The AI Input Class: Constitutional Urgency and Fair Licensing in AI Copyright Class Actions

Introduction

Generative artificial intelligence models are fast becoming a normal feature of the digital landscape.¹ So quickly, in fact, that it has proven impossible for lawmakers and regulators to fully conceive of, must less respond to, the host of issues raised by this new but artificially natural-feeling technology.² Indeed, AI’s creators are struggling with the impact of their own work.³ The rapidity with which generative AI is spreading, combined with lawmakers’ deep-rooted commitment—embodied in the Communications Decency Act—to minimizing regulation of the Internet, means that litigation, rather than regulation, will play the lead role in establishing legal limits on 21st-century AI systems in the U.S.⁴

Moreover, the sheer magnitude of people and entities affected by AI foundation models, and the depersonalized, algorithmic mechanisms that propel those models, mean that many legal claims against AI companies will only provide effective relief if they are brought as class actions.⁵ Without the benefits of aggregation, there is no viable way for the creative working class—from journalists, novelists, playwrights, poets, nonfiction writers, screenwriters, composers, lyricists, musicians, videogame developers, and cartoonists, to illustrators, painters, and artists—who allege unlawful activity by AI corporations to pursue compensation and/or injunctive relief against further harm. Even with class action lawsuits, a favorable outcome is uncertain.

We anticipate waves of class litigation by parties impacted by generative AI as they seek to shelter their businesses, livelihoods, and identities from the rising influence of artificial intelligence in everyday life.⁶ This Article addresses the first wave—intellectual property disputes against the AI companies themselves—which is already here.

Copyright—the protection for “the creative powers of the mind”⁷—is enshrined in the U.S. Constitution.⁸ To replicate natural-seeming language, AI models are dependent on complex, rich narratives—narratives that can be found in novels, short stories, academic articles, poetry, songs, and other copyright-protected creative output. So AI companies copied that output without slowing down to seek permission or offer compensation. The companies decided that it was better to ask forgiveness than permission.⁹ At a time when the humanities are already under attack, technology companies are eating them for breakfast.¹⁰

Over the past year, authors, artists, coders, publishers, and organizations representing them have filed suits arguing that AI foundation model companies, from OpenAI to Anthropic—backed by Big Tech money—are violating their copyrights by “training” their large language models on tens of thousands of copyrighted creative works without even the veneer of an attempt to obtain consent or provide compensation.¹¹ The claims in these lawsuits can be conceptually divided into two categories. The first category, and the focus of this Article, are what we call input claims—that is, claims alleging that AI companies’ initial copying of text, art, or code is a form of direct copyright infringement. The second category are output claims alleging the synthetic contents generated by AI models in response to users’ prompts violate copyright holders’ exclusive right to prepare derivative works under copyright law.¹²

Despite the enormous influence of a handful of class suits on the development of copyright law, there remains very little scholarship on the intersection of class actions and copyright.¹³ In this Article, we seek to advance the study of copyright class actions in the field of artificial intelligence. To that end, we argue that class action plaintiffs’ input claims are correct on the merits. Beginning in the 1990s with the Ninth Circuit’s decisions in the so-called MAI Trio, courts have consistently and adeptly used existing copyright law to preserve authorial rights against incursions by new duplicating technologies. Under that precedent, AI “training” constitutes unlawful copying. This same set of precedents also undermines AI companies’ fair use defense.

Procedurally, this Article argues that the Rule 23 class action device is an effective and appropriate tool with which to impose legal guardrails for direct infringement on AI companies. Examining earlier copyright class actions, we show that the AI plaintiffs’ direct copyright infringement claims—which we call “input” claims because they arise out of material that was input into AI engines—fall squarely within the letter of Rule 23. AI class actions also embody Rule 23’s spirit. Class actions are appropriate when a defendant’s conduct impacts a group of people in functionally the same way. This maps on precisely to the facts of what AI companies did. Moreover, the class action device can achieve a remedy for large numbers of people who otherwise would be without the resources or knowledge to vindicate their rights. That is also true here: Only as a group do creators have the clout to negotiate with companies valued in the billions of dollars.¹⁴

Importantly, input claims are only one facet of creators’ challenges to AI. Plaintiffs in AI cases have also alleged various downstream, post-“training” infringement claims, which we refer to as “output” claims.¹⁵ For example, plaintiffs have argued that AI engine outputs may themselves be infringing.¹⁶ Our focus here is on input claims, but two observations are in order. First, while class certification may be appropriate for adjudicating some output claims, requirements for individualized proof and/or individualized assessment of defenses will render class certification more challenging for such claims. Second, if (as we think likely) the parties reach a settlement in the pending cases, any such agreement is likely to encompass both input and output claims, which arise from the same basic factual predicate. Any release of future output claims will thus hover in the air during resolution of the input claims. In assessing the fairness of a settlement, the key question will be whether the agreement preserves authorial voice and independence into the future in a sustainable way.

With regard to both input and output claims, the class action device serves a vital forward-focused role. Xiyin Tang has argued that copyright class actions spur the creation of licensing schemes and generate blueprints for legislative gap-filling.¹⁷ That is already happening in the AI context. Ironically, at the very same time that AI companies are defending their “fair learning” justification for stealing from the creative working class, they are already entering into licensing deals with news publishers as well as user platforms such as Reddit.¹⁸ The difference in the class action context is that a court will need to approve any settlement; an important element of that judicial review will be ensuring that working class artists, and not only best-selling authors, receive fair compensation for any licensing agreement that results.

We conclude by centering the risk to human creativity that is at the heart of the pending class actions. Artificial intelligence inevitably has the buzz, and the buzzwords, of cutting-edge post-modern life.¹⁹ But generative AI raises questions at least as old and as important as the Constitution. And the Constitution’s far-sighted protection of intellectual property rights should not be forgotten in the heady rush of innovation. In fundamental ways, AI copyright class actions are not solely about the right of this or that author, or the economic impact of copyright infringement on individuals; rather, these claims are about a threat to the concept of authorship as a human endeavor, and to the vision of America’s founders that copyright will protect and incentivize human creativity.

Since the Founding, copyright law has withstood the challenges of successive new technologies. But in a world where 90% of online content will be synthetically generated by 2026, human creativity is endangered.²⁰ Thus, we conclude by arguing that there is a constitutional—and not merely a statutory or economic—urgency to protecting the human role, and human agency, in the process of creation. Class actions are a feasible, viable way to respect this constitutional urgency. We recognize that licensing agreements are the most likely, and perhaps the only, way to resolve the impasse between creators and AI companies.

But licensing regimes pose their own risks to the livelihood and dignity of artists. Fair licensing must do more than give authors a smidgeon of cash in exchange for their creative souls; it must address the constitutional urgency, and the unknowable future, brought on by generative AI. To qualify as far, we argue that a licensing regime must bear three hallmarks: 1) it must allow creators to opt out; 2) it must not allow licensees’ unilateral expansion of the scope of the license; and 3) it must provide a mechanism for attribution to human authors.

Part I of this Article illustrates how AI transforms, and taints, human creativity.

Part II briefly describes the pending intellectual property class actions currently pending against AI companies and describes their central copyright claims.

Part III traces how, beginning in the 1990s, courts have successfully applied existing copyright law to successive waves of technological disruption. It shows that, under these precedents, AI companies are liable for feeding copyright-protected creative works into their engines.

Part IV analyzes courts’ application of Rule 23 in pre-AI copyright class actions. Current AI class claims reflect the same fundamental pressure points as did earlier suits.

Part V argues that, just as existing copyright law can meet this moment, so too can existing class action doctrine. Using two current AI cases as examples—one from the Second Circuit and one from the Ninth—we argue that AI plaintiffs’ input claims fall in the heartland of what Rule 23 is designed to accomplish. We also recognize that class certification is likely to prove more difficult (though not impossible) for “output” claims—claims for derivative infringement and other downstream claims that arise out of the original unlawful copying.

Part VI contests the AI companies’ fair use defense, which comes down to one philosophy: “trust us.”

Part VII concludes by urging courts and Congress to reject fair learning in favor of fair licensing. It proposes three terms that can ensure that a licensing regime stays faithful to the constitutional essence of copyright law.

I. AI and the Risk of Post-Human Creativity

In a rock band, four or five members each play a role or more: a lead guitar, rhythm guitar, bass guitar, drum, keyboard, and lead vocal.²¹ A symphony requires many musicians with talent in their respective instruments to collaborate under a conductor’s baton perfecting a beautiful movement.²² Jazz musicians are known for their partnership merging their different styles and elevating their art or seizing inspiration from listening to other jazz musicians.²³ The “intimate exchange of the creative pair, such as John Lennon and Paul McCartney” illuminates the value of human connectivity for creative process and achievements.²⁴ Recording an album needs more than just the band but the producer, the studio, the engineer, the mixing engineer, and the mastering engineer.²⁵ Connecting humans for collaboration is essential for creativity.

Michelangelo is known for the breathtaking painting of the Sistine Chapel ceiling. But he could not accomplish the feat alone. He “worked with a team of assistants” to achieve the results.²⁶ Picasso made Cubism together with Georges Braque and rivaled with Henri Matisse leading to adversarial collaboration.²⁷ Albert Einstein accumulated inventive knowledge from reading and analyzing other people’s inventions while he was working as a patent examiner at the Swiss Patent Office,²⁸ and testing his idea for his theory of relativity in conversation with the engineer Michele Besson, “the best sounding board in Europe.”²⁹ Creativity is demonstratively the product of human connectivity and collaboration.

With generative AI, creativity may become post-human. A new rock song, a newly improvised jazz piece, a new symphony, among other myriad new music products, can be synthetically generated.³⁰ A known musician’s style can be imitated through AI generation.³¹ A famed vocalist’s unique sound can be generated because GPT-4o “can respond to audio inputs in as little as 232 milliseconds.”³² Connectivity and collaboration are not invited. All the models need is prompts.

Do you want to write in the style of Tolstoy’s novels—perhaps War and Peace, or Anna Karenina? AI models are standing by to assist in generating deepfake Tolstoy-like new text through a series of prompts: “1. Directly ask the AI to write like the writer you have in mind[,] 2. Describe what the writer’s voice is like, and ask the AI to write like that (preferably with examples)[, and] 3. Fine-tune the AI on a specific writer[.]”³³ Do you wish to draft a sitcom dialogue mimicking Jerry and Kramer in a new Seinfeld episode?³⁴ In minutes, AI models trained on novels and scripts written by human authors present you with synthetic literary text of the original authors’ style.

Writing pitches for sequels to popular movies in an instant? AI models’ performance in the movie sequels task garnered praises like “surprising,” “good,” “good enough that you could see someone buying a ticket to the film,” and “brilliant.”³⁵ For instance, in response to Casablanca in the prompt, AI wrote: “One possible plot for a sequel to Casablanca could explore what happened to Rick Blaine and Ilsa Lund after they parted ways in Casablanca. The sequel could pick up several years after the events of the original film, with Rick and Ilsa living separate lives in different parts of the world. Perhaps circumstances conspire to bring them back together, and they must grapple with the complicated feelings they have for each other and the choices they made in the past.”³⁶ Having ingested all the scripts and movies, AI models produce new pitches deriving from the originals.

AI can generate a painting that helps you win a jury prize.³⁷ Prompts, again, are what you input at the computer.³⁸ Learning how to paint from teachers, imitating styles from accomplished artists, studying their techniques, and listening to critiques from others, none of these steps are necessary. AI models trained on existing images from all sources are ready at your prompt command to generate whatever you desire and whenever you wish.³⁹ All you need is to type in the prompts, and generative AI programs will do the rest in seconds.⁴⁰

Learning to speak or sing beautifully in a foreign language typically takes years of learning from teachers and coaches. In many cases, the students may travel abroad to submerge themselves in living in foreign countries for the opportunities to connect with the people and culture while improving their language abilities. In the AI age, just ignore human connectivity and creativity regarding learning foreign languages. For instance, AI can generate an English-speaking rock band vocalist (who is real) rocking away fluently and with nuanced flair in Japanese and Korean languages (synthetically, of course).⁴¹

Albums of new songs by different artists can be synthetically generated without ever stepping into a studio.⁴² Google’s Dream Track AI tool excites performers to generate new content in their replicas.⁴³ The days of collaboration in the studio are numbered. Obviously, for now, those famous few get paid by Google’s Deep Mind AI for having their voices cloned as part of new synthetic content creation without collaborating with others in the studio. For many others with ordinary voices, there is nothing to collect.⁴⁴

Generating a haiku in seconds? Ditto. AI models ingest haiku and are trained to generate haiku, poems known for having 17 syllables, and apply the world’s shortest and clearest rules to compose new poems and evaluate their aesthetic. AI can generate haiku quicker than humans, and humans fail to distinguish which haikus are synthetic and which are human-created.⁴⁵ John Grisham took years to research and develop characters and stories for his books. However, AI can generate sequel chapters mimicking John Grisham’s style in an instant.

AI poses an existential threat to the creative working class, and also to the very concept of human creativity.⁴⁶ If we do not act soon, we will no longer be able to distinguish—or perhaps even remember—what human creativity is.

II. Copying and Fair Use in Disruptive Technologies

The Constitution’s Copyright Clause was drafted during the age of lithography—more than 50 years before the rotary press spurred modern print media.⁴⁷ Yet its language, and its mission, have remained relevant.⁴⁸ Copyright law has outlasted cycles of disruptive technological advancements. When personal computers first began to enter businesses and households, critics lamented that “as people and businesses often find the transition to computers difficult, integrating computer software into the realm of copyright law has proven to be a daunting task.”⁴⁹ Ultimately, however, both courts and Congress rose to the challenge, and copyright law has remained a relevant, practical tool for protecting authorial rights from technology-enabled infringement. Beneath each technological innovation, the heart of the issue has remained: the protection of human creative output. And in each instance, the Constitution’s Copyright Clause has served as a compass, guiding courts to encourage technological innovation while simultaneously protecting human creativity.

This Part recounts how courts have thoughtfully applied copyright doctrine—both affirmative claims, as well as the defense of fair use—to waves of disruptive digital technology. The precedents described below provide a firm basis for holding AI companies liable for their unauthorized “training” of AI models on copyrighted creative works, and for denying a “fair learning” defense. Simultaneously, however, this part recounts the Supreme Court’s somewhat expansive use of the fair use doctrine to protect some code-copying in its 2021 decision in Google LLC v. Oracle America, Inc.⁵⁰

A. The MAI Trio: Code Copying

The Copyright Act explicitly contemplates the evolution of technological methods of copying creative output. It defines “copies” as “material objects…in which a word is fixed by any method now known or later developed.”⁵¹ In the early 1990s, the Ninth Circuit penned three opinions later referred to as the MAI Trio, defining the meaning of “copying” under the Copyright Act during the rise of software and personal computers. These decisions, which granted property rights—and copyright protections—to the output of the software age, apply directly in the age of artificial intelligence.⁵²

1. Code Is a “Copy”

The first of the Trio, MAI Systems Corp. v. Peak Computer, established that even temporary copies of software code are protected by copyright law.⁵³ In so doing, MAI Systems defined the concept of “copying” for the digital age.

