ECPA Members Should Act Soon if Interested in Class Settlement
By Craig Gipson
After waiting years for resolution to the fair use questions surrounding artificial intelligence, some answers arrived quickly this summer. Federal courts issued two highly-anticipated decisions, including one in a class action suit that could involve ECPA members. Due to the pressing timeline of one case, this Legal Update begins with action your organization should consider within the next several days and follows with a discussion of the courts’ rulings.
Immediate (Class) Action
To start at the story’s end, the most pressing AI litigation issue for ECPA members is submitting information to potentially join the plaintiff class in Bartz v. Anthropic. After Judge Alsup certified the class in July, many commentators were surprised on August 27 when the parties informed the court that they had reached a tentative settlement. What many expected to be several more months of settlement negotiations and drawn-out litigation immediately shifted to an expedited schedule. The parties indicated they expect to finalize the settlement before September 5, and a hearing is now scheduled for September 8.
Once the parties finalize settlement terms and the court approves, procedural rules require attorneys for the class to undertake certain efforts in contacting class members. To facilitate these communications, one of the participating law firms established a Publishers’ Contact Information Intake Page to receive information about potentially joining the class settlement. Interested publishers may also reach out to the firm of Lieff Cabraser Heimann & Bernstein with questions at anthropiclawsuit@lchb.com or 1-800-254-2660.
If your organization believes its titles were affected (see more below), initiating communications in early September would be wise to ensure time for consideration before potential deadlines later in the fall.
Despite optimism about the settlement, court approval is not merely a rubber stamp in these cases. Courts must consider the fairness of the settlement to the entire class as well as antitrust and other legal concerns. Publishers may recall the court in Google Books rejected the parties’ first attempt at settlement in 2011 for this reason.
By submitting information to class counsel, you are not forfeiting any right to opt-out of the class. If eligible for class membership, your organization may hear from this law firm in the fall even without submitting anything now, but your contact information and documentation about potentially affected titles may expedite communications and provide more time for your organization to gather information and make decisions.
-
Why Are the Parties Settling?
Litigation is expensive and after Judge Alsup’s initial ruling (discussed below), both parties have something to lose. For Anthropic, facing 7 million instances of infringement at a statutory damages rate of $750 to $30,000 per instance, $150,000 if found to be willful, could be extremely damaging. For rightsholders, Anthropic could fight this litigation for years on appeal, chipping away at the number of titles or eligibility of class members. A sizeable settlement now that eliminates that downside risk may look very attractive.
Author advocate Dave Hansen pointed out that, cynically, it may benefit individual AI companies to reach class action settlements. They leave open the core copyright infringement question, which could still hamper competitors, while gaining access to a large corpus of works for training. Of course, settlement could also invite other lawsuits from third parties outside the settling class. Training datasets encompass so many works that it would be nearly impossible for a class to include every relevant rightsholder.
-
Who Can Be a Class Member?
An author or beneficial owner whose works were among those in the Pirate Library Mirror or Library Genesis shadow libraries downloaded by Anthropic may be in the class. We should learn more about exactly which titles Anthropic downloaded soon. If you suspect your organization’s titles are among those affected, include that title in the submission form to class counsel.
-
What Does Being a Class Member Require and Mean?
We should learn more once class counsel communicates with potential class members more broadly, but it may require evidence of rights to an affected title. If your organization’s titles are among those relevant to the case, you may need to demonstrate rights acquired via an author publishing agreement or similar grant of rights.
The settlement applies only to those who are ultimately found to be eligible and who do not affirmatively opt-out. A class action settlement binds class members but does not require that all potential class members remain in the class. Any affected party can affirmatively opt-out of the class and assert its own rights in a separate lawsuit. However, joining the class may be the simplest and least administratively burdensome way to receive compensation for Anthropic’s use of your organization’s published works. If you receive communication from class counsel, it is important to pay attention to deadlines and required replies so as not to be bound by the settlement unintentionally.
-
How Much Will Class Members Receive As Part of the Settlement?
Until the settlement is publicly announced, we will not know what the total amount is to be divided among class members. Nor will we know the number of participating class members.
