Tremblay v. OpenAI, Inc.

Background

Plaintiffs, described in the opinion as a group of authors, alleged that Defendants’ ChatGPT software relied on a large language model trained by copying and extracting information from large amounts of text. They alleged that ChatGPT could generate summaries of their copyrighted works and could produce text in the style of one of the plaintiffs or one of their works. The complaint included examples of prompts and ChatGPT responses in Exhibit B.

The discovery dispute concerned Defendants’ Request for Production No. 9, which sought nonprivileged documents and communications related to Plaintiffs’ investigation of the claims. Defendants narrowed the request to seek: (1) OpenAI account information for individuals who used ChatGPT to investigate Plaintiffs’ claims; (2) prompts and outputs from Plaintiffs’ pre-suit testing, including prompts and outputs that did not reproduce or summarize Plaintiffs’ works or otherwise support their claims; and (3) documentation of Plaintiffs’ testing process.

Defendants argued that Plaintiffs had placed their testing at issue by describing selected prompts and results in the complaint and Exhibit B. They also argued that the account information and complete testing data were needed to test the reported results fairly, including because account settings and prompt wording could affect ChatGPT’s responses. Plaintiffs argued that the undisclosed prompts and outputs were protected opinion work product because they reflected their lawyers’ thoughts and analysis. Plaintiffs said they had agreed to provide prompts and outputs supporting their claims but argued that negative or otherwise unsupportive results were not relevant and that Defendants could conduct their own ChatGPT testing.

Court’s analysis

The court explained that work product generally protects documents and tangible things prepared in anticipation of litigation or for trial. Fact work product may be discovered when the requesting party shows a substantial need for the material and cannot obtain its substantial equivalent without undue hardship. Opinion work product protects an attorney’s mental impressions, conclusions, opinions, or legal theories, although the protection is qualified and can be waived.

The court found that the account settings and negative test results were more like fact work product than opinion work product. It rejected Plaintiffs’ argument that disclosure of those facts would necessarily reveal their lawyers’ analysis. The court also ruled that, because Plaintiffs had included a large subset of the testing facts in the complaint and Exhibit B, Plaintiffs had waived work-product protection over the account settings and negative results that were reasonably related to the positive results they had placed at issue.

The court concluded that reviewing the negative test results and the account settings was necessary for Defendants to understand and meaningfully scrutinize the positive results. The court rejected Plaintiffs’ contention that the negative results were irrelevant merely because they did not support Plaintiffs’ claims, explaining that the information was relevant to Defenses. The court also rejected the argument that Defendants could simply conduct their own testing, because without Plaintiffs’ account settings and exact prompts, Defendants could not replicate the same results or effectively evaluate Plaintiffs’ testing.

Disposition

The court granted Defendants’ request to compel the OpenAI account information for the individuals who used ChatGPT to investigate Plaintiffs’ claims; the prompts and outputs from Plaintiffs’ pre-suit ChatGPT testing, including those that did not reproduce or summarize Plaintiffs’ works or otherwise support Plaintiffs’ claims; and documentation of Plaintiffs’ testing process. The opinion resolved this discovery dispute and did not decide the underlying claims.