The case concerned a dispute between MAI—a computer manufacturer and software developer—and Peak, a computer maintenance company.⁵⁴ When Peak provided maintenance and repair services to those of its customers using MAI computers, it necessarily ran MAI software.⁵⁵ MAI argued that this unlicensed use of its software constituted “copying” for purposes of the Copyright Act; it sought and obtained a permanent injunction against Peak.⁵⁶ The Court framed the question as whether “copying” occurs when a computer program is transferred from a permanent storage device to a computer’s random-access memory (RAM).⁵⁷ It answered that question in the affirmative.

To begin, the Court noted that it was “generally accepted that the loading of software into a computer constitutes the creation of a copy.” It applied this same logic to RAM copies.⁵⁸ A computer executing any program must load that program into RAM.⁵⁹ Peak argued that RAM copies were not “fixed” under copyright law. But the Ninth Circuit held that a RAM copy is “sufficiently permanent or stable to permit it to be perceived, reproduced, or otherwise communicated for a period of more than transitory duration.”⁶⁰ Therefore, it held that the “loading of copyrighted computer software from a storage medium (hard disk, floppy disk, or read-only memory) into the memory of a central processing unit “CPU”) causes a copy to be made.”⁶¹

The immediate impact of the case was that MAI could prevent Peak from using its software, thereby preventing Peak from offering its services as a competitor. The lasting impact was even broader. MAI was innovative, clear, and—given the lack of Supreme Court precedent in copyright—“far-reaching.”⁶² It was also problematic. Scholars have criticized MAI on textual and policy grounds.⁶³ As Danielle D’Onfro recently observed, the effect of MAI was to entrench restrictive intellectual property licenses into anything related to software—which is fast becoming everything.⁶⁴ Indeed, the Ninth Circuit now appears to offer a more nuanced take on software copies.⁶⁵

Nevertheless, what is good for the goose should be good for the gander. MAI’s broad conception of a fixed copy—which has tended to benefit corporate software developers—provides authors and other copyright owners with a strong shield against non-consensual copying by AI companies.

2. Limiting Fair Use for Code Copying

The second case in the MAI Trio, Triad Systems Corp. v. Southeastern Exp. Co., which reached the Ninth Circuit in 1995, narrowly construed fair use in the copying of software.⁶⁶ Triad manufactured computers for use by automotive parts stores.⁶⁷ Triad also developed software for its service technicians to use.⁶⁸ Southeastern competed against Triad for the business of servicing and maintaining Triad computers.⁶⁹ Upon discovering that Southeastern had been making RAM copies of Triad software when providing services to customers, Triad sued.⁷⁰ Following MAI, Triad moved for summary judgment on its copyright infringement claim; Southeastern countered with a fair use defense.⁷¹ The district court denied both motions, the case proceeded to trial, and the jury found that Southeastern had violated Triad’s copyrights.⁷² Following the verdict, the district court granted Triad’s motion for a preliminary injunction, pending the resolution of other issues in the case.⁷³ Southeastern appealed.

Applying MAI, the Ninth Circuit expressed “no doubt” that Southeastern had made infringing copies of Triad’s software.⁷⁴ The Court also noted that, for purposes of injunctive relief, there is a presumption of irreparable harm from copyright infringement, notwithstanding that a copyright plaintiff can theoretically be made whole through money damages.⁷⁵

Most significantly, the Ninth Circuit rejected Southeastern’s fair use defense.⁷⁶ Southeastern, the Court held, “has invented nothing of its own; its use of Triad’s software is…neither creative nor transformative, and does not provide the marketplace with new creative works.”⁷⁷ Southeastern was “simply commandeering its customers’ software and using it for the very purpose for which, and in precisely the manner in which, it was designed to be used.”⁷⁸ The Court also found that Southeastern’s conduct yielded no appreciable public benefit.⁷⁹ Accordingly, the Ninth Circuit rejected the fair use defense in RAM copy infringement cases.

In 1998—after MAI and Triad—Congress amended Section 117 of the Digital Millenium Copyright Act to prohibit software developers from using the RAM copy doctrine to eliminate competition from computer service companies.⁸⁰ But Congress did not eliminate the RAM copy doctrine itself; it merely provided an exception to it.⁸¹ Congress also did not overturn MAI’s holding that license agreements define the limits of licensees’ rights with regard to software.⁸²

3. The Age of Licenses

A decade later, the Ninth Circuit decided the final case in the MAI Trio, Wall Data Inc. v. LA County Sheriff’s Dept.⁸³ Wall Data developed RUMBA software and the Los Angeles County Sheriff’s Department purchased licenses for 3,663 computers.⁸⁴ But the Sheriff’s Department installed the software on 6,007 computers.⁸⁵ Wall Data brought a copyright infringement suit against the Sheriff’s Department, which raised “fair use” and “essential step” defenses.⁸⁶

The Court rejected fair use. It ruled that making verbatim copies of software, for the same purpose, was not transformative, nor was it an industry norm.⁸⁷ It also held that, were the practice of duplicating software in excess of one’s license to become widespread, it “could seriously impact” the market.⁸⁸

The Court also rejected an “essential step” defense under 17 U.S.C. § 117.⁸⁹ “Essential step” means the making of another copy of a software program in the utilization of the computer program in conjunction with a machine. The copy is used in no other manner. The Ninth Circuit held that the “essential step” defense only applies to an “owner” of a program, not to a licensee such as the LA Sherriff’s Department.⁹⁰ The Ninth Circuit cited MAI in support of its ruling that the Sheriff’s Department’s licensee is bound by the license terms, and as a licensee, the Sheriff’s Department did not have the statutory right to make extra copies of the software.⁹¹

The MAI Trio applied long-standing copyright law to burgeoning digital technologies. The Trio established that digital copying of copyrighted works in the computer software context is unlawful infringement, barring permission or a statutory exemption. It also empowered and protected software creators by allowing them to define the terms of use for their product through the (now ubiquitous) use of licenses. Notably, courts—not Congress—took the lead in analyzing the application of the Copyright Act to new technologies.⁹² Although Congress eventually weighed in, amending § 117, it did so only after the issues of copying and fair use of software had been explicated through an adversary process between those parties most directly impacted.

B. Cached Contents and Fair Use: Field v. Google

The explosion of networked computers in the early 2000s forced copyright law to update copyright doctrine once again. Companies such as Google, Yahoo, AltaVista, and Lycos developed technologies to capture the emerging search sector of the digital economy. To do so, they sought to copy, analyze, and cache every existing web page. Google’s automatic program for finding and cataloging web pages was called the “Googlebot.”⁹³ Search companies did not seek permission from the copyright owners of any of these web pages. Litigation soon followed, pitting copyright owners against the search companies.

In this battle, the search companies held their own.

When a user conducts a search at Google, the search engine displays its search results, comprising the title of a webpage, short snippets from the webpage, the URL for the page, and another link labeled “cached.”⁹⁴ Clicking on the “cached” link “directs an Internet user to the archival copy of a Web page stored in Google’s system cache, rather than to the original Web site for that page.”⁹⁵ A disclaimer explains that the page is a snapshot from the Google cache, not the original page.⁹⁶ The snapshot is therefore not current if the webpage has been updated by the owner.⁹⁷

The “cached” links served several purposes, according to Google, including: (1) providing archival copies that allowed users to access the pages in the event the website became inaccessible, (2) providing webpage comparisons, enabling users to determine how a particular webpage changed over time, and (3) identifying search query terms, thus allowing users to determine why a particular page was more responsive to their query.⁹⁸ Also, according to Google, the vastness of the internet prevented it from contacting website owners to ascertain whether they wanted their pages listed in search results or accessible through “cached” links.⁹⁹

In Field, the plaintiff—an attorney and amateur author—brought a copyright infringement action against Google for copying his website, on which he had posted 50 creative works, and storing these copies in a cache without authorization.¹⁰⁰ The plaintiff argued that Google directly infringed his copyrights when a user clicked on a “cached” link to the web pages containing the plaintiff’s copyrighted works and downloaded a copy of those pages from Google’s cache.¹⁰¹ That meant, according to the plaintiff, that Google itself had created and distributed copies of the plaintiff’s works.¹⁰²

Field did not prevail. The district court held that the user, not Google, creates and downloads a copy of the cached web page when the user requests a web page contained in the Google cache.¹⁰³ Without a user’s request, no copy is created or distributed; even with such a request, the “automated, non-volitional conduct by Google in response to a user’s request does not constitute direct infringement” under copyright law.¹⁰⁴ The district court also sustained Google’s implied license defense. At the time of the lawsuit, Field had an easy way to protect his digital material from being cached: he could have used a “no-archive” meta-tag, which would have informed Google not to display “cached” links to his page. Field did not do this.¹⁰⁵ Instead, the plaintiff “made a conscious decision to permit” Google to make available the cached links.¹⁰⁶

Most significantly, the district court found that Google’s copying and distributing of the plaintiff’s copyrighted works by allowing users access to them through cached links constituted fair use.¹⁰⁷ Under the first fair use factor—the purpose and character of Google’s use—the court observed that Google’s system cache serves a different purpose from that of the plaintiff’s original works. The plaintiff intended his copyrighted works to serve an artistic function.¹⁰⁸ But Google’s cache enabled users to access content when the original page is inaccessible, detect changes in a particular website, and allow users to understand why a page was or was not responsive to their query.¹⁰⁹ “Because Google serves different and socially important purposes…and does not merely supersede the objectives of the original creations,” the district court concluded that Google’s alleged copying and distribution of the plaintiff’s web pages was transformative.¹¹⁰

Assessing the three remaining fair use factors, the district court found that Google operated its cache system in good faith; that it provided a disclaimer and explanation; and that it followed industry protocols by not displaying “cached” links if the owner of a page opted out.¹¹¹

Field established a blueprint for other copyright infringement actions relating to cached contents.¹¹² For instance, the district court in Parker v. Yahoo! adopted much of Field’s reasoning and held that Yahoo!, a search engine with similar caching practices, had an implied license to display a cached copy of the plaintiff’s copyrighted works.¹¹³ AI companies will likely invoke Field to argue that their “training” is simply caching by another name. As we discuss further below, however, AI large language models are a far cry from Google’s caches.¹¹⁴

C. Licensing and Fair Use: American Geophysical Union v. Texaco, Inc.¹¹⁵

As courts expanded the power of licensors in copyright, they also incorporated licensing schemes into their concept of fair use. This trend is best embodied in American Geophysical Union v. Texaco, Inc., where the Second Circuit expressly considered the availability of a licensing mechanism in rejecting a fair use defense.¹¹⁶ In Texaco, academic researchers sued oil giant Texaco, alleging that the company routinely photocopied academic research publications for its scientists without permission from or compensation to the publishers.¹¹⁷ The court’s rejection of fair use in Texaco provides an important check on Field’s fair use approach.

In Texaco, the court questioned whether fair use could even apply to mechanical copying of an entire document.¹¹⁸ Even under the traditional four-factor fair use test, however, the court rejected the defense. It found that Texaco’s practice of making copies for individual scientists was “part of a systematic process of encouraging employee researchers to copy articles so as to multiply available copies while avoiding payment.”¹¹⁹ While the second factor—the nature of the copyrighted work—favored Texaco, the third and fourth factors tipped in favor of the plaintiff authors. The Court was particularly persuaded by the fact that Texaco copied each copyrighted work in its entirety. Texaco had attempted to reframe the question by noting that only a small fraction of any particular volume of the academic journal was copied—but the Court refused to take a macro view, given that “each article enjoys independent copyright protection.”¹²⁰

Most importantly, the Court found in favor of the plaintiffs on the fourth factor—purportedly the “most important” fair use element—because of the availability of the Copyright Clearance Center, a private licensing organization that could negotiate a royalty in exchange for additional copies.¹²¹ The Court noted that existence of “traditional, reasonable, or likely to be developed” licensing markets is an important factor in assessing the effect of an infringing activity on the market; if paying for a use is relatively easy, free use is, simply, “less fair.”¹²²

Some scholars have criticized Texaco for allowing the existence of a private licensing scheme to warp the traditional fair use analysis.¹²³ As discussed below, however, the doctrine of fair use must respond to current copyright culture. And within that culture, licensing organizations and agreements have become a routine and important way—perhaps the only feasible way—to mediate between copyright holders and digital innovators.

D. Digital Copying of Creative Works

Beginning early in the 2000s, technology enabled companies to duplicate creative works with increasing efficiency, opening the door to mass-scale digital duplication, resulting in individual as well as class suits seeking redress for copyright infringement by digital disruptors.

1. Digital Copying: New York Times Co. v. Tasini¹²⁴

In New York Times Co. v. Tasini, the Supreme Court confronted the first wave of large-scale digital replication of copyrighted works. The plaintiffs in Tasini had published their works in print periodicals. The periodicals’ publishers had, without consent, submitted the authors’ pieces to an electronic database, making them retrievable by users as individual works, detached from the volume in which they first appeared. The defendant publishers argued that this was a “revision” of a work authorized by copyright law. The Court rejected that argument.¹²⁵ Detaching these works from their context and dropping them into a massive database, the Court held, “effectively overrides the Authors’ exclusive right to control the individual reproduction and distribution” of their work.¹²⁶

The publishers offered several other defenses that recur in digital copying cases, none of which prevailed (and many of which were raised anew by subsequent digital disruptors). First, they claimed that the decontextualization of the copyrighted works could be solved technologically by enabling users to “manipulate the Databases” to retrieve all the works contained within the original volume.¹²⁷ The Court was unimpressed. Second, defendants argued that, at most, they could be contributory infringers; end users, rather than the Databases, would be committing the infringing acts. Again, the Court disagreed: “It is the copies themselves, without any manipulation by users,” that violate the Copyright Act.¹²⁸ Finally, the defendants argued that disallowing inclusion of these works in electronic databases would, as the Court phrased it, “punch gaping holes in the electronic record of history.”¹²⁹ The Court remained unmoved.¹³⁰ This, according to the Court, is what licensing agreements are made for.¹³¹

2. Fair Use: Google Book Project¹³²

Although Tasini was a Supreme Court decision, it is overshadowed in cultural memory by the Second Circuit’s adjudication of Google’s book project. Google began with an ambitious (and ultimately unrealized) plan to create a comprehensive database of books—a quasi-mythical universal library.¹³³ In 2004, announcing the “Google Books Library Project,” Google secured agreements with the New York Public Library, the Library of Congress, and several university libraries to digitally copy their collections, and it scanned more than twenty million volumes in its quest to create a massive digital book repository.¹³⁴ Left out of this vision was a process for seeking permission from, or paying compensation to, copyright holders included in the scanned book collection.

