[Editor’s Note: This article has been republished with permission. It was originally published June 19, 2026 on the Minerva26 Blog]

A first-of-its-kind decision out of the District of Connecticut held that AI prompts used by an expert witness are discoverable methodology under Rule 26 — and the district court stayed the order sixteen days later. Every litigator working with experts who use AI needs to understand what the court said, what CLF is now arguing on the stay, and what comes next.

Case: Conservation Law Found., Inc. v. Shell Oil Co., Civil No. 3:21-cv-00933 (VDO) Court: District of Connecticut | Decision: May 18, 2026 (docket text order, ECF No. 970) | Judge: United States Magistrate Judge Thomas O. Farrish

A landmark holding, on hold

On May 18, 2026, in a docket-text order with no accompanying written opinion, a federal magistrate judge held for the first time that an expert witness’s AI prompts are discoverable methodology under Rule 26. Sixteen days later, on June 3, the district court stayed the order pending review of the plaintiff’s Rule 72(a) Objection. That is the posture you are working with as you read this.

The case is Conservation Law Foundation v. Shell Oil, a Clean Water Act and RCRA citizen suit against the operators of a bulk petroleum storage terminal on the waterfront in New Haven, Connecticut. Plaintiff CLF retained Dr. Naomi Oreskes — the Harvard historian of science and co-author of Merchants of Doubt — as a testifying expert. Dr. Oreskes and her research assistant used AI to cull Shell’s document production down to a working subset for her analysis. Shell moved to compel production of the AI prompts and outputs Dr. Oreskes’ team used. Magistrate Judge Thomas O. Farrish granted that motion at ECF No. 970, in a docket-text order. CLF filed both an emergency motion to stay and a Rule 72(a) Objection on the day compliance was due. Judge Vernon Oliver granted the stay.

The framework Magistrate Judge Farrish laid down — that an expert’s AI methodology is fair game under Rule 26(b), that prompts are part of that methodology, and that a “search terms only” disclosure does not cover the field — is now public. If Judge Oliver vacates it in a published opinion, the framework loses its precedential force and may well fall out of the conversation. If he leaves it in place, modifies it, or addresses the issue in any unpublished disposition, the reasoning will continue to influence how other courts approach the question. Either way, the analysis is on the record now, and litigators advising on AI use cannot afford to wait for the district court’s ruling before acting.

This episode of Case of the Week walks through what the order held, what CLF is now arguing on the stay, what happens if no prompts exist, how Florida’s new AI accountability rule fits into the same conversation, and what every litigator advising clients on AI use should be doing about it this week. We have been building this body of law across several Case of the Week episodes — Episode 183 on Heppner and Warner v. Gilbarco, and a more recent episode on Morgan v. V2X, Inc. If you are looking for the foundational analysis of those decisions, those episodes are the place to start. The focus here is on what Conservation Law Foundation adds, and on what comes next.

The case underneath

Before we get to the discovery ruling, the underlying litigation matters. Conservation Law Foundation v. Shell is one of the most consequential environmental cases currently in active discovery in the United States. CLF, working with co-counsel at Motley Rice, filed citizen suits in 2021 alleging that Shell and the other operators of the New Haven petroleum terminal failed to design, maintain, modify, or operate that facility to account for the effects of climate change — sea-level rise, more frequent and more severe storms, storm surge — and that those failures create an imminent risk that the next major weather event will discharge oil and toxic chemicals into the surrounding waters and neighborhoods. A parallel CLF case proceeds against Shell in the District of Rhode Island.

These cases are among the first in the country to hold fossil fuel infrastructure operators accountable not for greenhouse gas emissions but for failing to prepare their physical facilities for climate change. Per CLF, they are also the first U.S. climate adaptation cases of their kind to reach the discovery phase. Dr. Oreskes — whose academic work has focused on the history of climate science and the manufactured doubt produced by fossil fuel companies — was retained to address what Shell knew, when it knew it, and what its internal records reveal about its understanding of the risks at issue. Working through Shell’s discovery production was central to her work. Using AI to make that document review tractable was central to her methodology.

That context matters because the stakes on the discovery ruling are not abstract. Whether Magistrate Judge Farrish’s framework survives district court review will shape both how Dr. Oreskes’ testimony comes in at trial and how every other testifying expert in active litigation handles AI from this point forward.

What Magistrate Judge Farrish actually held

The order is short. Three CLF arguments. All three rejected by Judge Farrish. The court ordered CLF to revise its discovery responses to disclose the AI prompts and queries Dr. Oreskes’ team used, with Rule 37(b) sanctions teed up if CLF later represents that nothing exists and that representation proves untrue. The compliance deadline was June 1.

CLF’s first argument was scope. CLF argued that AI prompts used by an expert are not within the scope of discovery at all. Magistrate Judge Farrish rejected that in a sentence, citing Macchia v. ADP, Inc., 711 F. Supp. 3d 162, 167-68 (E.D.N.Y. 2024), for the proposition that an expert’s methodology is fair ground for discovery. Under the facts here, the process by which Dr. Oreskes culled Shell’s production into a subset to be worked with is an aspect of that methodology. The principle is not new. Discovery of TAR protocols, search term lists, sampling rules, and coding instructions has been settled doctrine since Da Silva Moore in 2012. AI prompts are the next iteration of the same idea. What is new is the application to a testifying expert.

CLF’s second argument was the Rule 29 stipulation. The parties had agreed early in the case to limit expert discovery — specifically, that 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, would not be subject to discovery. CLF argued at the hearing that Dr. Oreskes’ AI prompts qualified as “notes” within that carve-out. Magistrate Judge Farrish did not agree. He had already addressed Rule 29 in this case at Conservation Law Foundation v. Shell Oil Co., 2025 WL 842278, at *3 (D. Conn. Mar. 18, 2025), holding that before a court denies otherwise-relevant discovery on a Rule 29 basis, the agreement “must be quite clear.” Calling an AI prompt a “note” is not so obvious as to be quite clear.

The practical move for litigators is straightforward and is the single most actionable thing in this opinion: Rule 29 agreements have to say what they mean. If you are negotiating a discovery stipulation today and you intend it to cover AI inputs and outputs, name them. Use “AI prompts,” “AI queries,” “AI outputs,” “embeddings,” and “model parameters.” Generic language — “notes,” “drafts,” “communications” — will not be read to sweep in AI artifacts. After this case, if it holds up, opposing counsel and the court will both insist on specificity. Magistrate Judge Braswell’s protective order language from Morgan v. V2X, Inc. is one starting point for that drafting, but it does not address experts specifically. You will need to extend it.

CLF’s third argument was that nothing exists to produce. CLF took the position that Dr. Oreskes did not use “prompts” at all — she used “search terms” — and that all search terms had already been produced. The court walked through the analytical framework that governs “nothing to produce” arguments. The default rule is that when a producing party tells the court the requested material does not exist, and the court has no reason to doubt that representation, the court will not order production of the impossible. That is

ACEDS

Message History

Delivery option(s)