did not want to be in federal court in the first place against defendants who appeared to be engaging in gamesmanship, the judges of the Northern District of Illinois may well have viewed dismissal for lack of standing as the best of all possible worlds. The data are potentially consistent with this explanation. In particular, appellate courts have been significantly more solicitous [so-liss-i-tus] of plaintiffs in data-protection cases, finding standing in 75% of cases, compared to only 51% of district courts. In data-breach cases, which likely present both the most technically complicated inquiries and the most diffuse harms, appellate courts upheld standing in 79% of cases, compared to a mere 41% of district court cases.

The greater the degree of uncertainty about legal precedents and the outcome of a trial, the more occasions there will be for p* > p* and the greater the number of trials that will take place relative to settlements. Alternatively, if the legal rules are clear and the parties agree on the relevant precedents and facts, p and p are not likely to differ significantly, so a settlement will result. Although no direct data on legal uncertainty exists, the passage of new laws and amendments to existing laws—particularly if they create new or expanded rights as opposed to clarifying existing laws and regulations are likely to be associated with greater legal activity. In the area of intellectual property, for example, we expect that statutory activity will increase the number of lawsuits because the statutes will cover a greater range of subject matter and activities. Less obviously, trials should comprise an increasing proportion of terminations because greater legal uncertainty, until a sufficient body of precedents interpreting the new statutes is established, increases the relative number of disputes in which the degree of mutual optimism (p* - p*)J, is greater than the costs of going to trial. An expansion in the level of economic activity associated with the creation of new intellectual property is likely to increase the number of potential disputes that have a positive expected value and hence the number of lawsuits and trials.

EXPANSION OF WAYS TO PROVIDE SHORT AND LONG-TERM RELIEF FOR OVERBURDENED COURTS To alleviate the burden on overburdened courts, a strategic plan by the Federal Judiciary focuses on expanding ways to offer both short and long-term relief. This involves innovative policies developed due to the impact of the COVID-19 pandemic, aiming to keep courts open and ensure equitable access to court-related services. Embracing technology has been pivotal for courts, not only in staying operational but also in advancing participation rates and facilitating efficient dispute resolution. These strategies are crucial steps toward achieving more effective court operations, relieving congestion, and improving overall fairness in judicial processes. Moreover, enhancing access to justice initiatives - such as improved evidence management and increased resources for conflict-free representation – play an essential role in this expansion effort. STRATEGIES FOR EFFECTIVE AND EFFICIENT MANAGEMENT OF PUBLIC RESOURCES Improving the use of judges and implementing are key strategies for maximizing public resources in civil courts. By streamlining processes and ensuring efficient allocation of resources, courts can better serve their communities and address the backlog of cases.

“We’re still waiting to see what proposals actually become law, but the massive number of A.I. bills introduced in states like California shows just how interested lawmakers are in this topic,” he said. And the state proposals are having a ripple effect globally, said Victoria Espinel, the chief executive of the Business Software Alliance, a lobbying group representing big software companies. “Countries around the world are looking at these drafts for ideas that can influence their decisions on A.I. laws,” she said. More than a year ago, a new wave of generative A.I. like OpenAI’s ChatGPT provoked regulatory concern as it became clear the technology had the potential to disrupt the global economy. U.S. lawmakers held several hearings to investigate the technology’s potential to replace workers, violate copyrights and even threaten human existence. The OpenAI chief executive Sam Altman testified before Congress and called for federal regulations roughly a year ago. Soon after, Sundar Pichai, chief executive of Google; Mark Zuckerberg, chief executive of Meta; and Elon Musk, chief executive of Tesla, gathered in Washington for an A.I. forum hosted by the Senate majority leader, Chuck Schumer, Democrat of New York. The tech leaders warned of the risks their products presented and called for Congress to create guardrails. They also asked for support for domestic A.I. research to ensure the United States could maintain its lead in developing the technology. At the time, Mr. Schumer and other U.S. lawmakers said they wouldn’t repeat past mistakes of failing to rein in emerging technology before it became harmful. Last month, Mr. Schumer introduced an A.I. regulation road map that proposed $38 billion in investments, but few specific guardrails on the technology in the near term. This year, federal lawmakers have introduced bills to create an agency to oversee A.I. regulations, proposals to clamp down on disinformation generated by A.I. and privacy laws for A.I. models. But most tech policy experts say they don’t expect federal proposals to pass this year. “Clearly there is a need for harmonized federal legislation,” said Michael Karanicolas, executive director of the Institute for Technology, Law and Policy at the University of California, Los Angeles. State and global regulators have rushed to fill the gap. In March, the European Union adopted the AI Act, a law that curbs law enforcement’s use of tools that can discriminate, like facial recognition software. The surge of state A.I. legislation has touched off a fierce lobbying effort by tech companies against the proposals. That effort is particularly pronounced in Sacramento, the California capital, where nearly every tech lobbying group has expanded its staff to lobby the Legislature. The 30 bills that were passed out of either the Senate or Assembly will now go to various committees for further consideration before the Legislature ends its session later this summer. Democrats there control the Assembly, Senate and governor’s office. “We’re in a unique position because we are the fourth-largest economy on the planet and where so many tech innovators are,” said Josh Lowenthal, an Assembly member and Democrat, who introduced a bill aimed at protecting young people online. “As a result, we are expected to be leaders and we expect that of ourselves.” Three of those bills are designed to protect actors and singers, living or dead. SAG-AFTRA, the union for actors and other creators, helped write a bill that would require studios to obtain explicit consent from actors for the use of their digital replicas. The union said the public has expressed strong support for intellectual property protections after high-profile conflicts between actors and A.I. companies. Last month, for instance, the actress Scarlett Johansson accused OpenAI of copying her voice to develop a voice assistant without her permission. OpenAI has denied the accusation. California is an important battleground because “the Legislature tends to be progressive and believes strongly in consumer protection and worker rights,” said Duncan Crabtree Ireland, the strategic and creative chief negotiator for SAG-AFTRA. “But it is also where five of the six biggest A.I. companies in the world are based.”