The Authors Guild, Inc., an organization of published authors, as well as several individual authors, challenged the project in a class action. Google asserted that its unauthorized activities—copying, storing, distributing, and displaying the copyrighted book—were protected by the doctrine of fair use.¹³⁵ The parties participated in extensive negotiations and entered into a proposed settlement.¹³⁶ That settlement would have granted Google broad (though not exclusive) rights over the books in its electronic database, including the right to sell subscriptions to an e-book database and the right to sell online access to individual books, in exchange for paying 63% of the revenues to copyright holders.¹³⁷ Objectors contested the settlement on the ground that it did not comply with Rule 23; they also alleged that it violated copyright, privacy, and antitrust law, among others.¹³⁸ The district court rejected the proposed settlement.¹³⁹

Google reduced the scope of its project so that users would only be able to see “snippets” of the copied texts in the public domain; what was originally a digital bookstore became a vast public index instead. The Second Circuit found that Google’s now-much-narrowed search and snippet functions satisfy fair use.¹⁴⁰ It held that the first factor—the purpose and character of the use—favored Google.¹⁴¹ Google’s use of copyrighted works “highly transformative” because its digitizing technology enables “a search for identification of books containing a term of interest to the searcher” and permits “a searcher to identify those that contain a word or term of interest”.¹⁴² With the snippet views, searchers can only “read snippets from the book searched” but gain “important value to the basic transformative search function, which tells only whether and how often the searched term appears in the book.”¹⁴³ Google’s tiny snippets show “the searcher just enough context surrounding the searched term to help her evaluate whether the book falls within the scope of her interest (without revealing so much as to threaten the author’s copyright interests).”¹⁴⁴ In other words, Google’s transformative use of the copyrighted works did not directly compete as substitutes for the plaintiffs’ books.¹⁴⁵

On the second factor, the nature of copyrighted works, the Second Circuit noted that “each of three” books scanned by Google is “factual” and concluded that nonfiction or fiction books are “not dispositive” in finding fair use.¹⁴⁶ The Circuit observed that “[n]othing in this case influences us one way or the other with respect to the second factor considered in isolation.”¹⁴⁷ Considering both the first factor and second factor of fair use doctrine together, the Circuit stated, “the second factor favors fair use not because Plaintiffs’ works are factual, but because the secondary use transformatively provides valuable information about the original, rather than replicating protected expression in a manner that provides a meaningful substitute for the original.”¹⁴⁸ The Circuit, however, did not provide extensive explanation. This seems to suggest that authors of published nonfiction books may face a conclusory finding of fair use under the second factor, or that the second factor is no longer relevant in fair use analysis.

For the third factor, the amount and substantiality of the portion used, the Second Circuit found that “[w]hile Google makes an unauthorized digital copy of the entire book, it does not reveal that digital copy to the public. The copy is made to enable the search functions to reveal limited, important information about the books.”¹⁴⁹ Accordingly, with respect to the search function, Google satisfied the third factor. As to the snippet view function, Google constructed the snippet feature in a “manner that substantially protects against its serving as an effectively competing substitute for Plaintiffs’ books”.¹⁵⁰ In fact, Google included many limitations on the snippet function:

These include the small size of the snippets (normally one eighth of a page), the blacklisting of one snippet per page and of one page in every ten, the fact that no more than three snippets are shown—and no more than one per page—for each term searched, and the fact that the same snippets are shown for a searched term no matter how many times, or from how many different computers, the term is searched. In addition, Google does not provide snippet view for types of books, such as dictionaries and cookbooks, for which viewing a small segment is likely to satisfy the searcher’s need. The result of these restrictions is,…, that a searcher cannot succeed, even after long extended effort to multiply what can be revealed, in revealing through a snippet search what could usefully serve as a competing substitute for the original.¹⁵¹

….

The other restrictions built into the program work together to ensure that, even after protracted effort over a substantial period of time, only small and randomly scattered portions of a book will be accessible.¹⁵²

With such limitations and restrictions imposed by Google in creating its search and snippet functions, Google satisfied the third factor for both the search and snippet functions.

Under the fourth factor, the effect of the use upon potential market or value, the Second Circuit ruled for Google because Google did not sell scanned books in competition with the plaintiffs. Specifically, the snippet views in the aggregate amount to “no more than 16% of a book.”¹⁵³ The Circuit also found that a loss of sale of a book occurs “in relation to interests that are not protected by the copyright” because “a searcher’s need for access to a copyrighted book will at times be because the snippet conveys a historical fact that the searcher needs to ascertain.”¹⁵⁴ In summary, Google’s making of a complete digital copy of the copyrighted works for the narrow and unchanging purpose of providing the public with its search and snippet view functions did not constitute copyright infringement.

Google, the defendant in both Field and Google Book Project, prevailed in its fair use defense in the second suit because both cached content and the snippet index involved limited, unchanging, and transformative use of existing copyrights. In neither case did the search company’s output interfere with copyright holders’ potential market of selling their books or negatively impact the value of copyrighted works. In the context of AI input class actions, by contrast, there are no firm guardrails on the use of the copied copyrighted works. There is also a strong profit motive to resist the formation of such guardrails. AI companies rely on existing literary text, images, and sounds to create AI models. Without human creative works to train AI models, there can be no AI models.¹⁵⁵ That creativity is the essence of what copyright law protects.

3. Fair Use of Code: Google LLC v. Oracle America, Inc.

In setting up its open-source Android platform, Google engineers wrote millions of lines of new code. But they also copied 11,500 lines of code from Java SE, a program ultimately owned by Oracle.¹⁵⁶ By using this code—which included tasks that were essential to using Java coding language at all—Google enabled “millions of programmers, familiar with Java, to be able easily to work with its new Android platform.”¹⁵⁷ Oracle sued Google, arguing among other things that Google’s use of the code violated copyright. The question before the Supreme Court was whether this copied code is subject to copyright protections (the trial court had found otherwise¹⁵⁸), and—if yes—whether Google’s actions were shielded as fair use.

Reasonable minds could and did disagree about these issues. One jury deadlocked on the question of fair use;¹⁵⁹ following an appeal, a subsequent jury found Google’s use of the code did constitute fair use.¹⁶⁰ On appeal again, the Federal Circuit reversed, finding “nothing fair about taking a copyrighted work verbatim and using it for the same purpose and function as the original in a competing platform.”¹⁶¹

The Supreme Court, in an archetypal opinion by Justice Stephen Breyer, reversed the reversal.¹⁶² It assumed, without deciding, that the code at issue was protected by copyright.¹⁶³ But it concluded that Google’s use of the 11,500 lines of code was protected by fair use.¹⁶⁴ The purpose of the judge-made, equitable doctrine of fair use, the Court noted, is to “permit courts to avoid rigid application of the copyright statute when, on occasion, it would stifle the very creativity which that law is designed to foster.”¹⁶⁵ In the Court’s view, Google’s use of the Java code merited that protection.

To a large extent, the Court’s fair use holding reflects underlying skepticism toward the concept of granting copyright protection to computer code. Although it assumed for purposes of argument that code is copyrightable, it strongly suggested that—to the extent it is protected at all—code should enjoy much thinner protection than fiction, and even than less-protected non-fiction.¹⁶⁶ In the Court’s view, the code at issue was a type of functional organizer—the equivalent of cabinets, drawers, and files.¹⁶⁷ On this basis, when analyzing the first fair use factor, the nature of the copyrighted work, the Court ruled that the code was, “if copyrightable at all, further than are most computer programs…from the core of copyright.”¹⁶⁸

The Court also indicated strong support for Google’s argument that its use of the Java code lines was transformative, thus tipping the second fair use factor in Google’s favor. In the Court’s view, Google’s use of the code furthered the development of computer systems in the new and important realm of smartphone platforms.¹⁶⁹ The Court found that the third factor—the amount and substantiality of the portion used—also favored Google: the Court found that the copied code was only 0.4 percent of the total set of Java’s code.¹⁷⁰ Finally, the Court found that Google’s use of the code lines posed little threat to Oracle’s market, because Oracle was “poorly positioned to succeed in the mobile market.”¹⁷¹ In other words, Google was not developing a market substitute for Oracle’s products.¹⁷² The Court also saw Oracle’s attempt to shield this code as a potentially monopolistic move that would thwart innovation.¹⁷³ It ruled in favor of Google on the ground of fair use.

***

The cases involving RAM copies, cached contents, digital libraries of scanned books, and lines of computer code, illustrate the ways that copyright law has adapted in response to the computer revolution in the last five decades. These precedents show that AI companies, in “feeding” copyrighted works to their systems, were engaging in what copyright law defines as unlawful infringement. Yet that is not the end of the matter. In the era of networked computers, courts—squeamish about squelching innovation—appear to have become more open to findings of fair use. Where Big Tech enjoyed fair use protection, no asking for forgiveness was necessary. But the scope of fair use doctrine remains ambiguous, and fair use decisions such as Field and Oracle apply imperfectly, if at all, to the wholesale copying of thousands of copyrighted works of fiction and non-fiction.

This uncertainty lays the foundation for the current wave of class actions by creators against AI companies. The technology is new, but the questions are familiar: what constitutes infringement? What infringement is protected by fair use? With one notable recent exception, courts in the many pending cases have yet to answer.¹⁷⁴

III. Generative AI and the First Wave of Litigation

The unexpected arrival of generative AI in 2022 as the world resumed normalcy after the COVID-19 pandemic sent a shockwave across the globe. Generative AI dominated headlines in late 2022 and early 2023. This new technology was lauded—and also criticized for trampling on copyright and privacy laws. This Section briefly describes the three primary types of cases in the current, first wave of litigation arising from generative AI.¹⁷⁵

A. Copyright Infringement Actions

Generative AI systems rely on data to train models for the creation and generation of new works that mirror the training data. The quality of that data matters.¹⁷⁶ The phrase garbage in, garbage out, first uttered by IBM programmer and instructor George Fueschel,¹⁷⁷ applies squarely to generative AI.¹⁷⁸ Indeed, AI developers actively seek to avoid poor data in training their models. This has led AI companies to use copyrighted content to train models, without taking the time to seek permission from, or offer compensation to, copyright holders. This type of infringement falls in the “input” claims category.

In general, copyright refers to a bundle of rights, namely, the right to make copies, the right to prepare derivative works, the right to distribute the works, the right to publicly perform the works, and the right to publicly display the works.¹⁷⁹ Before the arrival of generative AI, most copyright infringement cases centered on unauthorized copying and distribution. In American Geophysical Union v. Texaco, Inc., for example, the defendant’s employees used a photocopy machine to make unauthorized copies of journal articles.¹⁸⁰

The age of networked computers spurred the creation of new technologies for ease of copying and distribution, as seen in A&M Records, Inc. v. Napster where the defendant was accused of providing a central server that facilitated music file uploading and sharing among users without permission from copyright holders.¹⁸¹ As a result, fewer copyright disputes rested on violation of the right to prepare derivative work, and if they did, the focus was on situations where a subsequent work was allegedly derived from the original.¹⁸² Moreover, the development of well-recognized IP structures, such as the licensing model and movie right options, helped guard against infringing derivative works. For instance, J.K. Rowling sold both movie and merchandise rights based on her Harry Potter series.¹⁸³

In the AI era, copyright disputes have taken a darker turn. Pending copyright class actions zero in on the violation of both the right to make copies and the right to prepare derivative works. In Getty Images v. Stability AI, for example, Getty Images alleges that Stability AI copied more than 12 million photographs without permission for compensation, with the purpose of “build[ing] a competing business.”¹⁸⁴ In the course of so doing, Getty Images alleges, Stability AI also either removed or altered Getty Images’ copyright management information and infringed on its trademarks.¹⁸⁵ Now stuffed to the gills with rich and complex images created by human artists, Stability AI dazzles the public with its generative AI models, with which users can generate stunning AI images in seconds. In sum, unlike the copyright infringement actions of prior decades, AI copyright suits seek to protect creators’ rights to authorial creative integrity against a pending torrent of copycat creativity.

Two unsettled doctrinal questions dominate the pending copyright suits. The first is whether the AI companies’ actions constitute “copying” for purposes of copyright law. We argue above that, in light of a series of cases applying copyright law to emerging technologies, it does.¹⁸⁶

The second question is whether AI developers’ use of copyrighted works is protected by fair use. Courts, including the Supreme Court, have struggled to apply the traditional four fair use factors in this new context.¹⁸⁷ In Oracle, the Supreme Court held that fair use protected Google’s use of 11,500 lines of Oracle’s Java code.¹⁸⁸ But neither Oracle nor other cases address fair use when creative works are being copied and made available to the public wholesale, nor when the purpose of this copying is to enable the instantaneous creation of limitless derivative works that will directly compete against the works that were copied.

The pending AI copyright suits, which are seeking substantial money damages for past infringement,¹⁸⁹ leverage the utilitarian and incentivization theories underlying copyright infringement claims.¹⁹⁰ These theories, however, understate the core value of copyright to society.¹⁹¹ The heart of the copyright violations by AI companies is the complete disregard for creators’ originality, manifested as unique voices, styles, signatures, manners, and genres.¹⁹² Rich, diverse works of authorship enhance the progress of humankind. Encouraging that progress is an explicit purpose of the Intellectual Property Clause of the Constitution.

B. Claims Under the Digital Millenium Copyright Act

Open-source software developers are also in the first wave of generative AI class actions. In Doe v. GitHub, Inc.,¹⁹³ anonymous members of a proposed class of developers allege that AI companies have taken billions of lines of open-source code, without attribution or permission, to train their generative models, thus violating traditional copyright law as well as the Digital Millenium Copyright Act of 1998 (DMCA).¹⁹⁴ This type of claims is also in the “input” claims category.

Computer software developers spend significant time and creativity drafting software code.¹⁹⁵ Coders have adopted open-source licensing regimes to facilitate collaborative efficiency and creativity within the copyright law framework. For more than three decades, coders worldwide have followed the requirements of open-source license agreements, including GPL, MIT’s Open-Source License.¹⁹⁶ In these agreements, the copyright management information (CMI) such as the copyrighted work’s title, copyright registration number, the copyright owner’s name, the creator’s name, and terms and conditions for use of the work, are included as part of the licensed materials.¹⁹⁷ Under these agreements, billions of lines of code have become publicly accessible. Developers post code in so-called Git depositaries—hence the name GitHub. The open-source software movement fueled the rapid development of the World Wide Web, among countless other projects.

