In a decision issued yesterday, a Fed.R.Civ.P. 29 stipulation provided that expert’s “notes” were not discoverable. An expert’s prompts or queries were not “notes.” The court also held that the prompts were “reliance materials” that were discoverable under the Federal Rules of Civil Procedure:
Here, the Court concludes that artificial intelligence prompts or queries used by Dr. Oreskes or her team in the course of producing her expert witness report is discoverable under Rule 26(a)(2)(B)(i) because they comprise “the basis” for the opinions that Dr. Oreskes will express.
Conservation Law Foundation, Inc. v. Equilon Enterprises LLC, 2026 WL 3571573, at *4 (D. Conn. Nov. 20, 2026).
Discovery of expert’s use of AI, and the manner of use, is a hot topic:
THE FACTS: HOW WAS AI USED?
In Conservation Law Foundation (“CLF”), the District Court affirmed a Magistrate Judge’s “order compelling production of reliance materials related to the use of artificial intelligence.” CLF filed an environmental suit three companies challenging a bulk storage and fuel terminal in Connecticut.
As is relevant to the current objections, Plaintiff engaged Dr. Naomi Oreskes, the Charles Henry Lea Professor of the History of Science and Affiliated Professor of Earth and Planetary Sciences at Harvard University, to be an expert witness in this case. In writing the opening report, Dr. Oreskes relied on materials produced in this litigation, and employed a computational scientist, Dr. Alexander A. Kaurov, as a research assistant to assist with her review of materials produced in this litigation. After Plaintiff provided access to the documents produced in this action, Dr. Kaurov processed those documents using a dedicated Microsoft Azure cloud environment created solely for the project. To review and search the produced documents in this litigation, prompts, queries, and model outputs were used interactively as intermediate tools to identify potentially relevant documents for further review. Dr. Kaurov used ChatGPT to search for documents and to identify the documents for Dr. Oreskes to review. While doing so, Dr. Kaurov did not preserve complete native “prompt/output log” standalone files on any personal systems or in a cloud environment.
Id. at *2 (emphasis added). Chat GPT was used “on a secure server.” Id. at n. 14.
THE DISPUTE
Defendant’s position was based on Fed.R.Civ.P. 26(a)(2)(B):
On December 23, 2025, Defendants moved to compel production of artificial intelligence (“AI”) reliance materials. Defendants argued that Plaintiff violated Federal Rule of Civil Procedure 26(a)(2)(B) by refusing to produce the underlying facts, data, and documents that construct the methodology Dr. Oreskes relied on to use AI. Defendants noted that Dr. Oreskes relied on an AI workflow to form substantive opinions in her expert report and that she and Dr. Kaurov jointly developed prompts used to analyze Defendants’ document production.
Id. at *2. The parties conferred, were unable to reach agreement, and after briefing the Magistrate Judge granted the motion to compel. The matter was before the District Court on Plaintiff’s objections. The District Judge affirmed.
The materials had been requested in an email. The court “easily” rejected Plaintiff’s argument that it was error to enforce “informal” discovery requests, finding the argument was waived. Id. at *2. “But the Court exercises its discretion here to note that it was neither clearly erroneous nor contrary to law to compel production of materials requested through email.” Id. at *3.
THE FED.R.CIV.P. 29 STIPULATION
After rejecting an argument that the motion was untimely: “The Court similarly rejects the argument that the Magistrate Judge misinterpreted the parties’ stipulation under Federal Rule of Civil Procedure 29(b).” It wrote:
Here, there is no dispute that the parties are bound by a Rule 29 stipulation, which includes the following Non-Discoverability provision:
Non-Discoverability: Drafts of expert reports and communications between experts, employees and consultants of Party-related entities on expert opinions or settlement-related issues, including emails, notes, and other correspondence, shall not be subject to discovery. This includes any preliminary findings or informal discussions that have occurred between experts employees or consultants of Party-related entities.
Thus, the parties stipulated to remove certain categories of documents related to the expert report drafting process from discovery, including drafts, communications, emails, notes, and other correspondence.
