INSIDE THIS REPORT

• A Connecticut federal magistrate judge ordered a climate-science expert to hand over the AI prompts she used while building her report — an apparent first.
• The court says AI prompts count as "methodology," putting them squarely inside the zone that opposing lawyers are allowed to dig into.
• The order is on hold while a higher court reviews it, but attorneys nationwide are already rewriting how they manage expert witnesses who use AI.

By Samuel López | USA Herald - A federal discovery fight in Connecticut may have just rewritten the rules for every expert witness who opens a chatbot while working on a case.

In Conservation Law Foundation, Inc. v. Shell Oil Co., a case centered on whether a New Haven fuel terminal was properly fortified against storms and flooding, Magistrate Judge Thomas O. Farrish ordered the plaintiff to turn over the generative-AI prompts used by its expert, climate historian Dr. Naomi Oreskes, while she assembled her report. Legal observers are calling it the first ruling of its kind — and one that could ripple far beyond a single environmental lawsuit.

How the fight started

Dr. Oreskes and a research assistant had used a commercial AI tool to sift through a mountain of documents produced by the defense, narrowing it down to the material she'd actually rely on. She disclosed the tool's use in her report, along with some search terms. For the defense, that wasn't enough. After her deposition raised more questions than it answered, and informal negotiations stalled, Shell's lawyers filed a motion to compel the underlying prompts. The judge sided with them.

The court's logic was blunt: expert methodology has long been fair game in discovery, and the process of using AI to whittle down a document set is a methodological choice like any other. The plaintiff tried to shield the prompts under a stipulation covering an expert's private notes and drafts. The judge wasn't buying it — prompts, in his view, aren't idle scratch notes; they're the mechanism that decided which documents the expert ever laid eyes on.

That distinction matters enormously under Federal Rule of Evidence 702 and the Daubert reliability standard. AI systems are notoriously inconsistent — identical prompts can produce different results depending on the model version, and the path from raw data to a filtered output isn't always transparent even to the person running it. If opposing counsel can't see the prompts, the court reasoned, they have no way to test whether the expert's foundation was built soundly — or whether important documents got filtered out before anyone ever reviewed them.

Notably, the judge stopped short of ruling on attorney-client privilege or work-product protections, and the case in front of him involved AI used for document sorting, not for drafting conclusions. But legal analysts warn that logic rarely stays that narrow. If methodology is the test, prompts used for literature summaries, calculations, or hypothesis testing could face the same scrutiny down the road.

A preservation trap nobody saw coming

Perhaps the bigger headache for litigators isn't the ruling itself — it's the fact that almost no one is saving this material. AI chat logs and prompt histories are routinely deleted or never recorded in the first place. A discovery request for "the prompts" can balloon into demands for every iteration, every refinement, every system instruction ever typed — records that frequently don't exist. And a missing record can look just as bad in front of a judge as a damaging one.

There's also a murkier problem lurking underneath: prompts can carry an attorney's fingerprints. A query might be a neutral factual search, or it might smuggle in counsel's legal theory or strategic framing. Courts are only starting to sort out when that crosses into privileged territory.

Attorneys handling expert witnesses should treat AI use as disclosed-by-default, not hidden-by-default. That means asking experts early and often what tools they're using, preserving prompts and outputs as they happen rather than after a subpoena arrives, and drawing a hard line between AI used for organizing information versus AI used to shape opinions. Disclosures in expert reports should be specific enough to preempt suspicion, and experts should be coached that any prompt they type could eventually land in front of opposing counsel — a strong argument for keeping strategic or privileged language out of the AI window altogether.

The Connecticut order is currently stayed while the district court reviews the plaintiff's objection, so it isn't final law yet. But its central idea — that how an expert uses AI is part of their methodology, and methodology is discoverable — is simple enough that other judges are likely to find it persuasive.

ABOUT THE AUTHOR Samuel López is a Senior Legal Analyst, investigative journalist, and legal researcher with more than two decades of experience analyzing litigation and emerging legal trends.

👉 Follow us on X: https://x.com/RealUSAHerald