Big Tech companies, notably Microsoft, in teaming up with startup OpenAI, are releasing generative AI systems that are trained on billions of lines of code taken from these publicly accessible code depositaries. But coders allege that the companies have exploited this code without complying with the open-source license terms.¹⁹⁸ In fact, according to the class action complaint brought by coders, Big Tech companies removed or altered the copyright management information (CMI) from the licensed codes, cleansing them of copyright identity and protection.¹⁹⁹ The cleansed codes are used as training data for generative AI models. The output of the training data is new codes that are either copies of the original codes or derivatives thereof. Though AI companies have not made their codes public, the output exposes the scope and the damages to coders and authors alike. The output mirroring the training data demonstrates that when it comes to your children, you can immediately identify them because you were the one who gave birth to them. The coders seek to bring their grievances as a class.

C. Data Privacy Violations

The third type of AI class action challenges the unlawful taking of private information, including personally identifiable information, from hundreds of millions of internet users, including children of different age groups, across platforms and sites, without their informed consent or knowledge. The allegations in J.L. v. Google LLC and A.T. v. OpenAI exhibit this “move fast and break things” attitude.²⁰⁰ Indeed, Google, the defendant in one of the two data privacy class actions, has repeatedly engaged in the strategy of “Do First, Ask Forgiveness Later” that entangled the company in legal troubles with both regulators and rights holders.²⁰¹ The data privacy claims, however, is not in the copyright’s “input” claims category.

In J.L., eight plaintiffs filed a proposed class action against Google regarding its generative AI system Bard.²⁰² The plaintiffs alleged that Bard scraped personal data and property for its training data. Google moved to dismiss the complaint based on a wide range of procedural and substantive defects.²⁰³ Google asserted that the use of the data is necessary to train Bard and the lawsuit would “take a sledgehammer not just to Google’s services but to the very idea of generative AI.”²⁰⁴ Google’s response reflects the current arms race among Big Tech companies in their race to develop and train what will become the dominant generative AI models.

In another class action, A.T. v. OpenAI, the complaint alleges that OpenAI, together with Microsoft, has continually released new versions of ChatGPT, Dall-E, and Vall-E since the first version was released, jolting and entrancing the public with “human-like” generative AI capability.²⁰⁵ Subsequently, with ChatGPT 4.0, for instance, OpenAI charges a subscription fee of $20 per month per user. OpenAI is amassing billions of dollars from both its alliance with Microsoft and charging high user fees. According to the complaint filed in the class action, to develop ChatGPT, Dall-E, and Vall-E or generative AI products, OpenAI turns to the internet to steal private information, including personal data from millions of unsuspecting consumers worldwide. This stolen data, the complaint alleges, includes private information and private conversations, medical data, information about children, and other type of data OpenAI can scrape off the internet to feed its large language models and deep language algorithms.

In addition, the complaint asserts, OpenAI takes private and personal information from its 100 million registered users, without disclosing that to them. This information includes all input details, account information users enter for registration, names, contact details, login credentials, emails, payment information for paid users, transaction records, identifying data harvested from users’ devices and browsers, social media information, chat log data, usage data, analytics, cookies, keystrokes, typed searches, and other online activity data.²⁰⁶ Most significantly, OpenAI scrapes users’ digital footprints accumulated in real-time and as far back as fifteen years ago.

The allegations paint a stark picture that the massive personal data gives OpenAI sufficient information for its generative AI products to be created and trained to have the ability to “create our digital clones, including the ability to replicate our voice and likeness and predict and manipulate our next move,” “misappropriate our skill sets and encourage our own professional obsolescence.”²⁰⁷

***

AI models pose a threat to privacy and human creativity. In October 2023, President Biden signed an executive order directing agencies to develop “New Standards for AI Safety and Security.”²⁰⁸ But that directive is unlikely to provide compensation to those creators whose copyrights have already been disrespected, or to those people whose privacy has been violated. Especially given the pace and scale of AI companies’ efforts, litigation is the best option for copyright holders—including everyday people—to shelter their creative works and personal data from this new force. The first wave of AI class actions seeks to do just that.

IV. Class Certification in Pre-AI Copyright Suits

Class actions, governed in federal courts by Rule 23, are a well-established (if sometimes controversial) vehicle for groups of individuals alleging similar harm to seek redress for that harm in a single suit led by representative plaintiffs.²⁰⁹ The quintessential class action is for so-called “negative-value” claims, where—in the absence of a class—individuals would not have the incentive or resources to vindicate their legal rights in court.²¹⁰ AI class actions brought by creators whose works have been used to “train” large language models fit this framework perfectly. Yet class actions are high-stakes, uncertain endeavors. The very things that make class actions so valuable—their flexibility, their potentially broad preclusive effect, and their ability to generate productive settlement discussions—also make them risky for everyone involved.

Although group litigation has ancient roots, the modern class action dates to 1966, when rulemakers promulgated what is now Rule 23.²¹¹ From the outset, many proponents of Rule 23 conceived of it as a quasi-regulatory device, resolving concrete disputes in a way that could provide closure to large numbers of people, many of whom would lack the resources to seek individual vindication of their rights.²¹² Today, civil rights, consumer, and tort litigants argue that class actions are “an important substitute for, or addition to, public administration.”²¹³ In contrast to this quasi-regulatory conception, the defense bar has sought to depict Rule 23 as a mere joinder device, albeit one that, in their view, is being exploited by rapacious plaintiffs’ lawyers. David Marcus refers to this formalist conception as an “adjectival” view of Rule 23—one in which the procedure is subordinate to, and should not distort, the substantive law.²¹⁴

Inevitably, these theoretical tensions have played out in copyright class actions.²¹⁵ Even prior to the Roberts Court’s procedural revival²¹⁶—which has created heightened barriers to class certification—courts evinced skepticism toward copyright class actions. One court described a proposed copyright class unflatteringly as a “Frankenstein monster.”²¹⁷ Perhaps as a result, until recently, there were relatively few such suits, with the Google Book Project and Spotify cases likely the only ones to receive mainstream media coverage.²¹⁸

Yet over time many groups of creators did seek to bring class actions to vindicate their copyright claims. As described below, plaintiffs in earlier suits invoked different categories of Rule 23, sought different forms of relief, and met with varying degrees of success. Among those that failed, some founded on substantive grounds (such as a fair use defense) unrelated to class certification, while others were deemed not to have met the requirements of Rule 23. Many of these decisions were unpublished, which limited their precedential value.²¹⁹

Nevertheless, these pre-AI copyright class actions provide insight into the applicability of Rule 23 in today’s AI suits. In general, when courts concluded that the heart of a claim was a defendant’s allegedly infringing conduct, they tended to find Rule 23 satisfied. Conversely, courts focusing on the individualized nature of copyrights, or on individualized defenses such as notice or fair use, tended to find class resolution inappropriate.²²⁰ Predictably, then, plaintiffs’ theory of liability was rooted in the defendants’ conduct: the class argued that the defendant acted according to a uniform policy or practice, often with a dollop (or more) of bad faith. Conversely, copyright class defendants emphasized the scattered nature of the plaintiff copyright holders; they told a story of uniqueness, messy layers of contracts, and the need for individualized determinations of copyright ownership and fair use.²²¹ Details of Rule 23 aside, certification appears to depend on which of these conflicting narratives—and class action theories—prevails.

This Section describes how courts have applied the certification requirements of Rule 23(a) and (b) in pre-AI copyright class actions. It also explains the specific way that courts have applied Rule 23(e)—which governs approval of class action settlements—in the context of copyright. To approve a class action settlement, a court must find that the proposed class satisfies all of these provisions.

A. Copyright and Class Action Rule 23(a)

To obtain class certification—in order to litigate, or to obtain judicial approval of a settlement—a party bears the burden to satisfy the Rule 23 criteria. First, the proposed class representative must demonstrate that the claims meet the four requirements of Rule 23(a), which are referred to in shorthand as numerosity, commonality, typicality, and adequacy of representation.²²² Taken collectively, these four factors are intended to ensure that a proposed class is sufficiently large and yet sufficiently cohesive to litigate as a collective, through the proxies of class representatives.

Prior to the Roberts Court’s heightened scrutiny of class actions, 23(a) was not considered particularly onerous. This shoot-from-the-hip view of Rule 23(a) applied in early copyright class actions as it did more generally. For example, in Bernstein v. Universal Pictures, Inc., the district court provisionally certified a class of lyricists and composers in a suit alleging that movie and television producers used adhesive contracts to constrain plaintiffs from entering the industry.²²³ The Second Circuit ultimately rejected the suit on its merits, but the class certification appears to have been decided in a cursory preliminary order. Another court certified a class of composers alleging infringement by cable company Showtime/The Movie Channel in five breezy paragraphs (at least, breezy by civil procedure standards).²²⁴

Later decisions, however, gave more attention to 23(a) factors:

23(a)(1): To meet the numerosity requirement, a party seeking class certification must show that the class “is so numerous that joinder of all members is impracticable.”²²⁵ There is no numerical cutoff for a class; this finding is within the discretion of the court. Class sizes smaller than 50 people are unusual, but not unheard of.²²⁶ Copyright class actions have not typically struggled to meet the numerosity threshold, although in one early case, a court denied certification in a one-page order because the plaintiffs had failed to demonstrate the existence of a sufficient number of freelance photographers affected by the defendant’s alleged infringement.²²⁷

23(a)(2): This factor, which requires a showing of “questions of law or fact common to the class,”²²⁸ was long considered a minimal burden—a view that is reflected in early copyright class actions.²²⁹ In Ryan v. CARL Corporation, for example, the plaintiff class argued that the defendant’s document retrieval service, which functioned “like a private interlibrary loan service,” had provided over 700,000 copies of plaintiffs’ copyrighted publications to users of its service, making payments for only 461 of them.²³⁰ In affirming commonality, the court noted that “plaintiffs are essentially challenging a standard business procedure” with a “common core of salient facts.”²³¹ The same logic applied in American Geophysical Union v. Texaco, Inc., although that case settled without a judicial analysis of Rule 23, after courts in the Second Circuit ruled against Texaco’s fiar use defense,²³² Texaco’s activity of making copies of academic articles for its employees was uniform in its purpose and effect.

Since 2011, however, commonality has been a higher hurdle. That year, the Supreme Court decided Wal-Mart v. Dukes, a Title VII class action in which over a million employees claimed gender discrimination, the Court emphasized that not any common questions would do; claims must depend on a “common contention…that is capable of classwide resolution—which means that determination of its truth or falsity will resolve an issue that is central to the validity of each one of the claims in one stroke.”²³³

Unsurprisingly, Wal-Mart’s more stringent approach affected proposed copyright classes. For example, in Wu v. Pearson Education, Inc., the district court initially certified a class of copyright owners who alleged that Pearson had printed educational materials containing the owners’ work product in numbers in excess of the amount for which Pearson had paid.²³⁴ Subsequently the case was transferred to a different district court judge, who decertified the class, in part based on a finding that “layers of individually negotiated contractual arrangements” defeated commonality.²³⁵ Similarly, the district court in Football Association Premier League, Ltd. v. YouTube found no commonality in a suit by a “worldwide class” of copyright owners against YouTube, given the need for individual determinations of the validity of a copyright, notice to YouTube of infringement, fair use, damages, and other specific defenses.²³⁶ Commonality continues to pose a significant challenge to class plaintiffs, including those in pending copyright class actions.

23(a)(3): This provision asks whether “the claims or defenses of the representative parties are typical of the claims or defenses of the class.”²³⁷ Although there is some overlap between this provision and both (a)(2) and (a)(4),²³⁸ the focus of the typicality factor is on the congruency between the claim of the named plaintiff(s) and those absent class members: it requires that the claims of the class representatives are fundamentally aligned with those of the absentees.²³⁹

Analysis of typicality in copyright class actions has often been cursory. In one early suit by a proposed plaintiff class of music publishers against a proposed defendant class of producers and distributors, the court held that the single named plaintiff lacked Article III standing to make claims against defendants other than the one that had allegedly reproduced his copyrighted music.²⁴⁰ For that reason, the court held that the plaintiff also failed to show typicality.²⁴¹

But typicality may also be a sticking point in plaintiff-side class actions. For example, the court in Football Association minced no words in rejecting certification based on typicality as well as commonality.²⁴² The class claims in that case were dauntingly broad. The proposed class included copyright holders worldwide who fell into either of two categories: (1) YouTube had previously blocked infringement of their copyrights on its site, but there were subsequent infringing uploads; and/or (2) YouTube had allowed copyrighted musical compositions to be used without authorization.²⁴³ The court noted that, to establish liability, each class member of this proposed worldwide class would be required to show that YouTube had legal knowledge of the unlawful conduct.²⁴⁴ As to each allegedly unlawful download, moreover, YouTube would have a potential fair use defense.²⁴⁵ “The unique nature of each work and of its infringement,” the court stated, “cannot be obliterated by its inclusion in a sea of other claims.”²⁴⁶ Other courts have also evinced skepticism of typicality in broadly framed class claims.²⁴⁷

23(a)(4): The final, and most important, element of 23(a) is adequacy of representation. This provision encapsulates the need to ensure due process before allowing representative parties to litigate the rights of those who are absent.²⁴⁸

Adequacy contemplates two separate inquiries. The first—and the one most pressing in copyright class actions—is whether the named class representatives will fairly and adequately represent the class. To protect this fundamental requirement, courts must ensure there are no significant conflicts of interest between the named representatives and absent class members. In Amchem Products, Inc. v. Windsor, the Supreme Court rejected a massive asbestos-related class settlement on the ground that the class representatives, who had current asbestos-related injuries, were structurally inadequate to represent class members whose injuries might manifest in a more-distant future.²⁴⁹ Although the Court pointed to specific examples that gave rise to its concerns—such as the settlement’s failure to adjust for future inflation—its holding was premised not only on the terms of the settlement, but the structural conflict of interests between representatives and absent class members.