In this federal legislative vacuum, states are emerging as today’s AI regulators. Some of the laws passed include measures to protect consumer data privacy (OR SB-619), build institutional understanding of AI (LA SCR49), prevent election interference (MI HB-5144), and establish state AI task forces (IL HB-3563), offices (CT SB-1103), and advisory councils (TX HB-2060). Recently, Utah passed a broader law (UT SB0149) establishing liability, notice of interaction requirements, and an Office of AI Policy. However, in this emerging patchwork, one state is uniquely positioned to have a crucial impact on AI governance: California. The promise of California as an AI regulator Perhaps California’s efforts to influence AI policy, particularly, for frontier systems, is due to its status as an AI powerhouse, its large economy, and its Democratic penchant for regulation. The Golden State is home to 32 of Forbes’ top 50 global AI companies and leading frontier players such as OpenAI, Anthropic, Meta, xAI, Google, and Microsoft, among others. OpenAI kickstarted the generative AI race with its release of ChatGPT back in 2022, and its GPT-4 large-language model is still the top preference for AI users more than a year after its release. Anthropic is one of OpenAI’s major competitors; their recently released Claude 3 model outperforms GPT-4 on many standard performance benchmarks. Meta, the parent company of Facebook and Instagram, is a market behemoth that produces open models—those whose inner workings are released to the public for widespread inspection, modification, and execution—and it just introduced its long-awaited Llama 3 model. The presence of such major AI developers makes California an attractive jurisdiction for advancing responsible AI policy. Furthermore, California produces 14.5% of the United States GDP, and if it were a sovereign country, it would have the world’s fifth- or sixth-largest economy behind the U.S. (without California), China, Japan, Germany, and possibly India. California’s dual role as an AI hotspot and a powerful economy might enable their state AI regulation to accomplish many of the benefits of standard-setting legislation, such as requiring responsible development and deployment of frontier systems, mitigating theft of powerful dual-use models by malicious actors, and ensuring legal liability for AI harms. A newly introduced bill provides a model for such regulation and may garner many of these benefits for Californians and, incidentally, everyone else. In February 2024, State Senator Scott Wiener (D-San Francisco) introduced the Safe and Secure Innovation for Frontier Artificial Intelligence Models Act. The bill’s primary aim is to mitigate large potential risks posed by future frontier models, including the automation of large-scale cyberattacks and the production of novel biological weapons. It would address these risks by requiring the developers of a model trained using large amounts of computing power to demonstrate the safety of their system. If they were unable to make such a demonstration, they would be subject to additional requirements, including submitting yearly certifications of compliance with safety standards, implementing emergency shutdown protocols, and strengthening cybersecurity protections to prevent unauthorized access. The bill also includes several other key provisions to increase safety. It would require providers of computing clusters to verify the identity and goal of customers seeking large amounts of computing power for AI training. It would also mandate that developers report safety incidents, including hazardous use and model theft. The state attorney general would be empowered to prosecute AI-related damages and threats to public safety. Additionally, the bill would establish a subsidized public computing cluster to democratize access to expensive AI training—cutting-edge AI-specialized computer chips cost tens of thousands of dollars. To oversee all these requirements, the bill would create the Frontier Model Division within the Department of Technology. Through its ambitious and broad provisions, this bill, if passed, could capture the promise of AI concentration in California. Not only might California policies, like Senator Wiener’s, be able to achieve many of the benefits of national legislation, but passing such policies may also be easier in California than at the federal level. First, California’s government is overwhelmingly Democratic and therefore less subject to gridlock and polarization than is the case nationally. Governor Gavin Newsom is a Democrat and the state Assembly and Senate have wide Democratic margins. The absence of partisan-based gridlock makes it easier to pass laws and regulations in California than places with greater party competition. Second, California has a strong track record of passing groundbreaking technology regulation, such as an Internet of Things security law and the California Consumers Privacy Act (CCPA). These considerations may make it easier to advance AI policies in California than at the federal level.

