Tremblay v. OpenAI, Inc.

Background

Paul Tremblay and other plaintiffs sued OpenAI, Inc. and other defendants, asserting direct copyright infringement on behalf of themselves and a proposed class. They alleged that ChatGPT was powered by large language models trained by copying large amounts of text and extracting expressive information from it. They alleged that their copyrighted books were used in that process.

The plaintiffs’ complaint included prompts used to ask ChatGPT to summarize parts of their books and the resulting ChatGPT outputs. OpenAI served a document request seeking, as narrowed, account information for people who used ChatGPT to investigate the claims; prompts and outputs from the plaintiffs’ pre-lawsuit testing, including materials that did not reproduce, summarize, or support the claims; and documentation of the testing process.

The plaintiffs offered to produce complete prompt-and-output threads leading to the examples in their complaint, but refused to produce testing materials that were not used in the complaint or that did not reproduce, summarize, or support their claims. Magistrate Judge Robert M. Illman previously ordered production of the account information, the prompts and outputs from the testing—including negative results—and documentation of the testing process. The plaintiffs sought relief from that order under Federal Rule of Civil Procedure 72(a).

Legal standard

Under Rule 72(a), a district judge reviews a magistrate judge’s nondispositive pretrial order deferentially. The order may be changed if it is clearly erroneous or contrary to law. The reviewing judge may not simply substitute her judgment for the magistrate judge’s.

Court’s analysis

The court explained that the work-product doctrine protects materials prepared in anticipation of litigation. It distinguished factual work product from opinion work product, which includes an attorney’s mental impressions, conclusions, opinions, and legal theories. Opinion work product receives especially strong protection.

The earlier order had treated the ChatGPT account settings and negative test results as factual work product because they appeared to be bare facts. Judge Martinez-Olguin held that this misapplied the law. She reasoned that the prompts were queries crafted by counsel and reflected counsel’s mental impressions and opinions about how to question ChatGPT while pursuing the copyright claims.

The court also rejected the earlier order’s broad waiver ruling. Work-product protection can be waived when a party reveals or places protected material at issue, but the scope of any waiver must be limited to what fairness requires. The court explained that opinion work product is discoverable based on waiver only when the attorney’s mental impressions are at issue and the opposing party has a compelling need for the material.

The court found that the earlier order did not determine that counsel’s mental impressions were at issue, and the defendants had not shown a compelling need. The defendants’ assertion that the information was needed to test the plaintiffs’ allegations, and the earlier finding that the information might help explain the positive results, were not enough to justify extending the waiver to opinion work product that was not disclosed in the complaint.

Disposition

The court granted the plaintiffs’ motion for relief. It denied the defendants’ request to compel production of all negative testing results and the documentation of the testing process. To the extent they had not already done so, the plaintiffs were required to disclose the prompts, outputs, and account settings used for the inputs associated with the positive testing results referred to and disclosed in the complaint.