Greetings CIPAWorld!

Here are some exciting case updates involving Google. What started as a headline-making copyright case against Google just became required reading for anyone litigating under CIPA. So, you may be asking, what do copyright and CIPA have in common? Don’t worry… the connection will become clear as we explore these cases. In In re Google Generative AI Copyright Litig., No. 23-cv-03440-EKL, 2025 U.S. Dist. LEXIS 75740 (N.D. Cal. Apr. 21, 2025), a class of creators claimed that Google scraped their copyrighted works without permission to train its AI models. It was pitched as a massive data appropriation lawsuit. Still, the case stumbled temporarily because, as litigators know all too well, the Plaintiffs proposed an improperly defined class.

The Plaintiffs, a group of authors, illustrators, and content creators, accused Google of using their copyrighted materials to train its generative AI models without permission. I find this fascinating! While in law school, I wrote a white paper on this topic, examining the copyright implications of using creative works to train AI systems. The intersection of copyright law and emerging technologies presents novel legal challenges.

In this case, it was a sweeping theory of unauthorized data use, but the case ran into trouble the moment plaintiffs defined their class. They limited membership to individuals “whose exclusive rights under 17 U.S.C. § 106 in their registered works were infringed upon.” In re Google Generative AI Copyright Litig., 2025 U.S. Dist. LEXIS 75740, at *6. In other words, you were only in the class if Google violated your copyright.

It may seem straightforward to target affected individuals, but the Court immediately identified the problem. The class only included those who would ultimately prevail on the merits. As such, the Court couldn’t determine who was in the class without deciding if Google was liable to each potential class member. News flash… that’s what courts call a “fail-safe” class.

Judge Lee explained that “the Court cannot determine who is a member of the class without deciding the merits of each potential class member’s claim, including whether the potential class member has a valid copyright registration, whether Google infringed the class member’s work(s), and whether Google has a valid defense based on fair use or license.” Id. at *10.

As the Ninth Circuit explained in Kamar v. Radio Shack Corp., 375 F. App’x 734, 736 (9th Cir. 2010), a fail-safe class is impermissible because membership is conditioned on a legal finding. It’s circular. Because Plaintiffs’ proposed class was tied to the elements of infringement, the Court struck the class allegations under Fed. R. Civ. P. 12(f). See Google Generative AI Copyright Litig., 2025 U.S. Dist. LEXIS 75740, at *11. Judge Lee didn’t dismiss the case outright, but she gave Plaintiffs fourteen days to amend their definition.

The Court also offered a suggestion: reframe the class based on factual criteria. The revised definition proposed by Plaintiffs, “all persons or entities domiciled in the United States who owned a United States copyright in any work used by Google to train Google’s Generative AI Models during the Class Period,” was precisely that. Id. Judge Lee acknowledged that this revised definition “would not require an upfront determination by the Court that each potential class member will prevail on the merits of an infringement claim.” Id. This makes perfect sense. That’s the difference between a procedural dead-end and a viable class.

This issue isn’t unique to copyright litigation. I mean, this is CIPAWorld, right!? Plaintiffs continue to define classes as people “whose communications were intercepted” or “whose data was unlawfully shared.” These definitions don’t identify a group of people based on facts. They identify a group based on whether they’ve already proven their claim. That’s precisely what courts are rejecting.

I saw a nearly identical issue in In re Google RTB Consumer Priv. Litig., No.: 4:21-cv-2155-YGR, 2024 U.S. Dist. LEXIS 119157 (N.D. Cal. Apr. 4, 2024) a few weeks prior. That case focused on Google’s Real-Time Bidding (“RTB”) platform. Plaintiffs alleged that the system shared sensitive user data with advertisers through real-time ad auctions. The class was defined as Google account holders “whose personal information was sold or shared.” Sounds familiar, right?

Judge Yvonne Gonzalez Rogers found the class definition flawed. Like Judge Lee, she concluded that the definition was “fail safe” because it required resolving the merits. Specifically, whether Google “impermissibly shared” personal information, just to identify who belonged in the class. The Court stated: “The Court agrees with Google that, as written, the class definition is fail safe. The question on which this suit hinges is whether Google impermissibly shared its account holders’ personal information through RTB.” Id. at *18.

But that wasn’t the only issue the court addressed. Judge Rogers also cautioned that removing the contested phrases from the class definition might broaden the class so much that it would include users who weren’t harmed. The Court stated, “Defining a class so as to avoid, on one hand, being over-inclusive and, on the other hand, the fail-safe problem is more of an art than a science.” Id. at *17.

CIPA litigators should take note. These rulings aren’t just about definitions, but they’re about strategy. If the class can’t be defined in a way tethered to objective facts, plaintiffs won’t make it to the merits. Courts aren’t guessing anymore. They’re asking: Can we identify class members without deciding if the law was broken? If the answer is no, certification won’t happen.

The RTB case also surfaced another common problem in CIPA litigation: individualized consent. Judge Rogers denied certification of a Rule 23(b)(3) damages class because determining who saw disclosures and who didn’t would require user-by-user analysis. That inquiry would overwhelm common issues.

Still, the Court acknowledged that a Rule 23(b)(2) injunctive class could be appropriate. While Plaintiffs could not satisfy the predominance requirement for a damages class, the Court noted that prospective injunctive relief might proceed under a different analysis. A forward-looking injunction targets company practices going forward and doesn’t require resolving individualized consent issues for each user. But even injunctive claims must be grounded in a well-defined, objectively ascertainable class.

Despite presenting expert evidence involving millions of RTB bid requests, Plaintiffs faced one more obstacle. The Court was not persuaded that the data set reliably reflected the experience of the proposed class as a whole. Plaintiffs alleged that advertisers could determine what content users viewed and even infer their locations. But the Court held that this wasn’t enough. The data needed to be representative of the entire class experience, and the plaintiffs hadn’t met that burden. See In re Google RTB Consumer Priv. Litig., 2024 U.S. Dist. LEXIS 119157, at *32-33.

For defense counsel, the takeaway is that challenges to class definitions and evidentiary gaps remain powerful early tools to be utilized. Whether the issue is consent variability, class overbreadth, or sampling deficiencies, these rulings reinforce that procedural missteps can and often derail class actions before the merits stage. As CIPA litigation continues to sweep across California, these two Google rulings illustrate where cases are getting stuck. Defining your class around legal conclusions, relying on non-representative data, or ignoring consent variations are no longer technical errors. They are strategic liabilities.

Whether you’re responding to claims involving chat features, embedded scripts, or real-time data flows, the foundational question remains the same: who’s in your class, and how do you know? If answering that requires proving liability, the case may never reach certification.

As always,

Keep it legal, keep it smart, and stay ahead of the game.

Talk soon!