Courts have found adequacy to be an important consideration in copyright class settlements involving future licensing agreements. For example, the district court rejected a proposed 2011 settlement in Authors Guild v. Google (Google Book Project) in part because the “interest and values” of certain absent class members differed from those of the class representatives.²⁵⁰ Under such circumstances, the court found it unacceptable that absent class members might be “deemed—by their silence—to have granted to Google a license to future use of the copyrighted works.”²⁵¹ Similarly, the Second Circuit reversed approval of a proposed settlement in In re Literary Works in Electronic Database Copyright Litigation because the Court agreed with objectors that the named plaintiffs had a structural incentive to favor class members with registered copyrights over class members without any registered copyrights—a bias that was reflected in the structure of the proposed settlement.²⁵²

In both Google Book Project and In re Literary Works, the plaintiffs were subsequently able to satisfy the adequacy requirement, at least at the district court level. In Google Book Project, Judge Chin in 2012 approved Authors Guild’s motion for class certification notwithstanding Google’s claim that the class representatives were inadequate because a large percentage of absent class members in fact supported Google’s book project and/or believed they benefited from it.²⁵³ In dicta, the Second Circuit noted its view that Google’s adequacy argument “may carry some force,” but it vacated certification without analysis to prioritize resolution of the fair use question.²⁵⁴

The class in In re Literary Works also ultimately satisfied the adequacy requirement. The Second Circuit mandated the creation of sub-classes to ensure adequate representation of authors with only non-registered copyrights.²⁵⁵ Subsequently the district court approved a revised settlement agreement that included such a sub-class.²⁵⁶

The second facet of adequacy addresses whether class counsel has the requisite experience and resources to prosecute a complex action on behalf of a class. As of 2003, Rule 23 mandates that courts adjudicating class actions appoint class counsel.²⁵⁷ Since that time, it is unusual for a court to deny class certification on the ground that counsel is inadequate. But in one case—coincidentally decided in 2003—the court did deny certification in a copyright case based in part on poor representation.²⁵⁸

B. Copyright and Class Action Rule 23(b)

In addition to Rule 23(a), a proposed class must also comport with one provision of Rule 23(b), which delineates three major forms of the class action. Given the small number of proposed copyright classes that survived Rule 23(a), there is minimal caselaw analyzing Rule 23(b) in the context of copyright claims. That said, copyright plaintiffs have sought certification under all three Rule 23(b) class subtypes—a pattern that is replicated in pending AI class actions. As described below, the various Rule 23(b) categories differ in ways that are significant for copyright cases. For example, some Rule 23(b) categories prohibit class members from opting out; one category mandates such an opt-out right. Some categories allow money damages; others bar them. As described below, the AI copyright plaintiffs have pleaded certification under all these potential categories, leaving their options open.

Rule 23(b)(1): Although it comes first chronologically, Rule 23(b)(1) is a lesser known, somewhat obscure provision of the class action rule. It contains two subparts.²⁵⁹ Referred to colloquially as the “prejudice class” provision, (b)(1) authorizes class resolution in situations where individual suits reaching inconsistent results might prejudice either the party opposing the class (typically the defendant), or members of the class (typically the plaintiffs). Rule 23(b)(1) does not permit putative class members to opt out of a certified class; this means that a class certified under (b)(1) will bind—for good or for ill—all who are encompassed within the class definition. In addition, there is no requirement that class members receive individualized notice in (b)(1) classes, although Rule 23 does provide that a court “may direct appropriate notice to the class.”²⁶⁰

(b)(1)(A): This sub-provision of (b)(1) allows certification where individual suits might “establish incompatible standards of conduct.”²⁶¹ On its face, this language seems to contemplate class certification under a very broad set of circumstances. That is not true in practice. Most importantly, this provision does not make certification appropriate simply due to the risk of inconsistent jury verdicts. Construed that way, (b)(1)(A) would make most cases certifiable.²⁶² In practice, this Rule 23(b) category typically applies in contexts where a ruling against the defendant as to one party would force the court to take identical steps with respect to others similarly situated, such as in a case brought by one of multiple riparian landowners.²⁶³ In modern day, this subtype is most commonly used in ERISA cases, where a plaintiff’s action is brought on behalf of a retirement plan.²⁶⁴

The court in Showtime/The Movie Channel granted certification in a copyright suit under (b)(1)(A), based on its finding that individual suits by copyright owners against the defendant cable company might create contrary, seemingly arbitrary, results.²⁶⁵ Reflecting an earlier, more naïve approach to class action analysis, however, the court’s analysis was barely a paragraph long.²⁶⁶

(b)(1)(B): This sub-provision allows certification where individual judgments “as a practical matter, would be dispositive of the interests” of those not before the court.²⁶⁷ Just as with (b)(1)(A), stare decisis—or the possibility that different courts might reach different results—is not enough to justify certification under this sub-provision. As with Rule 23(b)(1)(A), absent class members do not have a right to opt out of a (b)(1)(B) class.

This is not a strong basis for class certification in the AI context. The paradigmatic use of (b)(1)(B) is—or was—where a defendant’s resources are inadequate to meet the substantial legal claims against it; in the absence of a class action, the first plaintiffs to reach judgment will deplete the available funds at the expense of those later in line. But forcing all affected parties to sue as part of a class raises due process concerns, especially given the lack of an opt-out right. In Ortiz v. Fibreboard, the Supreme Court cast a shadow on this type of class action. It rejected a settlement of an asbestos class action that would have been confined to the available insurance proceeds of the defendant (and not every available dollar); and it warned against “adventurous application” of Rule 23(b)(1)(B).²⁶⁸ As a result, courts rarely certify limited fund classes; like its counterpart (b)(1)(A), this subtype is predominantly used in ERISA suits.²⁶⁹

Courts have not certified copyright class actions under Rule 23(b)(1). In WB Music Corp. v. Rykodisk, Inc., the court refused such certification, noting that because each class member’s infringement claim was particular to them, resolution of one would not impair the rights of others.²⁷⁰ Similarly, the First Circuit rejected an attempt by a graphic artist to sue a defendant class made up of wholesalers and retailers who infringed the copyright to her wallpaper design.²⁷¹ In essence, the court found, the plaintiff was arguing, incorrectly, that stare decisis was sufficient to qualify as a prejudice requiring class resolution; the court was unpersuaded.²⁷²

23(b)(2): This class action category, referred to as the “civil rights” provision, applies where “the party opposing the class has acted or refused to act on grounds that apply generally to the class, so that final injunctive relief or corresponding declaratory relief is appropriate respecting the class as a whole.”²⁷³ This provision is a vehicle for obtaining declaratory and injunctive relief, rather than money damages.²⁷⁴ As with 23(a), there is no right to opt out of a certified (b)(2) class: a decision is binding on all those who fall within the class.

Rule 23(b)(2) may provide relief to groups of copyright owners in two ways. First, (b)(2) certification can provide the basis for class-wide injunctive relief—that is, it may be invoked to prevent future infringement. Second, parties and courts may invoke (b)(2) in so-called bifurcated proceedings. Under Rule 23(c)(4), a court may certify a class only for certain issues. Using this issue-class provision, a court could certify a class under (b)(2) for purposes of determining liability, while leaving damages or other downstream claims to be resolved through a different, perhaps individual, mechanism.²⁷⁵

Two early copyright class actions illustrate this. In one of the cases (b)(2) certification was not appropriate given the plaintiffs’ request for substantial damages awards.²⁷⁶ But the plaintiffs in Ryan v. CARL Corp. were successful in obtaining certification under (b)(2), despite potential for money damages. The Ryan class requested that the court bifurcated issues into liability—for which (b)(2) was appropriate—and damages, which could be adjudicated later, if necessary, either in the form of a (b)(3) class or in individual proceedings.²⁷⁷ The court in Ryan agreed: it approved (b)(2) certification given the “uniformity” of the defendant’s allegedly infringing practice and stayed consideration of a damages class for another day.²⁷⁸

Yet courts have been skeptical of (b)(2) copyright claims where plaintiffs cannot show a high degree of uniformity in the defendant’s relationship with class members. In Kihn v. Bill Graham Archives LLC, the Ninth Circuit reversed a district court decision certifying a class of musicians alleging copyright violations by a company that allows users access to recordings of live concerts from the 1950s to 1990s.²⁷⁹ The district court had certified the class under both (b)(2) and (b)(3); the appellate court rejected both.²⁸⁰ As to (b)(2), the Court of Appeals found that the result of certification—an injunction requiring the defendant to remove a whole category of recordings from its website—failed to recognize significant distinctions among class members.²⁸¹ For example, some of the musicians involved might have reached agreements with the defendant, and others may wish to do so in the future.²⁸² Under (b)(2), those musicians would have no opportunity to opt out of the class.²⁸³

23(b)(3): This final class action category, added in 1966, allows certification of a class for purposes of seeking money damages for individual class members.²⁸⁴ The pending AI copyright class actions all seek some form of money damages, which means that they will very likely need to contend with the thorny requirements of (b)(3). The Supreme Court has labeled (b)(3) “the most adventuresome innovation” of the modern Rule 23.²⁸⁵ As one scholar stated two decades ago, the (b)(3) standard is fuzzy at best; the rules for interpreting it “have received virtually no clarification from the Supreme Court, have bewildered lower courts, and have not attracted substantial scholarly scrutiny.”²⁸⁶ Notwithstanding the Roberts Court’s strong procedural bent, that remains largely true.

Because of inherent tensions between individuals’ right to their own day in court and the monolithic nature of a class determination, plaintiffs seeking (b)(3) certification must make additional showings to demonstrate that a representative action—rather than individual suits—is appropriate. Specifically, parties seeking (b)(3) certification must prove that common issues in the case predominate over individual issues, and also that a class action is superior to individual suits or other dispute resolution formats.²⁸⁷ Finally, a court certifying a class under Rule (b)(3) must also find that administering the class is feasible,²⁸⁸ although the Supreme Court has clarified that a court need not address trial manageability if it is certifying a class for purposes of settlement.²⁸⁹ Unlike in the other 23(b) categories, members of (b)(3) classes have the right to opt out of membership in the class.²⁹⁰

To show predominance, a proposed class must show that the class issues are not only common, but in fact are central—that they are the driving force in the dispute. Not surprisingly, therefore, proposed class suits with questionable commonality and typicality will inevitably founder on the extra predominance requirement.²⁹¹

In Google Book Project, for example, the district court used Rule 23(b)(3) to certify a class of copyright holders in their suit alleging that Google’s project to digitize and create a scannable index of books was a “massive copyright infringement.”²⁹² The court found that “[e]very potential class member’s claim arises out of Google’s uniform, widespread practice of copying entire books without permission of the copyright holder and displaying snippets of those books for search.”²⁹³ Google’s fair use defense, the court found, could also be resolved on an aggregate basis.²⁹⁴ The district court also found that a class action was a superior method of resolving the claims against Google, saving both time and money.²⁹⁵ On appeal, the Second Circuit vacated the certification order.²⁹⁶ The appellate court did not take specific issue with the district court’s analysis, although it intimated some skepticism.²⁹⁷ Rather, it ordered the district court hold its class certification in abeyance until after it resolved the fair use question.²⁹⁸

Class actions involve competing narratives—connectivity v. entropy. Copyright certification decisions reflect this tension. For example, in opposing (b)(3) certification, Google argued that copyright ownership was too varied and complex for resolution on a class-wide basis.²⁹⁹ But the district court found that those individual issues were subsidiary to the central question of the legality of Google’s book project.³⁰⁰ In contrast, in Waite v. UMG Recordings, Inc., the district court denied (b)(3) certification on the ground that individual issues—specifically, whether the musical compositions at issue were “made for hire”—prevented resolution of the class claims on common proof.³⁰¹ Resolution of these same connectivity v. entropy questions will determine class certification in the AI context.

C. Copyright and Class Action Settlements Under Rule 23(e)

Settlements of class actions are different—and more onerous—than the private settlements that are traditional in litigation. Because any settlement will bind absent class members to a judgment, Rule 23(e) requires a court to ensure that those absent class members have been adequately represented and that the proposed settlement is “fair, reasonable, and adequate.”³⁰² Reflecting the quasi-regulatory nature of class actions, the end-product of Rule 23(e) is a hybrid creature between a court judgment and a private settlement contract.³⁰³ Parties first bargain with each other; then they must persuade the court, over the voices of any objectors, that the settlement is worthy of a judicial stamp of approval.

This approval requires multiple steps. First, a court must determine that a proposed settlement class meets the Rule 23(a) and (b) requirements described above; the Supreme Court expressly admonished litigants and courts that the requirements apply—perhaps with even greater import—in the settlement context.³⁰⁴ Second, Rule 23(e) mandates that courts give appropriate notice to all class members.³⁰⁵ It must also hold a “fairness hearing” to evaluate the settlement, and to allow objectors to the proposal to have their say.³⁰⁶ In evaluating whether a proposed settlement is adequate, a court must consider, inter alia: whether the representatives fairly represented the class; whether the proposal was negotiated at arms’ length; and whether the relief provided for the class is adequate, taking into account factors such as the costs and risks of proceeding to trial and the method proposed for remitting the relief provided by the settlement (money damages or otherwise) to class members.³⁰⁷

Central to this fairness determination is an evaluation of whether there is a significant conflict among different subsets of the class. In Amchem Products, Inc. v. Windsor—the Supreme Court’s master blueprint for class action settlements—the Court rejected a global asbestos settlement on the ground that the agreement impermissibly privileged class members with current asbestos-related injuries over those whose injuries would only manifest in the future.³⁰⁸

Notwithstanding the relative paucity of copyright class actions, copyright class settlements have played an influential role in establishing the parameters for technology companies’ use of intellectual property. The Second Circuit’s analysis of the proposed settlement in In re Literary Works provides an apt illustration.³⁰⁹

Following the Supreme Court’s decision in Tasini,³¹⁰ the Second Circuit evaluated a proposed consolidated class settlement in a suit by authors whose works had been placed in electronic databases without their permission.³¹¹ The district court had referred the case to mediation; three years later, the parties had reached an agreement.³¹² They jointly moved the district court to certify the class and approve the settlement.³¹³ In re Literary Works contains two lessons for AI class actions:

1. Class Settlements Must Not Mask Significant Intra-Class Conflicts

The Second Circuit initially rejected the proposal on the ground that it did not adequately represent certain class members.³¹⁴ Under the terms of the proposed settlement, class members were divided into three classes.³¹⁵ Category A included class members who hold registered copyrights making them eligible for statutory damages.³¹⁶ Category B included those with copyrights registered too late to render them eligible for statutory damages, and Category C included those who have not yet registered their copyrights—a group which composed 99% of the total claims.³¹⁷ The Court found that there had been inadequate representation of the Category C group.³¹⁸ Noting that the interests of Category C should have been protected by a sub-class represented by independent counsel,³¹⁹ the Court remanded with instructions for the creation of three separately represented sub-classes.³²⁰ On remand, the district court approved a settlement with those sub-classes.³²¹ As discussed below, some of the pending AI suits have defined the class to include only creators who have timely registered their copyrights; other suits, however, are including both registered and non-registered copyright holders in the class, thus raising the likely need for sub-classes.³²²

2. Class Settlements May Release Defendants from Liability for Future Actions

Objectors to the proposed settlement in In re Literary Works also challenged the scope of the settlement release, which applied to claims that were beyond those that had been pleaded in the class action.³²³ Under the terms of the settlement, class members who did not opt out were prohibited from barring future use of their works by the defendant publishers, including the licensing of those works to third parties.³²⁴ Objectors argued that future infringements would be new and distinct harms, and they also objected to the provision allowing future sub-licensing.³²⁵ But the Second Circuit overruled the objections, finding that the settlement’s scope was acceptable and that the terms were reasonably within the factual predicate of the underlying suit.³²⁶

There may be limits to courts’ tolerance of future releases in some class settlements.³²⁷ In the proposed settlement of Google Book Project, for example, the district court rejected a release outlined in a 166-page settlement agreement, finding the scope of the release of future claims too broad—too unpredictably constraining on class members.³²⁸ But the court’s rejection of that settlement may be an outlier.³²⁹ Courts have approved copyright class settlements with releases in many other contexts.