A. OTHER TEXT AND DATA MINING (TDM) EXCEPTIONS As the benefits of AI technologies have become apparent, countries have amended their copyright laws to promote innovation and remain competitive in the AI/ML space.175 The United States already has comparably strong protections for data miners through the fair use doctrine.176 However, a general safe harbor for data mining would further the spirit of these protections by clarifying the interaction of copyright and AI for smaller innovators.177 Other jurisdictions have recently enacted text and data mining (TDM) exceptions that can provide guidance here. Japan was the first country in the world to update its copyright laws to include an exception for text and data mining.178 Article 47(7) was introduced in 2009 and authorized TDM by all users for all purposes, whether commercial or non-commercial.179 In line with Prime Minister Shinzo Abe’s objective of promoting AI and Big Data industries, the 2018 Amendment to the Copyright Act later introduced three provisions to clarify the law and remove perceived copyright barriers to AI: (1) Article 30-4 authorizes users to “analyse and understand copyrighted works for machine learning;” (2) Article 47-4 “permits electronic incidental copies of works;” and (3) Article 47-5 “allows the use of copyrighted works for data verification.”180 Japan’s three new copyright provisions that specifically reference acts relevant to the machine learning process demonstrate a national commitment to the flourishing of AI industries. Similarly, the United Kingdom has also enacted an exception for TDM purposes in its copyright laws.181 However, its exception is markedly narrower than Japan’s. Under § 29A of the Copyright, Designs and Patents Act 1988, copies made for TDM analysis do not infringe copyright, provided that the work is carried out “for the sole purpose of research for a non-commercial purpose.”182 The beneficiaries of the exception are also limited to those who have “lawful access” to the work in question.183 While providing legal certainty for researchers, the U.K. exception leaves out startups and entrepreneurs who aim to commercialize innovative machine learning technologies. Recognizing the legal uncertainty experienced by researchers who utilize TDM and a need for a harmonized exception among its member states, the European Union recently introduced two exceptions for TDM in Directive 2019/790/EU.184 Article 3 permits “reproductions and extractions made by research organisations and cultural heritage institutions in order to carry out, for the purposes of scientific research, text and data mining of works or other subject matter to which they have lawful access.”185 Article 4 confers an exception for “reproductions and extractions of lawfully accessible works and other subject matter” for commercial TDM uses but allows rightsholders to opt-out of the exemption.186 This limitation imposed on commercial TDM uses in Article 4 has been criticized for “effectively creat[ing] and legitimiz[ing] a derivative market for text and data mining, which right holders may wish to control, license[,] or even entirely prohibit.”187 While the TDM exceptions in the Directive recognize the importance of data mining in research and technology, the clear limitations in Article 4 may put commercial AI developers in the European Union at a disadvantage. Not to be left behind, other countries are also considering changes to their copyright law to provide legal certainty and encourage innovation and research in data mining.188 Australia is considering several TDM-related exceptions as part of an overall fair use exception.189 Canadian scholars have advocated for a fair use regime or a specific TDM exception for commercial uses.190 And Singapore is expected to move forward with proposed TDM exceptions enabling broad uses for both commercial and non-commercial contexts.191 As countries amend and clarify their laws to allow technological flourishing, the United States should also consider revisiting its copyright laws.

Both of the traditional rights regimes involve inefficiencies: (CP) potentially restricting the use of original content and both regimes trading off the incentives of the OCs and AI. The question is whether a different mechanism can create a more favourable balance of incentives for both original content investment and AI training quality while preserving the use of original content in consumption and AI training. Recall that, in the small AI model context, the AI provider would agree not to use content if the potential harm from leakage of that content were high. Of course, here, the extent to which leakage might be harmful to individual content providers is not known ex ante. It is, however, known ex post. This suggests a potential mechanism that involves liability assessed ex post is feasible. That mechanism would proceed as follows: 1. If ρi(1) < Γ, the AI provider is not liable for any damages from the use of i’s content. 2. If ρi(1) ≥ Γ, the AI provider must pay full compensation to i of an amount D(ρi(1)) = 1 4ρi(1)ˆxFU Here (FU) stands for ‘fair use’, and ˆxFU is the chosen content quality under that regime. The level of damages, if triggered, allows the content provider to earn 1 4 ˆxFU as if the leakage never occurred.11 The following proposition characterises the outcomes under (FU) and compares it to traditional rights regimes. Proposition 4 Suppose that AI earns non-negative profits under the (FU) mechanism. In the large AI model, for each OC, i, ˆxi,CP ≥ ˆxi,FU ≥ ˆxi,NC and for AI, ˆyFU ≥ ˆyCP ≤ ˆyNC. Using the specific function forms, expected social welfare under (FU) will exceed that under (CP) if and only if: γ > E[ρ(1)|Γ]2 + E[ρ(1)|Γ] − E[ρ(1)](1 − E[ρ(1)|Γ]) 4(1 − E[ρ(1)|Γ]) where E[ρ(1)|Γ] = � Γ 0 ρi(1)dF(ρi(1)). The proof (in the appendix) compares the first-order conditions for each type of investment and follows on by calculating social welfare under (FU) and compares it to those under traditional rights regimes. The condition on AI profits arises because it is possible that Γ is such that: u(ˆyFU) − cAI(ˆyFU, ˆxi,FU) < (1 − F(Γ)) � 1 Γ 1 4ρ(1)ˆxi,FUdF(ρ(1)) in which case AI would not be finally viable because expected damages are above its profits from AI. Below, when we examine what Γ is chosen by the planner, this feasibility constraint needs to be satisfied. The (FU) mechanism has two advantages over traditional rights regimes. First, under the (FU) mechanism, AI chooses to use all content in training the AI (i.e., s = 1).12 Thus, the prices for content and the AI have the same structure as under (NC) but, importantly, in terms of the consumption of that content, all consumers will do so (potentially) with half of the market purchasing from OC and the other half relying on the AI product. Thus, the (FU) mechanism allows content to be used at its socially desirable level for AI training and also to be consumed relatively widely. Second, this has an impact on incentives. As all content is used in AI training, AI training quality under (FU) is the same as that under (NC) for fixed content quality and, therefore, highest among the three regimes compared as