Things just went from bad to worse for OpenAI.
Judge Stein issued a second opinion related to OpenAI’s and Microsoft’s motions to strike allegedly new allegations made in the Consolidated Class Action Complaint by book authors. (Judge Stein’s first opinion denied OpenAI’s motion to dismiss the “output” copyright infringement claim.)
OpenAI won a modest victory: Judge Stein did strike new allegations related to new models of OpenAI: “GPT-4V, GPT-4.5, GPT-5, and “derivatives” and “successors” of any model.” These were not a part of the prior complaints.
But everything else was a loss for OpenAI: Judge Stein allowed the Class Plaintiffs’ theory of infringement based on downloading copies (from so-called shadow libraries) as a separate claim of infringement apart from training. Such downloading was already alleged in the prior complaint but it was framed in terms of training. In the Consolidated Class Action, the downloading became a distinct theory of infringement. Judge Stein held: “the Consolidated Class Action Complaint’s claim of infringement based on OpenAI’s alleged download and reproduction of plaintiffs’ books is not a new claim in violation of the Court’s May 22 order.”
Judge Stein also rejected Microsoft’s motion to strike allegations related to GPT-4o and GPT-4o Mini. OpenAI conceded they were part of the original action. And “Microsoft slices the baloney too thin.”
Is the MDL Litigation v. OpenAI now Bartz II or Bartz II-plus?
The way things are playing out, it looks like In re OpenAI Copyright Infringement Litigation is potentially turning into Bartz II or even Bartz II-plus.
The Class Plaintiffs (represented by the same lead counsel Justin Nelson of Susman Godfrey) now have at least 3 different theories of infringement v. OpenAI:
- Downloading from Shadow Libraries claim [Judge Stein just allowed]
- Training model claim
- Output claim [Judge Stein just allowed]
Bartz involved only 1 and 2.
So, OpenAI has its work cut out for it. Even if OpenAI wins its defense for fair use for AI training, it still must win on the Shadow Library claim by convincing Judge Stein that the correct approach is to treat the initial download as part of the purpose of training (as Judge Chhabria ruled in Kadrey v. Meta). But, if Judge Stein follows Judge Alsup’s approach in Bartz v. Anthropic, the case against OpenAI may turn into Bartz II.
For the purposes of statutory damages, the Class Plaintiffs get recovery only for each work infringed, no matter how many times infringed. So the Plaintiffs don’t even need to win on the AI training issue to maximize their potential statutory damages. Indeed, the Plaintiffs winning on training wouldn’t even add any additional statutory damages to any recovery they may get if they prevail on the Shadow Library claim.
Based on Judge Stein’s opinion yesterday, the output claim doesn’t sound very strong. Even if meritorious, it would also add nothing more in terms of statutory damages. In order for the model to output an infringing copy, the copy must have been included in the training datasets, which presumably were compiled from the alleged use of Shadow Libraries. But the output claim does provide an alternative path to proving infringement. Ergo Bartz II-plus.