D. Administrative Rule 23 Requirements

In addition to satisfying Rule 23(a) and one of the categories of Rule 23(b), a class action must be administrable—that is, a court must find that the parties are capable of implementing a judgment in a way that is fair to absent class members. Several practical elements of Rule 23 are particularly relevant to copyright class actions. Although pre-AI copyright opinions do not expressly address all these administrative elements, they are discussed here for purposes of setting forth the main points of contention in many certification disputes.

1. Defining the Class: Ascertainability and the “Fail-safe Class”

Over the past decade, some federal courts have read an “implied requirement of ascertainability” into Rule 23.³³⁰ The basic concept is not controversial. To be certified, a class must be “sufficiently definite so that it is administratively feasible for the court to determine whether a particular individual is a member.”³³¹ To ensure this, a class must be defined by objective criteria.³³² Sometimes this can be complicated. For example, in a consumer class action, it may be difficult to determine by an objective standard who purchased a particular food item or cosmetic.

As it is applied by some courts, ascertainability can interpose a high, seemingly arbitrary barrier to class certification.³³³ The Third Circuit is the leader of this aggressive approach.³³⁴ That Court defends its rigorous ascertainability requirement on the ground that it protects absent class members, ensures due process to the defendant, and eases the cost and administrative burden of an ill-defined class.³³⁵ This is purportedly necessary in order to determine the preclusive scope of a judgment—“who gets the benefit of any relief and who gets the burden of any loss.”³³⁶ The term ascertainability is nowhere in the text or rulemaking commentary to Rule 23.

But the Circuits do not agree on the nature and scope of an ascertainability requirement—or even on whether it really exists—and the Supreme Court has yet to weigh in. The Second and Ninth Circuits, where most of the major AI class actions are pending, have both expressly rejected the Third Circuit’s ascertainability requirement as grounded in policy rather than the text of Rule 23.³³⁷ Therefore, while ascertainability—sometimes called “administrative feasibility”—may pose a challenge to class action litigants, the AI class actions are not likely to founder on this ground.

Related to ascertainability is the argument that a court should not certify a “fail-safe” class.³³⁸ A fail-safe class is one in which membership in the class depends on whether a party has a valid claim.³³⁹ In other words, if a class member loses their claim for relief, then by definition they were never properly a member of the class, and therefore are not bound by res judicata.³⁴⁰ Neither the Second nor the Ninth Circuits has ruled on the existence or import of the concept of a “fail-safe” class.³⁴¹ In any case, however, as discussed below, the AI plaintiffs are not proposing such an ipse-dixit, preclusion-proof class. If courts in the pending AI cases certify classes of copyright holders, the members of those classes—that is, those creators who do not exercise their right to opt out³⁴²—may either prevail or take nothing, depending on courts’ views of the merits.

2. The Order of Operations

Under Rule 23, district courts have wide discretion as to how they manage class actions.³⁴³ Therefore, while Rule 23 provides that courts should assess class certification at “an early practicable time,”³⁴⁴ many courts allow certification to be decided after decisions—or even judgment—on the merits.³⁴⁵

This procedural flexibility inevitably interjects another axis of strategy and uncertainty into class actions, including the pending AI copyright cases. For example, Texaco, discussed above, was brought as a class action. However, the parties stipulated to try the question of fair use prior to addressing class certification. Once the Second Circuit denied Texaco’s fair use defense—basically ensuring that Texaco would be held liable for copying research publications for its employees—the case settled.³⁴⁶

As discussed below, litigants in the pending AI copyright class actions have taken different procedural pathways. Those decisions have pragmatic and substantive implications.

V. Certification of AI Input Classes

>Against the above background, multiple content creators have filed proposed class actions in federal courts against generative AI companies, seeking relief for various forms of copyright infringement, sometimes accompanied by state law claims. These lawsuits have sprung up primarily in two locations: New York and California, although two are in Delaware.³⁴⁷ The New York cases, all in the Southern District, include actions brought by fiction authors, non-fiction authors, and journalists.³⁴⁸ On the west coast, several suits—including those brought by visual artists,³⁴⁹ music publishers,³⁵⁰ coders,³⁵¹ and additional suits by authors³⁵²—were brought in California; all are pending in the Northern District.

In addition to the suits seeking formal class certification, there are a handful of cases against AI companies brought by single organizations or corporations that support, aggregate, and disseminate the work of a large number of creators. These include cases by The New York Times, the Center for Investigative Reporting, and Getty Images.³⁵³ Because these suits do not implicate the technical or logistical complexities of Rule 23, courts in these suits may reach decisions on the merits, including the applicability of fair use, earlier than their class-action cousins. If that happens, those precedents will influence the default expectations—and therefore the settlement options—of AI litigants across the board.

In the class suits, the parties’ litigation strategies reflect those adopted by early copyright class litigants discussed above.³⁵⁴ The plaintiff creators argue that the defendants’ rapacious copying of literary works is the defining conduct in the dispute, providing the basis for class certification and liability. As their complaint states, “OpenAI and Microsoft have built a business valued into the tens of billions of dollars by taking the combined works of humanity without permission.”³⁵⁵ In contrast, the defendants—while trying to evade liability entirely based on fair use³⁵⁶—have generally made blanket denials regarding the class-wide actions necessary to support class certification.³⁵⁷ The success of these opposing strategies will largely depend on courts’ views of the underlying intellectual property claims. As is always true of class actions, procedure and substantive law are intertwined. But even if the copyright claims appear to chart new ground—something we dispute—the class certification issues are, mostly, old hat.

For purposes of analyzing the likelihood of obtaining class certification, we divide AI claims into two categories. In the first category, input claims, the plaintiffs have a strong claim for class certification.³⁵⁸ The plaintiffs’ input claim is that AI companies committed direct copyright infringement by duplicating copyrighted works, without asking permission or offering compensation, in order to “train” their AI foundation models. To be sure, these are not slam-dunk claims; the AI defendants are straining mightily for a fair use defense, and the plaintiffs must obtain evidence proving which works were copied—evidence that they claim the defendants possess. Nevertheless, we believe these input claims are not only meritorious but also that they fit solidly within the Rule 23 class action framework. By way of illustration, we describe two input class claims below.

Several of the pending AI class suits seek a remedy not only for the initial copying of their works, but also for the “market usurpation defendants have enabled by making Plaintiffs unwilling accomplices in their own replacement.”³⁵⁹ We refer claims based on this market usurpation as output claims.³⁶⁰ Output claims, often alleging derivative infringement³⁶¹ or asserting violations of the Digital Millenium Copyright Act, allege that AI companies are using creators’ content—whether computer code, plot devices, literary text, images, or even name—to instantly produce and market derivative content without permission, attribution, or compensation. Although output claims are not the focus of this Article, we note that they are likely to require more individualized evidence and be subject to more individualized analysis (for example, of fair use, or substantial similarity). This is likely to make class certification of such claims challenging.³⁶² In all likelihood, however, the strength of those claims will never be tested: the lingering uncertainty and risk for all parties will provide a strong incentive to reach a settlement of all outstanding claims.

Below are two illustrative examples of input claims where we believe class certification is appropriate. As these examples demonstrate, AI copyright classes are not cookie-cutter. To the contrary, parties and attorneys have adopted quite different litigation strategies in these cases.

A. The Input Class: Direct Copyright Infringement in Authors Guild v. OpenAI, Inc.

In Authors Guild v. OpenAI, a professional writers’ organization³⁶³ as well as two groups totaling approximately 30 plaintiffs—including best-selling fiction authors such as George R. R. Martin and Jodi Picoult, as well as non-fiction authors including Stacy Schiff and Jia Tolentino—seek to represent classes of fiction and non-fiction authors in consolidated suits against Open AI and Microsoft.³⁶⁴ Among other claims, the plaintiffs allege that OpenAI and Microsoft committed direct copyright infringement when they scraped class members’ works in order to “train” their AI large language models GPT-N and ChatGPT.³⁶⁵

The consolidated complaint defines the fiction class as “[a]ll natural persons in the United States who are the sole authors of, and legal or beneficial owners of Eligible Copyrights in, one or more Fiction Class Works; and all persons in the United States who are the legal or beneficial owners of Eligible Fiction Copyrights in one or more Fiction Class Works held by literary estates.”³⁶⁶ The non-fiction class definition is the same for non-fiction copyright owners. For purposes of these claims, eligible class members are limited to those who timely registered their copyrights. In their Answers, OpenAI and Microsoft repeatedly deny that any works were infringed.³⁶⁷

In their pleadings, the defendants also assert blanket denials regarding class certification.³⁶⁸ Assuming that the plaintiff authors succeed in defeating a motion for summary judgment on the question of fair use, however, we believe plaintiffs’ direct infringement claims merit class certification.

1. Order of Operations—Fair Use First

As an initial matter, the parties to the consolidated Authors Guild suit—now in the discovery process in the S.D.N.Y.—followed the blueprint established in Texaco and negotiated an agreement to allow the district court to first resolve the defendants’ fair use defense, and only then to take up class certification if necessary.³⁶⁹ In exchange for this concession by the plaintiffs, the defendants agreed to forego filing Rule 12(b)(6) motions to dismiss, and to comply with an expedited briefing and discovery schedule.³⁷⁰

However it is resolved, the Authors Guild court’s fair use ruling, or a similar ruling in another case, will set the tone for the remainder of the case. As in Texaco, a rejection—even a partial one—of fair use will put the authors in a strong settlement position. A recent decision by the District of Delaware rejecting an AI company’s fair use defense is one such positive development for creators. In Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., Thompson Reuters, owner of legal search engine Westlaw, alleged that Ross had unlawfully copied Westlaw’s headnotes in order to train Ross’s competing product.³⁷¹ Ross had initially sought to license Westlaw’s headnotes. When Westlaw refused, Ross hired a company to create “Bulk Memos” to train Ross’s AI model; the Bulk Memos were created from Westlaw headnotes.³⁷² The district court found over 2,000 acts of direct infringement.³⁷³

The court rejected Ross’s claim of fair use. It held that “using Thomson Reuters’s headnotes as AI data to create a legal research tool to compete with Westlaw” was not transformative.³⁷⁴ Further, it found that the copying could have an effect on Westlaw’s market for its own current and future products.³⁷⁵ The court therefore granted summary judgment to Thompson Reuters/Westlaw as to the specific infringed headnotes.³⁷⁶

While this is a promising example of a court rejecting a fair use defense by an AI company for training on copyrighted material, the decision in Thompson Reuters is limited by its context: legal search engine competitors.

If the court in Authors Guild does not reach the same conclusion, however, and it finds that the defendant’s “training” constitutes fair use, either for all authors or (for example) only for the non-fiction authors, the settlement dynamics will shift in favor of the AI companies. Neither ruling is likely to destroy the incentive to reach some sort of settlement in these cases; the possibility of appellate or Supreme Court reversal will remain a threat. But the fair use determination will set the tone.

That said, no order of operations is risk-free here. Copyright class action plaintiffs on the west coast have opted for the traditional certification-first route. In Tremblay v. OpenAI, Inc., the defendants have already succeeded in dismissing some of the plaintiffs’ claims,³⁷⁷ although not the direct infringement claims.³⁷⁸ That dismissal will narrow the scope of potentially valuable discovery. Litigating class certification is complex and costly. In addition, the presence of a potential fair use defense may muddy the certification analysis. The possibility that the defendants have a valid fair use defense as to even some class members might hinder a court from finding that the plaintiffs have met their burden to show Rule 23 commonality and typicality. In contrast, the Authors Guild plaintiffs will have the Rule 23 wind at their back if they survive a fair use motion for summary judgment.

The decision to resolve fair use prior to certification has other collateral consequences as well, because there are competing class actions in different jurisdictions, with potential overlap in class members. Before class certification, absent class members have not yet been joined to a suit.³⁷⁹ If the Authors Guild case suffers a serious setback prior to certification, or if certification is delayed, absent class members may migrate to Tremblay or a different case that reaches class certification and settlement sooner.

2. Authors Guild Satisfies Rule 23(a) Factors

To succeed on certification, plaintiffs must show that their class claims meet the four Rule 23(a) requirements of numerosity, commonality, typicality, and adequacy of representation.³⁸⁰ As one scholar has noted, these quippy labels “mask the complexity of the case law and scholarship that has accreted over generations to give these terms their meanings.”³⁸¹ Notwithstanding this complexity, and assuming that fair use is not an obstacle, the Authors Guild plaintiffs have strong arguments for all four 23(a) factors.

