A federal court just delivered one of the clearest messages yet on AI in litigation: if an expert used AI to do the work, the prompts may be discoverable. In Conservation Law Foundation, Inc. v. Shell Oil Company et al., Magistrate Judge Thomas O. Farrish ordered the Plaintiff to produce the prompts its expert used in preparing her report, treating them as part of the expert’s methodology rather than protected drafting material. That puts AI prompts squarely into the discovery fight.
Background
The dispute came out of expert discovery in CLF’s case against Shell and other defendants. CLF’s expert disclosed that she used AI to help analyze the Defendants’ document production. Working with a research assistant, the expert used OpenAI models to filter and categorize documents before reviewing the results herself. At her deposition, the expert described the tool as a “sieve” for identifying potentially relevant documents. But when asked for “[a] list of the Bates numbers of all Defendants’ produced documents that were uploaded, ingested, or otherwise made available to any AI tool,” CLF declined.
After months of meet-and-confer sparring, the parties narrowed the fight to a simple question: did CLF have to turn over the prompts and queries tied to that AI-assisted workflow?
The Court heard argument on May 14, 2026, and answered yes four days later.
The Defendants’ Motion
Defendants framed the issue as classic Rule 26 discovery. If the expert used AI to narrow the document universe, they argued, then the prompts, outputs, and related workflow details were fair game.
They also raised a spoliation angle. In a declaration, the expert’s research assistant said he had not preserved a complete native prompt/output log. Defendants seized on that statement and argued sanctions should be on the table.
The motion also had a data-security dimension. Because the expert had used “ChatGPT on a secure server” to analyze confidential materials, the Defendants wanted more detail about the setup and whether logging or model-training features had been disabled.
CLF’s Response
CLF pushed back, calling the requests overbroad, duplicative, and outside the proper scope of discovery.
Its core point was that the AI did not write the report. CLF said the tool merely applied search terms to a large document set, flagged potentially relevant materials, and left the real analysis to human review. In CLF’s telling, this was more like Technology Assisted Review than generative AI opinion-making.
CLF also leaned on the parties’ FRCP 29 stipulation, arguing that AI inputs, outputs, and related materials counted as protected expert notes or drafts.
And CLF said there was nothing more to give. It maintained that “Defendants have the list of documents that make up the universe of Defendants’ materials [the expert] was working from,” that the Defendants “have the search terms [the expert] provided to the AI tool,” and that the methodology had already been fully explored in discovery.
On the security front, CLF responded that the documents were handled in a private Microsoft Azure environment and were not used to train outside models.
The Court’s Analysis
At the hearing, the fight boiled down to three arguments. CLF lost on all three.
First, CLF said AI prompts fell outside Rule 26(b). The Court disagreed. The Court reasoned that an expert’s methodology is discoverable, and using AI to cull an opposing party’s documents was part of that methodology. Relying on Macchia v. ADP, Inc., 711 F. Supp. 3d 162, 167–68 (E.D.N.Y. 2024), the Court treated AI-assisted analysis as a step in the expert process.
Second, CLF argued the prompts were shielded by the Rule 29 stipulation covering expert “notes, drafts, or communications.” The Court rejected that too, finding it was not clear enough that those terms covered AI prompts.
Third, CLF said no additional materials existed because the expert used only “search terms,” not “prompts.” But according to the Court, a declaration from the expert’s research assistant referring to “prompt[s]” gave the Defendants an “evidence-backed reason for doubting CLF’s representation.” The Court noted that a good-faith statement that nothing exists usually ends the issue, but not when the record gives the other side real reason to doubt it.
Implications for Practitioners
Although this is only one decision from one court, litigators should take note when they retain testifying experts who use AI. This decision demonstrates that some courts may treat AI prompts and related materials as part of an expert’s discoverable methodology. If AI helped sort documents, test inputs, or shape outputs that informed the opinion, it is possible that the other side will ask for the receipts.
Given this possibility, practitioners should ask questions to understand how their experts plan to use AI at the outset of an engagement. They should also carefully consider whether to instruct their testifying experts to preserve information about their use of AI workflows. Parties should also consider the AI overlay in the negotiation of discovery stipulations.
This case also serves as a good reminder of the importance of being accurate and precise in communications with adversaries and the Court regarding discovery disputes. The Court rejected CLF’s attempt to characterize the disputed material as “search terms,” in part because the research assistant’s declaration called them AI “prompts.”
Bottom line: as testifying experts incorporate AI into their methodologies, there is a significant possibility that opposing counsel will seek discovery into their experts’ use of AI. Lawyers should plan for that now.