What Did the Courts Decide About AI Training and Fair Use: The Bartz and Kadrey Decisions
Bartz and Kadrey share many similarities in being closely-watched opinions on the question of copyright infringement and AI training. But the distinct facts and different legal theories presented in each case prevent a true apples-to-apples comparison. Appeals and potential settlements will impact where the law ultimately lands, but on the primary question at issue, courts appear favorable toward the argument that training AI models on copyrighted works constitutes fair use. Not all is lost, however. One court expressed greater skepticism on the legality of training AI models on pirated copies and creating a perpetual, electronic library of pirated downloads.
As in all fair use analysis, the courts evaluated the four factors contained in the Copyright Act. And as in almost all fair use cases, the first factor—the nature and character of the use—and the fourth factor—the effect on the market for the original work—were determinative.
-
Bartz v. Anthropic
Anthropic became well-known for developing one of the leading frontier AI models, Claude. Already a high-interest AI-copyright case, Bartz gained even more attention from legal observers in July when Judge Alsup agreed to certify a class of rights owners affected by Anthropic’s alleged infringing actions. The prospect of a class action suit with potential damages in the billions raised eyebrows across the publishing and technology sectors. These high stakes likely pushed each side toward the pending settlement discussed above.
Each party asked the judge to grant summary judgment, essentially arguing that with the evidence already presented, copyright law weighed so strongly in their favor that the judge could issue a final opinion. In response, Judge Alsup gave each party a potential win and a potential loss: he unequivocally found that training AI models on copyrighted works is a transformative use, supporting a finding of fair use. But he also displayed skepticism toward Anthropic (i) downloading seven million pirated copies for the purpose of training, and (ii) maintaining those copies in a central electronic library for any future use, even after deciding some would not be included in a training dataset.
In an atypical approach, Judge Alsup broke his first factor analysis into two distinct categories: (1) the type of use—training versus indefinite storage and undefined use, and (2) the method of acquisition—purchase and electronic scanning versus downloading pirated copies. For the type of use, Judge Alsup wrote strongly in favor of Anthropic on the first factor, finding use of legitimately acquired copies for training to be highly transformative. But he also found that training was the second step in Anthropic’s process. First, it built a general-purpose electronic library with pirated copies, copies which could easily have been lawfully acquired. This distinguished Anthropic’s actions from the 2015 Google Books case when the tech giant scanned millions of print copies without permission. Publishers and authors objected to Google’s large-scale reproduction effort but no one questioned whether the works were legally obtained. Google purchased or worked with libraries to build out its dataset. Libraries lawfully assembled their collections for the purpose of scholarship, and digitized copies for accessibility and to guard against loss or damage. Conversely, Anthropic maintained permanent copies of pirated books, even without a discernible use, and allowed broad access to its library of works.
On the method of acquisition, the court questioned whether Anthropic could ever meet its burden of showing why downloading pirated copies was reasonably necessary to facilitate a subsequent fair use. The court pointed out that other accepted fair uses do not excuse an infringing origin. Writing a book review, researching the facts of a book, or creating an AI model do not require downloading an illegal copy. Weighing the transformative nature of the use with the questionable means of obtaining copies, the court split on the first fair use factor. For lawfully acquired titles used to train AI models, the court leaned toward fair use. For illegally acquired titles used for training or indefinite storage in a “central library,” the court weighed in favor of infringement.
On the fourth factor (market harm), the court equivocated on the issue of AI training. Finding that training AI models did not serve as a direct substitute for the original work, it acknowledged the developing market for AI licensing. However, it pointed out that this was not a market to which authors are entitled under the Copyright Act.
But for the category of storage in an electronic central library, Judge Alsup was more definitive: Anthropic has no rebuttal to the charge that it stole works to build its collection. Anthropic attempted to argue that it could have purchased used copies of the works, which would not have resulted in payment to the authors and therefore would have achieved the same result. The court acknowledged this possibility and recommended that perhaps Anthropic should have done so rather than download pirated copies for free.
Judge Alsup granting each side a win and leaving open the possibility of a major loss likely incentivized the settlement discussions progressing to their current state.
