A recent federal district court ruling suggests that generative AI prompts and outputs are discoverable in litigation — even those that were part of pre-suit investigation. Parameters around use of this type of evidence are still relatively unclear. Nonetheless, litigants should be thinking carefully about privilege, spoliation, reliability and authentication issues that will inevitably arise around AI-generated materials.
Use of prompts and outputs as evidence has arisen most prominently in the context of copyright infringement lawsuits that large classes of authors and creators such as Sarah Silverman and Ta-Nehisi Coates have filed against AI companies.
In Tremblay v. OpenAI, filed in the U.S. District Court for the Northern District of California last June, the authors accused OpenAI of using their copyright-protected materials without permission to train its ChatGPT platform. They alleged that both copying at the input/training stage and at the output stage constituted infringement.
The court granted OpenAI's motion to dismiss the majority of the authors' claims, finding the authors had failed to allege facts sufficient to support their allegation that ChatGPT outputs were "substantially similar" to the original inputs. However, the court allowed the authors to replead with additional support for their claims.
In hopes of avoiding dismissal the second time around, the authors alleged in their amended complaint that particular prompts in ChatGPT would cause ChatGPT to reproduce portions of their books almost verbatim. This additional detail elicited discovery requests for pre-suit testing of ChatGPT.
OpenAI argued to the court that the authors should be required to hand over the ChatGPT account information for anyone who entered the prompts listed in the authors' complaint, as well as all prompts and ChatGPT outputs that the authors or their lawyers tested, but did not ultimately include in the complaint.
The authors argued that because the ChatGPT prompts were entered in connection with preparing the lawsuit, the prompts, outputs, and related information were protected from discovery by the attorney work product doctrine. That doctrine protects documents that were prepared by a lawyer, or by someone else at a lawyer's direction, in connection with anticipated or pending litigation.
A magistrate judge handling discovery disputes in the case agreed with OpenAI, finding that the authors had waived work product protection by referencing the prompts and outputs in their complaint. The authors appealed the ruling to the district judge, who overturned the magistrate judge's decision.
The district judge concluded that the authors' waiver of the work product doctrine should be limited only to those prompts and outputs actually referenced in the complaint, as well as the ChatGPT account setting in place when those prompts were entered. Although this decision put some guardrails around the magistrate's original ruling, it signals that other plaintiffs looking to sue generative AI platforms for infringement are likely going to find themselves in the difficult position of having to reveal information that might otherwise be considered privileged in order to plead viable claims against the platforms.
In the United States, AI companies do not have to disclose what data their generative AI programs were trained on, though there has been legislation proposed to address this issue at the federal level. For now, plaintiffs alleging infringement generally have no direct way to show whether their works were copied for use in AI training. As a result, they have to offer indirect evidence of copying by, among other things, prompting the AI programs to see if the resulting output is similar to their works.
Under the Tremblay district judge's approach, plaintiffs who make those types of allegations will at least have to turn over all prompts and outputs referenced in the complaint and certain account settings. But if judges in other parts of the country find the magistrate judge's reasoning more persuasive, plaintiffs may find themselves being forced to turn over the keys to their generative AI program accounts. Plaintiffs will either have to take care in drafting their complaints to try to avoid a privilege waiver, or otherwise be ready to give the AI company their AI program account information.
The Tremblay ruling will reverberate far beyond the copyright context as use of ChatGPT and similar programs becomes relevant and discoverable in all types of litigation. For example, where AI is used for employment screening, records of the AI program usage and outputs may be highly relevant in discrimination cases. And where a vendor uses generative AI as part of its provision of services to a client pursuant to a contract, the generative AI prompts and outputs could be relevant to a lawsuit for breach of that contract.
This triggers preservation obligations to avoid sanctions for spoliation of evidence. If they are not already doing so, companies that use generative AI should take measures similar to those implemented for email retention. They should ensure that employee use of generative AI is appropriately monitored, and that prompts and outputs are preserved for a set period of time or, when litigation is imminent, indefinitely.
Moreover, when litigation has commenced or is imminent, companies or their lawyers often issue "litigation hold" notices to advise company employees to preserve all relevant documents and communications. Most litigation hold notices have already been updated to cover relatively recent forms of communication (e.g., social media), but companies and lawyers that have not already done so should update their litigation hold notices to cover generative AI prompts and outputs where applicable.
Litigants should also consider what steps may be required to authenticate AI-generated evidence for use in court. This past spring, the U.S. Judicial Conference's Advisory Committee on Evidence Rules heard testimony from experts on AI to evaluate whether the current Federal Rules of Evidence are equipped to address use of AI-generated material as evidence.
The committee released a report considering questions such as whether experts or validation studies are required to testify to reliability of AI-generated content, and whether a special rule should be created to address authentication of potential "deepfakes." Report of the Advisory Committee on Evidence Rules for the Judicial Conference of the United States' Committee on Rules of Practice and Procedure, May 15, 2024 (https://bit.ly/4dYuUCJ).
No definitive changes have been made. Because rulemaking is such a lengthy process, it is likely that courts will be forced to start weighing in on the application of existing rules in the context of this rapidly evolving technology.
On a piecemeal basis, some courts have already issued standing orders requiring disclosure of the use of AI to draft pleadings, and/or requiring attorneys to certify the accuracy of AI-generated material before submitting to the court. Meanwhile, state bars, including those in Florida and New Jersey, have issued ethics opinions addressing use of AI. Florida Bar Ethics Opinion 24-1, 2024 Florida Court Order 0005 (C.O. 0005), issued Jan. 19, 2024; Preliminary Guidelines on the use of Artificial Intelligence by New Jersey Lawyers, 2004 New Jersey Court Order 0004 (C.O. 0004), issued Jan. 24, 2024.
Although the technology is still in its relatively early days, generative AI presents new issues and challenges to which U.S. courts, companies, lawyers, and others are all trying to adapt. The Tremblay rulings are a reminder that, as with any new technology, individuals and businesses need to consider how generative AI impacts their practices and whether those practices need to change to address generative AI.
Originally published by Reuters and Reuters' Westlaw Today
The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.