The CLF court explained:
It was not clearly erroneous for the Magistrate Judge to distinguish AI prompts or queries from the categories of documents in the agreement here, where the agreement does not even mention the use of artificial intelligence. “Prompts,” as referenced here and in Dr. Kaurov’s declaration, refers to how users interact with ChatGPT. ChatGPT’s official website indicates that “(p)rompting is how you tell ChatGPT what you want to know, make, or change” and that “[a] prompt can be a question, an instruction, or a goal.” The website further provides that, after reviewing the response from ChatGPT, users can “use follow-up messages to shape the result.” Against this backdrop, it was not clearly erroneous for the Magistrate Judge to conclude that AI prompts are distinguishable from an expert’s “notes” related to draft expert opinions and to reject the interpretation of the stipulation in a way that excluded AI prompts from the universe of discoverable materials.
Id. at *4.
THE FED.R.CIV.P. 26 ANALYSIS
Having decided the issue based on the Fed.R.Civ.P. 29 stipulation, the CLF court also rejected Plaintiff’s objections on the merits, under Fed.R.Civ.P. 26(b)(relevance).
As to Rule 26(b), this Court concurs with the reasoning provided by the Magistrate Judge in granting the motion to compel. The Magistrate Judge did not clearly err in concluding that an expert witness’s methodology is fair ground for discovery. Plaintiff’s objection does not instill in the Court a conviction that the Magistrate Judge erred, because this Court concludes that the process by which Dr. Oreskes culled down the defendants’ document production into a subset to be worked with falls within the scope of discovery, as it is nonprivileged matter that is relevant to a party’s claim or defense and proportional to needs of the case. Fed. R. Civ. P. 26(b). [Magistrate] Judge Farrish adequately considered how Defendants satisfied the “relatively low threshold for a party to show that the material sought is relevant to any claim or defense in the litigation.”
Id. at *4 (emphasis added).
The CLF court then looked at Fed.R.Civ.P. 26(a)(2)(B) and held that it also compelled disclosure of “artificial intelligence prompts and queries used by Dr. Oreskes or her team in the course of producing her expert witness report.” It wrote that “Witnesses who are specially employed to provide expert testimony in a case have an affirmative duty to provide a report containing ‘a complete statement of all opinions the witness will express and the basis and reasons for them;’ and “the facts or data considered by the witness in forming them. Fed. R. Civ. P. 26(a)(2)(B)(i), (ii)….” It explained:
ChatGPT’s website indicates that “
rompting is how you tell ChatGPT what you want to know, make, or change” and that, as part of the prompting process, users can “use follow-up messages to shape the result.”
To review and search the produced documents in this litigation with ChatGPT, Dr. Oreskes or Dr. Kaurov used prompts or queries and model outputs as an intermediate tool to identify potentially relevant documents for further review. At her deposition, Dr. Oreskes testified that she used “AI as a sieve to identify all the documents that could be relevant to a particular question” and to “go through all of the documents provided by the Defendant to find all of the documents that mention sea level rise.” To do this, Dr. Oreskes discussed with her research assistant what the prompts would be and reviewed the documents after the analysis was run. Thus, it appears that the prompts and queries comprise the basis for Dr. Oreskes’s opinions.
Id. at *4 (emphasis added).
THE “PROMPTS NO LONGER EXIST” ARGUMENT
Plaintiff contended that the prompts no longer existed. “Finally, the Court rejects Plaintiff’s argument that the Magistrate Judge erred in compelling production of materials that, based on counsel’s representations, do not exist.” Id. at *5.
[Magistrate] Judge Farrish did not clearly err in concluding that there was an evidence-backed reason for doubting Plaintiff’s representation that materials never existed. Dr. Oreskes’s assistant, Dr. Kaurov, referenced “prompt[s]” in his declaration. Specifically, to review and search the produced documents in this litigation, Dr. Kaurov represents that prompts or queries and model outputs were used interactively as an intermediate tool to identify potentially relevant documents for further review. Indeed, Dr. Oreskes discussed Dr. Kaurov, her research assistant, what the prompts would be and reviewed the documents after the analysis was run. The testimony under penalty of perjury provided by Dr. Kaurov and Dr. Oreskes comprise “solid evidence” that AI prompts or queries used by Drs. Oreskes and Kaurov existed at one time, and thus, an order of production is warranted.
Id. at *5. Plaintiff’s objections to the Magistrate Judge’s order were overruled and the order affirmed in its entirety.