Sarah Silverman and her fellow author plaintiffs are fighting a judge’s recent order requiring them to disclose the prompts and outputs they used in preparation for filing their class action lawsuit against ChatGPT owner OpenAI. The judge is giving OpenAI until July 24 to respond to the plaintiffs’ argument that the material should be shielded as work product.

The magistrate judge’s ruling is the latest in a series of setbacks suffered by the plaintiffs in Tremblay et al. v. OpenAI et al., who filed suit in the U.S. District Court for the Northern District of California just over a year ago against several entities associated with ChatGPT. As we have previously reported, the plaintiffs allege that OpenAI infringed their rights in copyright-protected content by using it as training material for ChatGPT. The court granted OpenAI’s motion to dismiss many of the plaintiffs’ claims, but gave the plaintiffs an opportunity to replead. The plaintiffs did so in March, and OpenAI has moved to dismiss that amended pleading.

The parties participated in a settlement conference in June, but did not settle. The latest motion to dismiss is still pending. Even so, discovery is proceeding in parallel.

OpenAI issued a request seeking access to all non-privileged documents and communications related to the plaintiffs’ pre-suit investigation, including OpenAI account information as well as prompts and outputs from pre-suit testing of ChatGPT. Not surprisingly, the plaintiffs objected on the basis that the material constitutes privileged work product.

The magistrate judge presiding over the case agreed that the material was indeed work product—but also found that the plaintiffs had waived protection by including certain of the test result outputs in their complaint. Specifically, plaintiffs attached summaries of their copyright-protected works to the pleading in order to illustrate their allegation that OpenAI must have used the original works as training material for the platform. The exhibits also includes specific examples of prompts used to generate those summaries.

According to OpenAI, the prompts included in the complaint were “preferred, cherry-picked results,” and OpenAI should be able to access the full threads of prompts and outputs that led to those examples. The court agreed, concluding that the plaintiffs had placed ChatGPT prompts and outputs at issue, and thus had waived work product protection over them. It rejected the plaintiffs’ argument that revealing all of the information requested, including account settings and negative test results (i.e., those unrelated to the examples highlighted in the complaint), would reveal their attorneys’ thought processes and interpretation of legal theories. The court held that “an evaluation of those account settings and negative test results is a necessary component of Defendants’ ability to understand Plaintiffs’ positive results and to fairly subject those results to meaningful scrutiny.”

Plaintiffs have now filed a motion in the district court seeking relief from the magistrate’s decision, and the court is allowing OpenAI to respond. Meanwhile, the court has set another settlement conference for September 13, 2024.