OC content has a higher quality under (FU) than (NC). OC content is higher than under (NC) because if potential harm from AI training exceeds a certain threshold, each OC expects to receive a damages payout. Thus, from their perspective, while under (NC) their potential appropriation is 1 4(1 − E[ρi(1)]) for each unit of xi, under (FC) it is F(Γ) 1 4 � 1 − E[ρi(1)|Γ] � + (1 − F(Γ)) 1 4, a higher marginal return as E[ρi(1)|Γ] ≤ E[ρi(1)] (with an equality if Γ = 0). These advantages hold for any Γ < 1. But what level of Γ would be optimal? Note that expected social welfare is higher as Γ is reduced, and there is a lower threshold of harm that can trigger compensation. This has the effect of raising both content provider incentives and AI training incentives. The latter effect arises because, regardless of Γ, all content is available for AI training. Moreover, a lower Γ does not change the total number of consumers who consume original content (that is, 1 2 + 1 2E[ρ(1)]). Note that as Γ → 0, ˆxFU → ˆxCP and ˆyFU → 1 + γ 1 4. This is feasible if AI profits (u(ˆyFU) − cAI(ˆyFU, ˆxFU) − � 1 0 1 4ρ(1)ˆxFUdF(ρ(1))) remain positive. Compared with (CP), Γ = 0 behaves like a regime where all content is copy-protected. However, it is permissive of infringement with compensation, which generates the benefits of original content for AI training without diminishing the incentives of original content providers to invest in quality. In effect, the difference is akin to a prohibition versus a Pigouvian-like tax on the use of original content where external effects are internalised through ex-post prices. In practice, an (FU) mechanism is respectful of transaction costs associated with asserting copyright protection when limited harm occurs. For this reason, it also rationalises the human scenarios explored earlier in that those scenarios are very unlikely to cause commercial harm to original content providers, even in the aggregate. 5 Conclusion There is a sense in which the copyright challenges associated with generative AI are not that different to those that arose with digitisation and the rise of the Internet. In each case, a new technology lowered the cost of utilising original content while enhancing its potential scope or value. For generative AI, this arises because such AI can be trained on original content but also, precisely because that content is widely available, reproduce that content within AI products themselves. This threatens existing business models for original content providers that, not surprisingly, they are protective of. Copyright law played a role in the evolution of digital technologies. However, one thing that was preserved was long-standing copyright protections, including the ability to prevent use except where that use was regarded as fair. This paper has analysed generative AI, taking into account the particular role of such technologies as well as how generative AI products threaten original content creators’ commercial activities. It is demonstrated that different rights regimes can lead to distinct social welfare outcomes, but these rest firmly on the ability of original content providers and AI providers to negotiate over the use of copy-protected material in AI training. When they can negotiate – something that requires knowledge of what content is proposed to be used in AI training as well as what potential harm to original content providers’ commercial activities it might cause – then traditional full copyright protection leads to stronger incentives to produce original content and that this determined the social welfare properties of such protections. The conditions that allow for this, however, require that the AI models be small in the sense of using and relying on a relatively small corpus of content. Such “fine-tuned” or RAG models are currently being developed. The value of copyright protection, however, rests on the potential for commercial harm. If that potential is low, as it arguably is for similar human uses of content in their own “training,” then a strong protective regime may not be of much social value. Consequently, when the AI models are large in the sense of using an almost unimaginable amount of content in training, the type of rights regime imposed can have important social welfare consequences. Because negotiations are prohibitive as it is hard to identify copyright use, let alone determine clearly whether copy-protected content has been used in training, copyright protection, while potentially strengthening incentives to create original content, serves to limit its use both for consumption and training, which involves adverse social consequences. No copyright protection does not involve those costs, but if the threat to original content providers’ commercial activities is high, it could undermine the production of original content and, in the process, its uses. However, if the threat to those commercial activities is low, as it arguably is for equivalent human uses of content, then a permissive regime for the use of content in AI training is socially desirable. For that reason, this paper identifies an alternative mechanism based on fair use. That mechanism permits the use of copy-protected content in AI training but subjects AI providers to damages should the realised commercial harm of content providers be large. Thus, the purpose here is to provide some insurance. The presumption is that content used in AI training is, for the most part, not likely to damage original content providers’ commercial interests. However, should it turn out to do so, they will receive compensation. Insured against such risks, the original content providers’ incentives to create content are undiminished. This represents a practical way of respecting copyright ownership while also allowing its large-scale use in AI training in a way that minimises transaction costs that might arise from injunctive legal action or license negotiations. It is also conceivable that copyright holders could opt out of this type of regime, although the technical requirements to do so would likely require further development as it would require content used to be traceable. When a new technology arises that creates copyright challenges, a clarification of the rights regime can often lead to institutional and technological developments to “make it work.” This happened for music rights with radio and public broadcasting in the form of collecting societies. For AI, Besen (2023) argues that the large volume of data lends itself towards collecting copyright payments by collective societies as seen in other domains such as music. He believes that this would have to be established by government regulation. However, in other areas, the developments were technological. For instance, Google developed a rights management system for YouTube that allowed people to post content with copy-protected materials but for rights holders to be notified and then to utilise a revenuesharing arrangement (Gans (2015)). By settling on a policy approach to generative AI, even if some inefficiencies remain at present, these may incentivise new ways of minimising those inefficiencies.

