The class of plaintiff authors seeking to hold OpenAI liable for copyright infringement has faced yet another setback. The U.S. District Court for the Northern District of California has knocked out the majority of their claims, refusing to accept the blanket allegation that “every output of the OpenAI Language Model is an infringing derivative work.” However, the court has allowed the plaintiffs another chance to cure many of the deficiencies in their pleadings, so the battle is not yet over.

As we’ve previously reported, named plaintiffs including Paul Tremblay, Sarah Silverman, and Michael Chabon have filed class action lawsuits against several companies associated with popular Large Language Model tools like ChatGPT. The lawsuits claim that because the defendants copied their original works of authorship to use as training material for the LLMs, the AI companies are liable under the federal Copyright Act and various state tort laws. For a quick recap of the theories they are asserting, check out our recent AI Update.

This week’s ruling in the Tremblay/Silverman/Chabon cases—which the court intends to consolidate into a single case moving forward—echoes a similar ruling in Anderson v. Stability AI, which involved analogous claims related to visual art instead of written works. As we reported in November, the judge in that case found the complaint “defective in numerous respects,” but gave the plaintiffs a chance to replead. The plaintiffs have filed an updated complaint in that case, but we have not yet seen whether any of the amended claims will be upheld.

The Tremblay plaintiffs will also be forced to replead if they want to continue to pursue the bulk of their claims. Indeed, the court nearly granted OpenAI’s motion to dismiss five of plaintiffs’ six claims in its entirety, leaving a portion of Count IV intact and dismissing the rest with leave to amend.

Here’s a quick rundown on the court’s decision as to each of these claims:

  • Direct copyright infringement (Count I) – SURVIVES
    • OpenAI did not move to dismiss on this claim (we suspect this is because they plan to assert a fair use defense at summary judgment).
  • Vicarious copyright infringement (Count II) – DISMISSED
    • OpenAI argued that this claim fails because plaintiffs did not make a threshold showing of direct infringement.
    • The court agreed, finding that in order to make out a direct infringement claim, plaintiffs would have to show that the output from the LLMs is “substantially similar” to the original works that were allegedly input as training material, not just that the works were copied initially at the input stage.
  • Violation of DMCA Section 1201(b) (Count III) – DISMISSED
    • This claim alleges OpenAI violated the Digital Millenium Copyright Act (“DMCA”) by removing identifying information such as title, author, and copyright notice information for the original work.
    • In its order, the court points to the fact that excerpts of the copyrighted works cited in plaintiffs’ own complaint include “multiple references to Plaintiffs’ names, suggesting that OpenAI did not remove all references to ‘the name of the author’” as alleged.
    • The court ultimately dismissed the claim to the extent it is based on knowing and intentional removal of copyright information, and also to the extent it is based on distribution of derivative works without the copyright information included. The court stated that it is not sufficient to allege merely that “every output from the OpenAI Language Models is an infringing derivative work.”
  • Unfair competition (Count IV) – DISMISSED IN PART
    • The court held there is no “unlawful” activity to support this claim because it had dismissed the predicate DMCA claims, and any resulting injury would be purely speculative.
    • The court held there is no “fraudulent” activity to support this claim due to a lack of specific fraud allegations.
    • However, the court upheld the portion of the claim based on “unfair” activity, concluding that when assuming the truth of plaintiffs’ allegations, use of copyrighted works to train LLMs for commercial profit could be considered an unfair practice. (Of course, the court observed in a footnote, this claim is based on activity prohibited by the Copyright Act, and thus may ultimately be preempted on that basis).
  • Negligence (Count V) – DISMISSED
    • The court found that plaintiffs had not identified any basis for OpenAI’s legal duty to “maintain and control” information contained in the copyrighted books, and denied the claim on that basis.
  • Unjust enrichment (Count VI) – DISMISSED
    • This claim was dismissed because plaintiffs failed to allege that OpenAI “unjustly obtained benefits from Plaintiffs’ copyrighted works through fraud, mistake, coercion, or request.”

The plaintiffs now have until March 13, 2024 to file an amended complaint. It will be interesting to see how they attempt to reframe these claims in light of the court’s harsh ruling.

Other parties may be watching closely as well: Tremblay and the other California plaintiffs have moved to stay cases filed in New York by George R.R. Martin and others against OpenAI until the California suits are resolved, or to have those New York cases transferred to California so they can all be resolved together.