Rule 23(a)(1)—the requirement that the class be “so numerous that joinder of all members is impracticable”³⁸²—is satisfied. Notwithstanding the weak protests of the defendants,³⁸³ numerosity is not seriously in doubt.³⁸⁴ Relatedly, by limiting the class membership to creators in the U.S. with registered copyrights,³⁸⁵ the plaintiffs have defined the class in a way that is objective and verifiable, thus satisfying any potential ascertainability arguments. (Note: This is not necessarily true of the Tremblay plaintiffs in the Northern District of California, who are seeking to represent all copyright holders, whether registered or not.³⁸⁶)

The Authors Guild plaintiffs can also likely meet their burden to show commonality under Rule 23(a)(2). Predictably, the plaintiffs’ theory of the case centers on the unifying and uniform act of AI companies duplicating copyrighted works for purposes of “training” large language models. Specifically, the Consolidated Complaint alleges that “[d]efendants copied and data-mined the works of [class members], without permission or compensation.”³⁸⁷ In so doing, the plaintiffs allege, the registered copyrights of all class members “were infringed in materially the same way.”³⁸⁸

At some level, this allegation may appear to be circular; it begs the question of what constitutes a material infringement. But the plaintiffs’ claims are more straightforward than that: they allege that the AI defendants unlawfully and directly “copied Plaintiffs’ works and then fed them into their ‘large language models.’”³⁸⁹ This is a simple, unifying claim of liability for direct infringement. Whether or not the defendants’ actions are in fact unlawful, copying is the central question in the case for all class members.³⁹⁰

Plaintiffs’ direct infringement claim also satisfies Rule 23(a)(3)’s typicality requirement, which tends to merge into the commonality analysis.³⁹¹ Typicality asks whether the claims of the class representatives align with those of the absent class members. In this case, the question is whether the fiction author representatives’ claims mirror those of absent fiction authors, and whether the same dynamic is true for the non-fiction representatives vis-à-vis the absent non-fiction class members. It appears that they do. The class representatives allege that the defendant AI companies used all class members’ copyrighted works in the same way and for the same general (albeit broad and evolving) purpose.³⁹² Beneath the technological complexity of AI, the direct infringement claims are simple and uniform—unlike the “Frankenstein” claims that the district court rejected in Football Association. In fact, the Second Circuit has found typicality in class actions with far more divergence among class members’ experiences than exists here.³⁹³ Moreover, unlike in Football Association, the classes in Authors Guild are defined concretely and objectively. They are limited to persons in the United States who possess (or are beneficial owners of) registered copyrights.³⁹⁴ The claims are cohesive.

Finally, and for many of the same reasons, the Authors Guild classes satisfy Rule 23(a)(4) adequacy of representation. As just discussed, it does not appear that either the Authors Guild or the individual class representatives have interests that are misaligned with those of the absent class members. In the Google Book Project case, Google alleged that many plaintiff class members supported Google’s efforts and disagreed with the class claims, thus calling adequacy of representation into question. Although it decided the case on other grounds, the Second Circuit acknowledged that Google’s argument “may carry some force.”³⁹⁵ As of now, however, there is no significant outcry by creators who favor giving their intellectual property to AI companies at no cost. To be sure, there are academics and policymakers who favor AI companies’ free use of copyrighted material.³⁹⁶ But there is as of yet no sign of a serious fissure among creators that would undermine adequacy of representation. In fact, the Library Copyright Alliance, which supports a finding of fair use for AI copying, submitted comments to the Copyright Office urging that office to defer to courts’ resolution of these questions.³⁹⁷ As to the second element of (a)(4) adequacy—the quality of the lawyering—there is no serious likelihood that the court or the defendants will argue that the renowned attorneys representing the class are in any way inadequate.³⁹⁸

At first glance, it may also appear that the fame and wealth of named plaintiffs like John Grisham and Jodi Picoult could undermine typicality and/or adequacy in a class composed largely of minor authors. After all, Grisham and other best-selling authors might argue that the defendants’ actions resulted in vastly larger damages. But that should not prove to be a barrier. Copyright law provides for two possible forms of damages: actual and statutory.³⁹⁹ It is often difficult to calculate actual damages. In lieu of that calculation, a plaintiff may elect at any time before final judgment to receive statutory damages for each act of infringement.⁴⁰⁰ The precise amount of a statutory damages award depends on a court’s assessment of the seriousness of the infringement and the defendant’s ability to pay; it does not depend on a showing of the damages suffered by the copyright owner.⁴⁰¹

The Authors Guild Complaint notes that, should the class elect statutory damages, “the damages inquiry will likewise be common, if not identical,” among all members.⁴⁰² The simplicity of statutory damages bolsters the classes’ claims of commonality, typicality, and adequacy. It also raises the specter of a massive damages award.

3. Authors Guild Satisfies Rule 23(b)

The Authors Guild complaint asserts that class certification would be appropriate under all of the Rule 23(b) categories.⁴⁰³ In addition, it alleges that, as an alternative to certification of the entire dispute, the court could bifurcate the proceeding into liability and damages under its power to certify a class only as to certain issues.⁴⁰⁴ This would entail first certifying a liability-only class under Rule 23(b)(2), for the purposes of resolving the common question of whether the defendants are liable for infringement for copying class members’ works,⁴⁰⁵ while leaving the question of damages to another process on another day.

Despite the class certification laundry list in the pleadings, however, the plaintiffs’ most straightforward and effective option is to obtain certification under the most rigorous of the Rule 23(b) options, Rule 23(b)(3), and they are in a strong position to do that on their direct infringement claims. This form of certification had distinct advantages. First, it will allow any objecting authors to opt out of the class and either pursue claims elsewhere or drop any potential claim—thus minimizing the force of any arguments that the class members are not united. Second, unlike (b)(2) classes, in which class members may only obtain declaratory or injunctive relief, (b)(3) classes may seek money damages.

As explained above, to obtain certification under Rule 23(b)(3), the Authors Guild class representatives will need to show what are referred to in shorthand as predominance and superiority. This they can do.

Rule 23(b)(3) Predominance

The predominance requirement requires plaintiffs to show that the common issues are at the heart of their case, and that any differences among plaintiffs are of only secondary importance. As discussed above, this provision is pragmatic rather than formalistic; courts have discretion as to how they interpret and apply it.⁴⁰⁶

But predominance should not be an insuperable barrier to certification of input copyright classes. As of now, the Authors Guild plaintiffs’ strong arguments in favor of commonality and typicality carry over directly to this predominance argument: The plaintiffs are alleging identical conduct by the AI defendants, and they have invoked their right to seek statutory damages, which would minimize the need for thousands of fact-specific individual determinations.

Rule 23(b)(3) Superiority

The superiority requirement asks whether a class action is the best option for resolution of the claims at issue.⁴⁰⁷ Rule 23(b)(3) lists five non-exclusive factors that courts should consider in evaluating superiority, including whether individual class members would have an interest in proceeding separately; the nature of any ongoing litigation in other courts; the appropriateness of the forum, and potential difficulties in managing a class suit.⁴⁰⁸

As one scholar has recently observed, courts apply this requirement in different and sometimes haphazard ways.⁴⁰⁹ But the murkiness of superiority in class action doctrine writ large should not detract from its application to this case: a class action is the superior way to resolve the case of thousands of identically situated creators and to give repose to AI defendants. The fiscal and logistical impossibility of individual small-time authors taking on AI companies makes it applicable.

Moreover, the district court in this case has already, albeit indirectly, analyzed some of the superiority factors. The west-coast plaintiffs in Tremblay moved to intervene in Authors Guild for the purpose of moving to dismiss, stay, or transfer the New York action to the Northern District of California.⁴¹⁰ In denying the motion, the district court stated its view that the Authors Guild classes had brought suit in an appropriate forum.⁴¹¹ The court’s order also makes clear its view that the “nature of any litigation concerning the controversy already commenced by or against members of the class is not a barrier to the New York proceedings.⁴¹² This finding closely parallels the superiority finding required under Rule 23(b)(3)(B).⁴¹³

B. The Input Class: Andersen v. Stability AI, Ltd.⁴¹⁴

The Authors Guild case concerns AI models’ training on text; Anderson v. Stability AI is about images. The named plaintiffs in this proposed class action—artists including Karla Ortiz, Gerald Brom and Jingna Zhang⁴¹⁵—allege that defendants Runway, Stability AI, DeviantArt, and Midjourney downloaded billions of copyrighted images without permission for the purpose of creating and/or using an AI product called Stable Diffusion.⁴¹⁶

Specifically, the plaintiffs claim that Runway worked with Stability to help train and then distribute Stable Diffusion, an AI image generator whose work products are “in the style of” the named plaintiffs and other artists.⁴¹⁷ The training images came from so-called LAION datasets.⁴¹⁸ Midjourney trained on Stable Diffusion, and DeviantArt used the model in its own image-generating product, DreamUp. In addition, plaintiffs allege that defendant Midjourney “has repeatedly promoted the use of artist names—including Plaintiffs’ names—within text prompts as a means of getting better results.”⁴¹⁹ These companies now market and sell AI image products; the plaintiffs allege that “what they’re really selling is copyright infringement as a service.”⁴²⁰

The named plaintiffs have alleged several causes of action against each defendant. In addition to their input claim—direct copyright infringement—they assert claims for induced copyright infringement, vicarious copyright infringement, violation of the DMCA and the Lanham Act, as well as common law claims for unjust enrichment.⁴²¹ They characterize defendants’ AI image-generating products as “copyright-laundering devices, promising customers the benefits of art without the costs of artists.”⁴²²

Unlike in Authors Guild, the parties in this suit did not stipulate to a non-traditional order of operations. Instead, the defendants moved to dismiss for failure to state a claim. Details aside, the plaintiffs’ claims largely survived this motion. The district court found that “the plausible inferences at this juncture are that Stable Diffusion by operation by end users creates copyright infringement and was created to facilitate that infringement by design.”⁴²³

With regard to the specific claims, none of the defendants moved to dismiss the direct copyright infringement claims based on scraping of and “training” with copyrighted images. The court rejected defendants’ motion to dismiss plaintiffs’ induced copyright infringement claims.⁴²⁴ For all defendants, the court dismissed the plaintiffs’ DMCA claims with prejudice, and their common law unjust enrichment claims with leave to amend.⁴²⁵ The court ruled variously on the other claims against each defendant:

• Stability AI: the district court granted the defendants’ motion to dismiss the plaintiffs’ DMCA claims as well as their unjust enrichment claims.⁴²⁶ The court declined to dismiss the plaintiffs’ induced infringement claim.⁴²⁷
• Runway: the district court denied Runway’s motion to dismiss direct infringement claims based on plaintiffs’ argument that Stable Diffusion is itself an “infringing statutory copy” of plaintiffs’ works, and/or that distributing Stable Diffusion is the equivalent of distributing plaintiffs’ works.⁴²⁸ It also denied Runway’s motion to dismiss the induced infringement claims.⁴²⁹
• Midjourney: the district court denied Midjourney’s motion to dismiss plaintiffs’ copyright claims; Midjourney had contended that the plaintiffs’ pleadings did not plausibly allege that the artists’ work was covered by registered copyrights.⁴³⁰ The court also denied Midjourney’s motion to dismiss plaintiffs’ Lanham Act trademark claims.⁴³¹
• DeviantArt: the district court denied DeviantArt’s motion to dismiss copyright claims.⁴³² It granted the motion to dismiss plaintiffs’ breach of contract claims and unjust enrichment claims.⁴³³

The named representatives in Andersen are seeking class certification under Rules 23(b)(2) and (b)(3).⁴³⁴ They break their (b)(3) class into three sub-classes, based on from which database the copyrighted images were downloaded.⁴³⁵ They also seek separate certification of a (b)(3) class composed of all artists whose names were invoked in prompts by the Midjourney Image Product.⁴³⁶ These multiple class claims are more nuanced than those in Authors Guild. But the Rule 23 analysis for direct infringement is functionally the same.

Discovery is ongoing. Based on currently available information, however, the plaintiffs can satisfy Rule 23(a) and (b) for their direct copyright claims. As with Authors Guild, numerosity is not controversial: the plaintiffs allege that billions of copyrighted artistic works, paired with descriptive captions, were copied without permission.⁴³⁷ And as with Authors Guild, the driving force of the litigation—defendants’ allegedly unlawful replication of plaintiffs’ copyrighted works—meets the commonality and typicality requirements.⁴³⁸ The First Amended Complaint states the common question as “whether Defendants violated the copyrights of Plaintiffs and the Class when they downloaded and stored copies of [class members’] Works…and…used copies of the Works to train AI Image Products.”⁴³⁹ This allegation of centralized and uniform conduct is the heartland of commonality and typicality. Finally, as of now there are no allegations that either the named representatives or the highly competent attorneys representing the class are inadequate to represent the rights of absent class members.

The logic of Authors Guild also supports certification of the Andersen plaintiffs’ (b)(2) and (b)(3) classes for direct copyright infringement. Again, the centerpiece of certification under both subprovisions is the uniform conduct of the defendants—conduct that affects the plaintiff visual artists in more or less identical ways. In fact, while analytically Authors Guild and Andersen are very similar, it is likely to be easier—more intuitive—to show copying of images than it is of texts.⁴⁴⁰

Of course, as with Authors Guild, the plaintiffs in Andersen will need to prevail against a fair use summary judgment motion. But, as explained below, we argue that fair use should not be extended to give enormous, for-profit technology companies free license to replicate copyrighted works.

VI. Stealing is Not Fair Use

In the early days of computers, Steve Jobs was known for bragging, “Good companies copy, great companies steal,” while Bill Gates begged hobbyists to stop copying Microsoft software without paying. Ironically, Apple subsequently sued Microsoft for stealing the displays of Apple’s Macintosh computer user interface in Windows 2.0. Perhaps both companies are great, and stealing is at the heart of Big Tech, both then and now.

Fair use, however, was not at issue in that stealing case. Stealing is and should never be deemed fair, as both great companies understand the fundamentals of good business and corporate reputation. Wholesale stealing, in particular, like what gave rise to the AI input class actions, is never fair.⁴⁴¹

The defendants in the AI input class actions are riding on their fair use defense to avoid infringement liability. In addition to Oracle, they will rely on Field v. Google and Authors Guild v. Google for support. Unlike Google in these three cases, the AI companies are not engaging in “caching” webpages, nor indexing snippets for users to search, nor using a miniscule amount of essential code; instead, they are brazenly stealing wholesale literary text and images to train their models. AI models need copyrighted content and without the creative works of authorship input into the AI models, there will be no AI models. The stealing itself is not transformative. The act of copying or scraping creative works belonging to others is not transformative. The act at that moment does not create anything new to benefit anyone except AI models that have been valued at billions of dollars and enriched Big Tech and Big AI companies. Without acknowledging the existence of copyrights and without compensating creators, the stealing avoids payments to creators and disregards existing human creativity.

Furthermore, like the defendant in American Geophysical Union v. Texaco, Inc. who made copies of academic articles for their corporate use and faced the court’s ruling that the act of copying the articles was not transformative, AI defendants copy the entirety of literary text and images for their own corporate use of building their AI models likewise is just plain copying without pay. Also, the stealing in the AI input class constitutes some exceptions under copyright law as seen in the MAI Trio cases, because the stealing is massive, permanent, and intentional to build powerful models of artificial intelligence for worldwide domination.

At the moment when the stealing occurs, the input claim of copyright infringement is established. That means the act at that moment of copying and scraping does not transform the literary text or images into something else except AI models, which are trained to commit the second type of copyright infringement by violating the right to prepare derivative works. AI companies will attempt to conflate the input and output claims, injecting end users as a shield for liability. At the input stage, however, there are no users to enter the prompts because the AI companies are busy stealing copyrighted works to complete the training of the AI models. Users enter the scheme only after AI companies debut their AI models long after the AI models have ingested and trained on all the copyrighted works.