DMCA Safe Harbor. The Digital Millennium Copyright Act (“DMCA”) is a U.S. law created to address digital copyright issues that came about with the advancement of technology. The DMCA safe harbor provisions protect online service providers from legal responsibility for copyright infringement claims. DMCA protections might vary depending on a number of considerations, but we emphasize that they are not guaranteed for all model deployments. We examine several of these considerations here. DMCA protections for generative foundation models are uncertain. At first glance, it may seem like the Digital Millennium Copyright Act (DMCA) would protect machine learning model hosts, at least when users upload the model to be hosted (e.g., model repositories with a user interface, like HuggingFace). Like in other hosted sites, they would need to meet the relevant requirements like using a registered agent under DMCA §512(c)(2). Then they could put up a take-down request form and add filters for the offending model (or model output) when served with a take-down request under the DMCA §512(c) safe harbor.26 An internet company that has a notice-and-takedown scheme in place is not liable for hosting infringing content posted by a third party. But it is not obvious that the DMCA safe harbors apply to generated content. For example, in 2019, Amazon lost on a motion to dismiss when its algorithms selected copyrighted material to host on the website (Williams-Sonoma, Inc. v. Amazon.com, Inc., N.D. Cal. 2021, Order on Motion to Dismiss). The court was unconvinced that Amazon was eligible for safe harbor under the DMCA. They stated that to establish safe harbor the content must be stored “at the direction of the user” (Mavrix Photographs, LLC v. Livejournal Inc., 9th Cir. 2017, at 1052). This may mean that generated content does not have the same safe harbor and that post-hoc take-downs are not sufficient to reduce liability.27 As such, filtering of generated content before a takedown request is ever received may be more important while the courts determine the applicability of DMCA to generated content. It might matter where the model and data comes from. This also implies that DMCA protections may vary on who the model host and creator are. For example, a website hosting models uploaded by users might find it easier to argue for DMCA protection because the website itself is not creating or selecting the content (assuming that it follows other requirements like using a registered agent under DMCA §512(c)(2)). On the other hand, if a company were to create and host a model that itself selects content provided by others, like Amazon did in Williams Sonoma v. Amazon, it is unclear whether courts would agree that DMCA protections would apply. Another unclear variation on DMCA eligibility rests on the source of the training data: is the data user-contributed or creator-contributed? This might mean that, for example, DMCA safe harbors might be more likely to apply when a model is refined via Reinforcement Learning from Human Feedback (RLHF) (Ouyang et al., 2022) based on user ratings or updated automatically via user-generated data. These modes of method training and deployment are more akin to users uploading content to YouTube, and might help with arguments for DMCA protections. Things might become more murky if, for example, an RLHF model is trained on user data but this data is modified by creator-hired annotators. If model creators themselves curate and scrape data, then host the model themselves, this might be more akin to the Williams Sonoma case (which, again, is not a settled rule or a binding decision, but shows the variation of outcomes that is possible in the current state of the law). It is unclear what’s the best mechanism for DMCA takedowns with generative models. Even if it applies, it is unclear how the DMCA notice-and-takedown scheme would work as applied to foundation models, but the most likely “take down” approach might actually look more like output filtering with a safe harbor. As we will discuss in §4.2, instance unlearning is a nascent research area and retraining a model without a taken down datapoint could be exceedingly costly. The most likely approach in the near term is to “take down" model outputs that were too similar to a copyrighted work via a filtering mechanism since foundation models generate content on demand for users rather than hosting persistent content that can easily be taken down. But new research is needed to identify new and improved mechanisms for handling takedown requests in this relatively new setting. The Copyright Office has a DMCA anti-circumvention exemption for text and data mining. Finally, the Copyright Office also provides an exemption to DMCA’s anti-circumvention requirements in §1201 in the case of non-commercial data mining.28 This may allow non-commercial researchers to, say, remove digital rights management software to train on video or text content.29 Overall, DMCA protections are far from guaranteed, so model creators and deployers cannot rely on its safe harbor provisions to reduce liability. Instead, they must take a more proactive approach. Moreover, some previously-discussed provisions add to potential liabilities like §1202, creating additional compliance challenges

