In Re: OpenAI, Inc. Copyright Infringement Litigation is another proposed class action by authors involving alleged infringement by large language models (LLMs). Here, the class would represent authors and copyright holders of fiction and non-fiction books who assert claims on their own and on behalf of the proposed class.
Plaintiffs alleged OpenAI infringed their copyrights by downloading and reproducing plaintiffs’ works and then using the reproduced work to train OpenAI’s LLMs. OpenAI brought a motion to dismiss, claiming that the plaintiffs did not plausibly allege substantial similarity between ChatGPT’s output and the plaintiffs’ works.
The underlying class action complaint included allegations that, when prompted, ChatGPT accurately generated summaries of several of George R.R. Martin’s works, including those associated with A Game of Thrones. They allege that ChatGPT also generated detailed outlines for a prequel and alternate sequel using some of the same characters and imitating the fantasy names and aesthetic that made A Game of Thrones so popular.
In considering the motion, the SDNY noted that when a work has both protectable and non-protectable elements, the “more discerning observer” test is used to gauge substantial similarity between the elements that are worthy of copyright protection and those that are not. This is a more rigorous test than the “ordinary observer” test, which is used when there are no non-copyrightable elements mixed in with the copyrightable elements.
After excluding things like tropes and scènes à faire, the court paired down the copyrightable elements and concluded that a more discerning observer could easily perceive that the detailed summaries provided by ChatGPT did in fact infringe on Martin’s original works. Therefore a reasonable jury, the court said, could easily determine that the output was substantially similar to Martin’s original work. OpenAI’s motion to dismiss the substantial similarity claims was denied.
The motion to dismiss also challenged the fact that plaintiffs had failed to attach any examples of the infringement to their brief. The court threw that out, reminding OpenAI that, at the pleading stage, a court may consider materials outside of but sufficiently referenced in the pleadings.
Plaintiffs had attempted to argue that OpenAI’s motion to dismiss was barred by a stipulation between the parties in another lawsuit, Authors Guild v. OpenAI. There, the parties had agreed not to move to dismiss the then-currently pleaded claims. This court disagreed, saying that lawsuit was irrelevant to the proposed class action at hand because these pleadings consolidated several cases and were different from the complaint in the earlier case.
Typically analyses of substantial similarity involve side-by-side comparisons of the original and allegedly infringing works. However, the court’s declining to cite to a side-by-side comparison of the works was a testament to the original work in question (as the Game of Thrones books were the basis for the popular TV series); rather, the court simply recounted the infringing output, evidently expecting the Zeitgeist to take care of the rest.
Though the lead plaintiff in this case is “unimaginably successful” (in the author’s own words), creators of less popular works will theoretically benefit if the plaintiffs prevail in this case. I wonder, though, whether even another case’s proposed class action settlement for a three thousand dollars for each work is significant for individual authors, who typically share large portions of all income with their publishers and book agents. If AI is poised to replace works of a lifetime with a fresh version in minutes, compensation for the authors may become very, very limited.