-
Kadrey v. Meta
Kadrey is not a class action case but involves the direct claims of 13 authors against Meta and its Llama AI models. Overall, Kadrey is a less friendly opinion to rights holders than Bartz.
In Kadrey, as in Bartz, the court found the training of AI models on copyrighted works to be transformative under the first fair use factor. Judge Chhabria found that the purposes of an AI model are wholly distinct from those of the allegedly infringed works. He cited AI purposes such as using a chat bot to find recipes, get tax or medical advice, translate documents, or conduct research as distinct from those of a single book. The court paid lip service to the commercial nature of Meta’s venture but found that the “highly transformative” nature of the use ultimately weighed in favor of fair use.
While the court took pains to check the necessary jurisprudence boxes—claiming that the transformative nature was not necessarily determinative—it treated the problems of commerciality and bad faith (i.e. using pirated copies) with kid gloves. As for commercialism, the court acknowledged the billions of dollars at stake for the AI industry, but alluding to its later discussion of the final fair use factor, found that the commercial nature of Meta’s use is less important because there is no resulting market harm to the copyrighted works.
Judge Chhabria then funneled the bad faith act of shadow library downloads through a kind of rights laundering filter: that the focus of inquiry should be on the resulting creation benefiting the public, and thus protected by copyright law, rather than on the “morality of the secondary user.” Departing from Judge Alsup in Bartz, Judge Chhadria found that the infringing party acting in bad faith “shouldn’t be especially relevant” if the new expression does not substitute for the original work. That blatant, unquestioned infringement can somehow be cured by a subsequent fair use is a bitter pill for rights holders to swallow. The court expressing “no market harm” when the bad actor avoided purchasing millions of copies is difficult to reconcile. In future cases, rights holders will likely press for the adoption of the Bartz view.
-
Market Harm Problems
The courts’ conclusions on the transformative nature of AI and authors and publishers’ exclusive right of reproduction failed to surprise many legal commentators. But perhaps more concerning was the judges’ dismissive language under the fourth factor toward the effect of AI on relevant markets.
Generally, the judges considered three types of harm: (1) direct substitution by the models creating verbatim or substantially similar works; (2) licensing market harm by destroying the market for licensing works for AI training; and (3) indirect substitution by AI creating so many works that it dilutes the book publishing market altogether.
While publishers express concern with all three, the judges disagreed as to which posed the greatest problem and evidence demonstrating damages. In Bartz, Judge Alsup dismissed concerns about indirect substitution, likening it to schools teaching more children to write stories. While Judge Chhadria in Kadrey thought the risk of displacement by indirect substitution is very real.
Discouragingly, both cases dismissed the second market harm of licensing works for AI purposes. The Kadrey court argued that every favorable fair use decision destroys a potential licensing market for rightsholders; that is the very nature of fair use. However, its concern with the circular nature of this argument ignored that this market is not theoretical but is already very much established. Evidence of numerous licenses from large media organizations to AI companies is readily available. While Bartz highlighted that this is not a market the Copyright Act grants rightsholders, it is one which the relevant parties themselves created. For the law to interject and destroy a market already worth tens or hundreds of millions of dollars to rights holders seems antithetical to the purposes of copyright.
Conclusion
Settlement and appeals notwithstanding, training AI models on copyrighted works likely falls within the fair use exception now. Each case is fact specific so rightsholders may still emerge victorious in certain fights, but those are likely the exceptions. But in the doom and gloom, at least the added insult of training on illegal copies may be prohibited. If Bartz settles—and every indication is that it will—there will be no final opinion on the pirated works question. But it would require a brazen tech company to attempt it unless it was courting another round of copyright litigation.
For now, publishers should move quickly to consider the Bartz class settlement by visiting the Publishers Contact Information Intake Page. You may also reach out to your organization’s counsel or Flagler Law Group with questions. Registering your organization’s interest in early September will help begin the process of communication and determining class membership.
Publishers should also continue to evaluate AI licensing opportunities. Just because training AI models is fair use, does not mean that all AI-related uses are fair. If Bartz’s finding regarding its compilation of a central library stands, AI companies may still approach rightsholders to license content for retrieval-augmented generation and reference purposes.