By expanding the capabilities of models, machine learning researchers and practitioners have stumbled into the muddy waters of fair use. As a result, websites like Getty Images have banned AI-generated content (Vincent, 2022), and lawsuits have been filed against products using foundation models, including against GitHub Copilot, Stable Diffusion, OpenAI’s ChatGPT, Meta’s Llama, and more (DOE 1 v. GitHub, Inc., N.D. Cal. 2022; Andersen et al. v. Stability AI et al., N.D. Cal. 2023; Vincent, 2023; Chabon v. OpenAI, Inc.,, 2023; Kadrey v. Meta Platforms, Inc., 2023). In this work, we shed light on this subject matter for machine learning researchers and highlight that significant additional work is required to de-risk foundation model deployments for generative use cases, focusing primarily on U.S. laws. First, we provide an overview of U.S. case law on the fair use doctrine.7 We draw analogies to foundation model use cases. We supplement these with a review of prior experiments, as well as novel experiments, and illustrate that foundation models can produce content that is sufficiently similar to copyrighted material. Furthermore, the case law suggests that even certain types of transformations of the training data would not be considered fair use. Thus, the risk of infringement is real, and fair use will not cover every scenario where a foundation model is created or used. The exact amount of risk is unclear, and the law will evolve with ongoing litigation. Second, we overview technical mitigation strategies that will reduce this risk in accordance with the current state of the fair use doctrine. Grimmelmann (2015) stated that “paying attention to robotic readership refocuses our attention on the really fundamental questions: what is copyright, and what is it for? To say that human readers count and robots don’t is to say something deep about the nature of reading as a social practice, and about what we want robots—and humans—to be.” Lemley and Casey (2020) suggested that humans and AI should be held to similar standards when it comes to copyright. If this is the case, it is the job of machine learning researchers and practitioners, working together with legal practitioners, to ensure that foundation models create transformative content which would pass muster under the same fair use analysis as provided to a human. To get there, new strategies and techniques will need to be developed, taking steps to ensure that foundation models behave in more transformative and novel ways. We call for more research to align technical mitigation strategies with fair use, including better output filtering mechanisms relying on higher-level semantics and new innovation in training-time techniques like extraction-preventative learning from human feedback. Developing these mitigation strategies is an important research challenge for machine learning and natural language processing and would bring practices in the two fields into better alignment with the law. Lastly, we argue that a co-evolution of technical mitigation strategies and law can help establish a middle ground where the positive impact of foundation models is realized while reducing the harms to data creators’ intellectual property rights. With the current uncertainties of fair use doctrine, as Sobel (2017) and others noted, the law may sway to one extreme or another. On one hand it could lead to overly permissive interpretations of fair use that could allow any generative AI use, disregarding the rights of data creators. Or it could lead to overly restrictive interpretations of fair use that could broadly prevent foundation model training and use, concentrating power among entities that have already acquired vast quantities of licensed data. By developing and deploying strong technical mitigation strategies, it may be possible to lessen the risk of such extreme legal outcomes. And the law should take into account the existence and strength of such technical mitigation strategies. This could involve a multi-pronged approach: considering technical mitigations in fair use assessments, clarifying the status of DMCA protections for foundation models, or developing DMCA-like safe harbors for deployments that use strong technical mitigation efforts, pursuing policy strategies for reducing harms to labor, and more. Realizing this middle ground requires the participation of a much broader community including the data creators impacted by foundation models, technologists, legal professionals, among many others. We encourage more multidisciplinary work to further the co-evolution of law, policy, and technical methods for mitigating intellectual property harms. Overall, the goal of this work is to act both as a guide and call-to-action for ML researchers and practitioners to actively pursue technical mitigation strategies. We hope that this guide helps instill a better understanding that fair use is not a panacea, and that a nuanced comprehension of the legal landscape’s intricacies is vital to effectively navigate potential pitfalls and uncertainties. Furthermore, this work may also prove useful to lawyers and policymakers, providing them with more insight into potential technical details of foundation models, including technical mitigation strategies, and how they might play a role in the developing legal best practices and potential reforms.