The second factor of fair use is the nature of the copyrighted works. The creators will easily satisfy this factor against fair use because their literary text and nonfiction contain the level of creativity entitled to protection in fair use analysis as discussed in Harper & Row v. Nation Enterprises.⁴⁴² Likewise, on the third factor, the amount and substantiality of the portion used in relation to the copyrighted work as a whole favors the creators because the AI defendants steal the copyrights in their entirety. The last factor, the effect of the use on the potential market for the value of the copyrighted work, also favors against a finding of fair use. Indeed, as we saw in the discussion of the last factor by the court in American Geophysical Union v. Texaco, Inc., the existence of a licensing model to establish royalty for creators makes stealing unfair. Also, AI defendants themselves openly admitted the licensing model for non-literary content as they paid Reddit millions for the license to use the platform’s user content.

Stealing is never fair to victims, and society. Stealing is never fair when Big AI and Big Tech with enormous resources are teaming up to extract the heart and soul of the creativity class. Fair use defense is not available shielding AI defendants in the input claim. AI defendants must pay under a licensing model.

VII. Fair Licensing

The Framers enshrined the Intellectual Property Clause to “promote the Progress of Science and useful Arts” into the Constitution by securing “for limited Times” to authors “the exclusive Right” to their “Writings.”⁴⁴³ At the time of the Framing, “promote the Progress of Science” referred to “the creation and spread of knowledge and learning” from works of authorship.⁴⁴⁴ The young nation needed authors and new works of authorship as expressed and embodied in literature, music, theatre, journalism, fine arts, and architecture for development and growth.⁴⁴⁵ The works of authorship, they believed—key contributors to humanity and to defining a new nation—must be cultivated and incentivized through uniform federal protection.⁴⁴⁶ Without legal protection, others could freely copy authors’ works, reducing the incentive to create works of authorship in the young nation and thereafter.⁴⁴⁷ Executing this constitutional mandate without delay, the new Congress passed the first federal copyright law during its first session in 1790.⁴⁴⁸

Copyright law embraces human creativity, and the level of creativity necessary for a work of authorship qualified for protection is minimal. The modicum of creativity means new works of authorship are constantly in the stream of creation from the Founding time to the present. As long as the works possessing a modicum of creativity are independently created by the authors and fixed in a tangible medium of expression, the works enjoy the exclusive rights enshrined in the Constitution. An email, a tweet,⁴⁴⁹ a poem, a song, a drawing, a painting, an essay, an investigative report, a photograph, a blueprint, a chevron design on a cheerleader uniform,⁴⁵⁰ a musical score, a dance, a recording, a video, a game, a homemade movie, a pamphlet, and a compilation of data,⁴⁵¹ regardless of its aesthetic appeals, high or low culture, captivating or distasteful, are entitled to federal protection. Human creativity, in all forms, drives the nation’s soul, as the Constitution’s “for the progress of Science” recognizes.

AI companies and their Big Tech backers, as seen in the class action complaints, brazenly take works of authorship, from literary to nonfiction, photographs, and images to music works, to develop their dataset. They steal to build their models. They disregard the constitutionally mandated copyright protection for human creativity as expressed in eight categories of works of authorship.⁴⁵² Their AI foundation models generate output infringing on copyrighted works and populating the online content with synthetic commons where fakes are quickly replacing human-created text, images, and sounds. The theft of human creativity should be addressed by compensating the authors for their losses. Before we discuss “how,” a detour on copyright law on damages is illuminating.

Federal protection of works of authorship encompasses a robust enforcement system against infringers. Indeed, copyright law accords the copyright owner with the option to elect statutory damages in the amount of no less than $750 and no more than $30,000 per work being infringed.⁴⁵³ If the infringer committed the act(s) willfully, the statutory damages per work can increase up to $150,000.⁴⁵⁴ Statutory damages are only available in copyrights, not other types of intellectual property infringements, because Congress recognized that actual damages and profits are difficult to establish by copyright owners, and desired to compensate the copyright owners for the losses to their creativity as expressed in works of authorship. Statutory damages also serve to prevent copyright infringements which are easy to accomplish by copyists. To obtain statutory damages, copyright owners must register their copyrights within three months of publication or before the infringement begins.

Returning to “the how,” we suggest fair licensing to compensate the creative working class who are the authors of copyrighted works of authorship being taken with neither permission nor compensation from AI companies and their Big Tech backers. What is licensing? What is fair?

Licensing is an old business model in the intellectual property field. Bill Gates sold licenses or the right to use Microsoft’s software which is protected under copyright law to end-users who are individuals and business entities.⁴⁵⁵ In demanding that people pay for his copyrighted software instead of using it without permission, Gates asked, “Who can afford to do professional work for nothing?” and sparked the licensing of software.⁴⁵⁶ Microsoft dictates the prices of different licenses, and the individuals and business entities pay. The prices are fair market value because Microsoft, the seller, and the individuals and business entities, the buyers, are unrelated, and they agree to engage in the transactions without pressure.⁴⁵⁷ In other words, Microsoft and other Big Tech companies know the licensing of copyrights well and their lawyers understand the clients’ licensing-centered business.

Moreover, the music industry is known for its savviness in licensing copyrights; their experience can shape licensing deals. Further, under the compulsory licenses for music, everyone can use others’ songs for programming, digital, and physical sound recording releases, without permission from the composition rightsholders but must pay the statutory mechanical rate for each copy sold on the song covered. The rate is 9.1 cents per song or 1.75 cents per minute of playing time, whichever is greater.⁴⁵⁸ Spotify, Apple Music, and other digital streaming services pay blanket licenses for mechanical royalties, and parties on both sides of blanket licenses can furnish the licensing terms as samples to AI copyright class action cases.⁴⁵⁹

Likewise, the number of licensing deals AI companies have secured with various social media platforms, media entities, and publishing companies intimates the current pricing lists of different types of content for AI companies to train their models.⁴⁶⁰ These deals demonstrate that AI companies know how to value and pay for the content types that they need. What constitutes fair market value is what the parties decide in the deals that they both have reasonable knowledge of the copyrighted works, the use of the copyrighted works, and their value, and they act under no undue pressure.

Regarding representation in negotiating for fair terms in licensing for the creative working class, the lawyers for the plaintiffs in the AI class action cases are sophisticated law firms. They are capable of negotiating with AI companies and their Big Tech backers for a fair market value of different licenses of copyrights. Using the statutory damages provisions as guideposts, $750 to $150,000 per work of authorship, the parties through their lawyers can reach sensible licensing terms to achieve the widely recognized fair market value concept.

Moreover, for the licensing to be fair to the creative class who typically possess neither resources nor sophistication for negotiation, we suggest three key terms for consideration. First, the scope of the license grant must satisfy four components: (i) the royalty setting must be fairly commensurate with the individual creator’s works, (ii) the license grant must preserve the author’s control over the use of their copyrights in subsequent technologies, unknown at the time of the license but later developed, (iii) the authors are entitled to receive notice of any proposed changes to the scope of the license grant, and (iv) and licensees’ unilateral expansion of the scope of the license grant constitutes intentional copyright infringement. These components will help tilt the balance from AI defendants/licensees to the creators.

Second, the creators must also retain control over their copyrights in addition to the control relating to the scope of the license grant. Here, the authorship of the creators’ works must be acknowledged. This requires appropriate attribution to the original human creators who created the works before machines.

Lastly, creators’ autonomy must be respected. That means creators have the right to opt-in or opt-out of the licensing scheme at any time during the duration of the licensing scheme.

Licensing, though not a perfect model, is the only viable option to righting the wrongs, the stealing by AI companies of works of authorship produced by the creative class. The massive stealing must end. With the three key terms outlined above, we believe that the lawyers of both sides can negotiate towards an agreement to conclude the AI input claims, the first in the two steps in the preservation of humanity.

Conclusion

Upholding copyright law in the AI age will not stymie innovation; this is not about fiction authors barring technology from accurately diagnosing cancer. Nor is upholding copyright law against AI companies administratively impossible. In a world where AI can instantaneously accomplish so many tasks, it can also accomplish the task central to the constitutional concept of copyright: preserving and respecting the small, individual, and fundamentally human contributions of creators to national and international culture. The pending AI suits allow these individual creators to aggregate to take on AI defendants—the ultimate technological aggregators. We believe that courts can, and should, use long-existing copyright law to recognize and vindicate the copyrights of the creative working class.

****

VIII. Appendix A

Conducted on October 1, 2024, 12:45 p.m.

Sent by you: generate multiple choice questions for sorting Hogwarts houses

Sent by Copilot: Sure! Here are some multiple-choice questions to help sort someone into a Hogwarts house:

1. Which quality do you value most?
• A) Bravery
• B) Loyalty
• C) Intelligence
• D) Ambition
2. What would you do if you found a lost wallet?
• A) Try to find the owner yourself
• B) Hand it over to the authorities
• C) Look for clues to find the owner
• D) Keep it for yourself
3. Which magical creature do you find most fascinating?
• A) Phoenix
• B) Hippogriff
• C) Thestral
• D) Basilisk
4. How do you prefer to spend your free time?
• A) Going on adventures
• B) Helping friends
• C) Reading books
• D) Planning your future
5. What is your greatest fear?
• A) Failure
• B) Betrayal
• C) Ignorance
• D) Weakness
6. Which of these subjects would you excel in?
• A) Defense Against the Dark Arts
• B) Herbology
• C) Charms
• D) Potions
7. How do you handle conflict?
• A) Face it head-on
• B) Seek a peaceful resolution
• C) Analyze the situation
• D) Use it to your advantage
8. What is your ideal pet?
• A) Lion
• B) Badger
• C) Eagle
• D) Snake
9. Which of these quotes resonates with you the most?
• A) “Courage is not the absence of fear.”
• B) “Loyalty is the strongest bond.”
• C) “Knowledge is power.”
• D) “Ambition drives success.”
10. What is your approach to teamwork?
• A) Lead the team
• B) Support the team
• C) Strategize for the team
• D) Ensure the team wins

These questions can help identify traits associated with Gryffindor, Hufflepuff, Ravenclaw, and Slytherin. Have fun sorting!

****

Sent by you: generate crypto-related questions sorting Hogwarts houses

Sent by Copilot: Sure! Here are some crypto-related multiple-choice questions to help sort someone into a Hogwarts house:

1. Which aspect of cryptocurrency excites you the most?
• A) The potential for revolutionary change
• B) The sense of community and collaboration
• C) The innovative technology behind it
• D) The opportunities for financial gain
2. How would you handle a sudden drop in the value of your crypto investments?
• A) Stay calm and hold on
• B) Discuss strategies with fellow investors
• C) Analyze the market trends and data
• D) Look for ways to capitalize on the situation
3. What type of blockchain project would you most likely support?
• A) One that aims to disrupt traditional systems
• B) One that focuses on social good and inclusivity
• C) One that introduces groundbreaking technology
• D) One that promises high returns on investment
4. How do you prefer to learn about new cryptocurrencies?
• A) Through hands-on experience and experimentation
• B) By joining online forums and communities
• C) By reading whitepapers and technical documents
• D) By following market trends and expert advice
5. What is your approach to securing your crypto assets?
• A) Use a combination of methods to ensure safety
• B) Share tips and best practices with others
• C) Stay updated with the latest security technologies
• D) Invest in the most secure and profitable options
6. Which of these crypto-related activities appeals to you the most?
• A) Mining or staking
• B) Participating in airdrops and giveaways
• C) Developing or coding blockchain solutions
• D) Trading and investing
7. How do you view the future of cryptocurrency?
• A) As a tool for global change and empowerment
• B) As a way to build stronger communities
• C) As a field ripe for innovation and discovery
• D) As a lucrative investment opportunity
8. What would you do if you discovered a new, promising cryptocurrency?
• A) Invest immediately and take the risk
• B) Share the information with your network
• C) Research thoroughly before making a decision
• D) Evaluate its potential for high returns
9. Which of these statements resonates with you the most?
• A) “Fortune favors the bold.”
• B) “Together, we are stronger.”
• C) “Knowledge is the key to success.”
• D) “Ambition knows no bounds.”
10. What is your strategy for staying informed about the crypto market?
• A) Follow the latest news and trends
• B) Engage with the crypto community
• C) Study market analysis and reports
• D) Monitor investment opportunities closely

These questions can help identify traits associated with Gryffindor, Hufflepuff, Ravenclaw, and Slytherin in the context of cryptocurrency. Enjoy sorting!

****

Sent by you: Generate law school-related questions sorting Hogwarts houses

Sent by Copilot: Sure! Here are some law school-related multiple-choice questions to help sort someone into a Hogwarts house:

1. Which area of law interests you the most?
• A) Criminal Law
• B) Family Law
• C) Intellectual Property Law
• D) Corporate Law
2. How do you handle a challenging case?
• A) Tackle it head-on with determination
• B) Collaborate with colleagues for the best outcome
• C) Research extensively to find the best solution
• D) Strategize to turn the case in your favor
3. What motivates you to become a lawyer?
• A) Fighting for justice
• B) Helping those in need
• C) The intellectual challenge
• D) Achieving success and recognition
4. How do you prepare for a big exam?
• A) Study diligently and practice under pressure
• B) Form study groups and support each other
• C) Create detailed outlines and review notes
• D) Focus on key areas to maximize your score
5. What is your approach to legal ethics?
• A) Uphold justice at all costs
• B) Ensure fairness and compassion
• C) Follow the rules and guidelines meticulously
• D) Navigate the system to benefit your client
6. Which legal figure do you admire the most?
• A) Ruth Bader Ginsburg
• B) Thurgood Marshall
• C) Oliver Wendell Holmes Jr.
• D) Clarence Darrow
7. How do you handle courtroom pressure?
• A) Stay calm and assertive
• B) Rely on teamwork and support
• C) Use logic and evidence to persuade
• D) Use your wit and charm to influence
8. What is your ideal legal career path?
• A) Public defender or prosecutor
• B) Legal aid or non-profit work
• C) Academic or legal scholar
• D) Corporate lawyer or partner at a firm
9. Which of these quotes resonates with you the most?
• A) “Justice delayed is justice denied.”
• B) “Injustice anywhere is a threat to justice everywhere.”
• C) “The law is reason, free from passion.”
• D) “The ends justify the means.”
10. How do you approach legal writing?
• A) Write with passion and conviction
• B) Ensure clarity and empathy
• C) Focus on precision and detail
• D) Make persuasive and compelling arguments

These questions can help identify traits associated with Gryffindor, Hufflepuff, Ravenclaw, and Slytherin in the context of law school. Good luck with your sorting!