When Your Expert Uses AI, the Prompts Are Fair Game: Reading the CLF v. Shell Order

On May 18, 2026, Magistrate Judge Thomas O. Farrish granted a motion to compel in Conservation Law Foundation, Inc. v. Equilon Enterprises LLC d/b/a Shell Oil Products US (D. Conn. No. 3:21-cv-00933), ordering the plaintiff to produce the artificial intelligence prompts and queries its expert used to build her report. The ruling is short. Its logic is the part worth your attention.

CLF's expert, Dr. Naomi Oreskes, used an AI workflow to cull Shell's large document production down to a working subset for her analysis. When the defendants asked for the prompts and outputs behind that process, CLF resisted on three grounds. The court rejected all three, and the way it did so tells you how courts are likely to treat AI-assisted expert work going forward.

First, CLF argued that AI prompts fall outside the scope of permitted discovery. The court disagreed, holding that an expert's methodology is fair ground for discovery, and that the process by which Dr. Oreskes narrowed the document set was part of that methodology. The court did not treat AI as a special category. It treated the prompts as the mechanism the expert used to reach her conclusions, no different in kind from any other methodological step an opposing party is entitled to probe.

Second, CLF argued that the parties' Rule 29 discovery agreement shielded the prompts, characterizing them as the kind of "expert notes" the parties had agreed not to pursue. The court applied the rule that an agreement limiting otherwise-relevant discovery must be "quite clear" before a court will enforce it to deny discovery. Calling AI prompts "notes" was not obvious enough to clear that bar.

Third, CLF argued it had nothing more to produce, contending that Dr. Oreskes used only "search terms" rather than "prompts," and that the search terms had already been turned over. Here the court drew a useful line. The general rule is that a party's good-faith statement that responsive material does not exist resolves the issue, because no one can be compelled to produce the impossible. But that rule gives way when the requesting party has a reason to disbelieve the representation that is backed by solid evidence rather than mere suspicion. The defendants had that evidence: Dr. Oreskes's assistant had referenced "prompt[s]" in his own declaration. I appreciated the court’s approach: it ordered CLF to formally revise its Rule 33 and 34 responses, and if a diligent search turns up nothing more, to say so in a signed response, with Rule 37(b) sanctions available later if that representation proves untrue.

That third move is the practical teeth of the order. The court did not simply accept the semantic distinction between "prompts" and "search terms." It ordered CLF to respond to all discovery requests “that call for disclosure of any artificial intelligence prompts and/or queries used by Dr. Oreskes or her team in the course of producing her expert witness report.” Judge Farrish noted that, if CLF’s representations turn out to be untrue, the other party can move for sanctions.

Notably, the underlying motion also raised concerns about whether the expert uploaded confidential documents to an AI platform in violation of the parties’ protective order (a protective order enables parties to exchange sensitive and confidential information without it becoming public knowledge). The court's order, however, ultimately focused on discoverability of the expert's methodology rather than issues of privilege or confidentiality. I expect to see this issue raised again after this piece of the discovery process is complete.

The Practical Takeaway

The lesson from CLF is not that AI is dangerous; the lesson is that AI is a methodology, and methodologies used by testifying experts are subject to scrutiny.

Experts who intend to use AI should discuss that methodology with counsel early in the engagement, not after the report is drafted and the other party starts sending discovery requests.

Counsel needs to understand what tools are being used, what information is being uploaded, whether prompts and outputs are being preserved, and whether the workflow could potentially violate the party’s confidentiality obligations or the court’s protective order.

Just as importantly, experts should assume that prompts, queries, and outputs may eventually become part of the discovery record. If AI is being used to identify documents, analyze evidence, or otherwise assist in forming opinions, preservation of those materials may become as important as preservation of the underlying documents themselves.

The broader point is that courts are not creating special rules for AI-assisted expert testimony. They are applying familiar discovery principles to a new tool. An expert's methodology has always been fair game. CLF suggests that AI prompts are now part of that methodology.