Lessons can be learned from the implementation of TDM exceptions in other jurisdictions to craft a safe harbor for data mining in the United States. Such a safe harbor should (1) define “data mining” broadly; (2) clearly allow reproductions, derivations, and dataset-sharing for data mining purposes; (3) allow non-commercial and commercial uses to the same extent; and (4) be limited to functional, non-expressive uses of data mining. A safe harbor should broadly define “data mining” to cover a wide range of pattern extraction techniques, not limited to AI/ML. One potential definition is proposed by Jean-Paul Triaille: “The automated processing of digital materials, which may include texts, data, sounds, images or other elements, or a combination of these, in order to uncover new knowledge or insights” (although he uses the term “data analysis” instead of “data mining”).192 While other jurisdictions refer to “text and data mining,” “text” is redundant, as it is encapsulated in “data.”193 In crafting the safe harbor, what matters is not the exact technique through which patterns are extracted, but the automated process of gaining functional insights from expressive data. While this Note specifically considers data mining in a machine learning and artificial intelligence context, using a broad definition of data mining in a copyright safe harbor would not foreclose future technologies that employ other techniques for pattern extraction. A safe harbor should also be specific enough in the rights that are conferred to data miners to provide legal clarity. Given the intimidating U.S. copyright remedial regime and the unpredictable fair use doctrine, smaller innovators in the United States would benefit from an exception that clearly spells out allowed uses.194 Here, lessons can be drawn from Japan’s recent amendments to clarify its copyright exception, which provide an innovation-friendly and clear implementation of an exception.195 It should be clearly enumerated that reproductions and derivations (for labeling and annotation) of works are permitted in a data mining context. Further, the secure storage, retention, and sharing of datasets should also be permitted for verification purposes. Allowing researchers and regulators the ability to scrutinize how data is being used would promote fairer, safer, and higher-quality AI products.196 Policymakers should also be careful to not impose limitations on commercial uses, such as opt-out mechanisms, that could unduly stifle innovation contrary to the expressed policy of copyright law in the United States.197 Unlike the exceptions of the United Kingdom and European Union, commercial uses should be permitted to the same extent that research and non-profit uses are allowed. The U.S. fair use doctrine already includes a copyright exception for research uses and considers nonprofit uses favorably.198 Thus, the proposed safe harbor would need to specifically free up commercial uses of data mining. Rightsholders should not be allowed to opt-out of uses of their works for commercial data mining purposes because doing so would legitimize a licensing market for machine learning training data that could prompt further legal confusion, put smaller innovators at a competitive disadvantage, and lead to lower quality, biased AI systems.199 Finally, and most importantly, the safe harbor should be limited to functional, non-expressive uses of data mining. Some emerging AI technologies use expressive data to generate expressive works and are not covered within the scope of this Note.200 Because expressive AI-generated works could compete with and effectively replace original works, it would be more difficult to justify a categorical safe harbor for these works without unduly encumbering authors and creators. Instead, the fair use doctrine should be applied on a caseby-case basis for expressive uses of data mining. In sum, a data mining safe harbor should (1) define “data mining” broadly; (2) clearly allow reproductions, derivations, and dataset-sharing for data mining purposes; (3) allow non-commercial and commercial uses to the same extent; and (4) be limited to functional, non-expressive uses of data mining. As the United States strives to maintain its dominance in artificial intelligence,201 it must consider how existing laws enable or stifle technological progress. Copyright law presents a potential barrier for AI growth when machine learning models are trained using expressive data. Fundamentally, data mining is not copyright infringement. While the fair use doctrine provides a degree of flexibility in U.S. copyright law, certainty in the law is desirable for smaller actors. Practically, the fair use doctrine is unpredictable and has been stretched beyond its limits in new technological contexts. Establishing a clear right to use copyrighted materials to train functional machine learning models is consistent with the goals of copyright law and would ultimately remove barriers to innovation. Rather than rely on courts to rule on important issues of technology policy, Congress should be more proactive in enacting a data mining safe harbor. A safe harbor would make legislatively clear what should already be clear given the scope of copyright law: data mining is not infringement at all.

As demonstrated throughout this work, the risk of copyright violation and litigation, even with fair use protection, is a real concern. To mitigate these risks, we recommend that foundation model practitioners consider implementing the mitigation strategies outlined here and pursuing other novel research in this area. There is significant, exciting, technical research required to make technical mitigation strategies robust and aligned with fair use doctrine. We reinforce that machine learning researchers must play a role in providing viable mitigation mechanisms to demonstrate that models are truly covered by fair use. Preventing extreme outcomes in the evolution of fair use law by advancing mitigation strategies. Legal scholars have noted that there might be two extreme outcomes for fair use and machine learning (Sobel, 2017). On one hand, there is a possibility that courts may rule that foundation models are widely acceptable under fair use regardless of the likelihood of infringement or efforts at mitigation, which could have adverse effects on the income of data creators and disregard the ethical and moral rights attached to their work. On the other hand, there is a possibility that courts may declare that generative foundation models cannot be trained on unlicensed copyrighted data in most cases. This scenario could lead to a concentration of power for companies that have retained licenses to large amounts of data; companies like YouTube or Facebook might be able to leverage large amounts of user-contributed data where others would be shut out of model training. Neither of these two outcomes is ideal. As litigation progresses, identifying mechanisms to prevent extreme outcomes will be critical. For example, it is important to understand that not all forms of foundation models are inherently infringing and that some of their potential risks can be effectively managed through technical means. With better demonstrations of co-evolving technical mitigation strategies, the law might find a middle ground that allows model training and deployment with sufficient effort to implement objectively strong mitigation strategies. Courts may consider the reasonable efforts of model builders and deployers to mitigate copyright risk, both in deciding fair use and in determining whether they can face indirect infringement liability. Trademark courts have taken a similar approach, for example in Tiffany v. eBay (2d Cir. 2010). As such, advancing research in this area (with methods such as improved similarity metrics) may help in preventing extreme outcomes in legal settings.42 We should not over-zealously filter. There must be a balance to filtering. Well intentioned but strict filtering mandates adopted by other countries have been criticized and criticized for their impacts on free speech (Schmon et al., 2022). Similarly, YouTube’s content ID system, a large-scale filtering approach, has been criticized for not following fair use standards and being overaggressive in its filtering (Bartholomew, 2014; Boroughf, 2015). Levendowski (2018) points out that restrictive views of fair use doctrine can exacerbate biases and that fair use can help create fairer systems. While mitigation strategies will help prevent undesirable outcomes, it is important to develop strategies that carefully align with fair use standards, as we have previously discussed. This means that factual content should not necessarily be filtered, neither should parodies, or short form regurgitation used for commentary. And evolutions of fair use doctrine or further policymaking should consider the distributive effects of preventing access to certain types of data for model creation. Policymakers could consider how and if DMCA (or similar) safe harbors should apply to foundation models. As we have seen, there are various ways, including filtering, to mitigate the risk of copyright infringement in the output of foundation models, but none will entirely eliminate the risk of liability. Even when trained on presumably permissively licensed datasets, for example, it is difficult (if not impossible) to determine the provenance of every piece of data and filter it out. Users might post content to seemingly permissively-licensed databases that they do not actually have the rights for. There may even be uncertainty about whether a piece of content is actually in the public domain or whether that status has been revoked.43 And even if foundation model practitioners implement strong mitigation strategies, the amorphous nature of fair use doctrine may make it difficult to know what kinds of content will be covered by fair use ex ante. With the uncertainty of DMCA protections (discussed in § 3), the law may need to adapt to this reality, and it could do so, for instance, by clarifying the role of safe harbors for models that implement sufficiently strong mitigation strategies. Policymakers could make clear that DMCA protections apply to this setting or they could identify other more suitable safe harbor mechanisms. This may provide more balance than general-purpose text and data mining exemptions seen in other countries, but again are not a panacea. Such safe harbors would have to be structured to consider the strength of the implemented mitigation strategies to ensure that they are not abused.

Last year, the Copyright Office issued guidance on the registration of works that contain AI-generated elements titled Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence. In the guidance, the Office explains that applicants have a duty to disclose the inclusion of AI-generated content in a work submitted for registration and to provide an explanation of the human author’s contributions to the work. Other notable requirements are that for AI-generated content, registrants must use the standard application, and in a situation where registration has already been granted (but AI-generated content was not disclosed), the applicant should correct the public record by submitting a supplementary registration. Unfortunately, despite the Copyright Office’s efforts to provide clarification, somewhat unsurprisingly, the Guidance has left many unanswered questions and resulted in some confusion on how the standards set forth in the guidance will be applied in practice. For example, there are several inconsistencies between the guidance and parts of the Copyright Office Compendium on registration guidelines that must be clarified. For instance, the Compendium says that unclaimable material should be disclaimed when it represents an “appreciable portion” of the whole work, whereas the guidance says that AI-generated content that is more than de minimis should be explicitly excluded from the application. These are two different standards that must be reconciled. In a webinar explaining the Guidance, the Office attempted to define what it meant by de minimis and described how it compared to “appreciable amount,” but unfortunately this raised additional questions. Use of the de minimis standard to separate material that should be disclaimed from material that should not be is not only confusing but also an inappropriate standard. The term “more than de minimis” is used in other contexts in the law where it does not involve a standard of separable copyrightability. For instance, in the context of joint authorship, a more than de minimis contribution may suffice to constitute joint authorship while not rising to the level of independent copyright protection, and, in the context of an infringement analysis, whether a portion of a work used without authorization is “more than de minimis” may have different implications. Because the de minimis standard varies in different contexts, the Office should not use the term as the standard for determining when it is necessary to disclaim material in a registration application. It would make sense for the Office’s guidance to simply confirm the existing standard articulated in section 621.2 of the Compendium, which explains that “[u]nclaimable material should be disclaimed only if it represents an appreciable portion of the work as a whole.” The guidance also includes the requirement that registrants must use the standard application when registering a work with AI-generated content, which, among other things, prohibits the use of group registrations and the benefits that flow from them, making it more challenging and economically infeasible for certain creators to register their works with the Office. Lastly, there is significant concern amongst many in the copyright community about retroactive application of the Copyright Office’s guidance. For creators and organizations with a vast portfolio of registrations, the threat of invalidation or cancellation is a major concern, especially when the guidance on where to draw the line regarding what to disclaim is unclear. Many copyright owners are worried that the new guidance will be misused by overly aggressive litigators to challenge the validity of every copyright registration if they believe AI was used even slightly and was not disclosed, which in turn might make litigation even more expensive than it currently is.