An expert may see AI as a faster way to find the right documents. Opposing counsel may see the prompts as a map of how the expert built the universe of evidence. That tension is at the center of Conservation L. Found., Inc. v. Shell Oil Co., No. 3:21-cv-00933 (VDO), ECF No. 970 (D. Conn. May 18, 2026) (Farrish, M.J.).
There, an expert and her research assistant used generative AI to narrow a large document production before the expert completed her analysis. Defendants moved to compel, seeking in part the methods the expert used to cull the production. The court concluded that the process used to reduce the document set was part of the expert’s methodology — and, therefore, fair ground for discovery under Federal Rule of Civil Procedure 26(b).
The plaintiff said the expert had used search terms, not prompts, and that all responsive terms had already been produced. The court recognized that a movant needs evidence, and not just suspicion that a party has failed to make a complete production. But, here, the research assistant’s declaration referred to the use of “prompts.” The court viewed that as an evidence-backed reason to require more than the plaintiff’s assurance that nothing else existed.
The court ordered plaintiff to serve revised discovery responses that either disclosed any additional prompts or queries used by the expert or her team — or formally confirmed, after a diligent search, that none existed. The court cautioned that plaintiff could be sanctioned if that representation was later shown to be untrue.
The order is stayed while the district judge considers the plaintiff’s objection, so this remains a developing issue. The operative order also expressly addressed prompts and queries; it did not command production of every AI output.
Still, the practical message is difficult to miss.
AI Does Not Have to Write the Expert Report to Affect the Expert’s Work
If it helps decide which documents the expert sees, which authorities are surfaced, or which data points receive attention, it may become part of the expert’s method, opening up opposing counsel’s right to test.
Get Permission Before the First Prompt
An expert — or anyone working for the expert — should consider discussing the introduction of an AI tool into an engagement with retaining counsel before doing so. Retaining counsel, in turn, might wish to obtain or confirm the client’s approval. That discussion may be useful before case information is entered into the tool. Counsel, the client, and the expert might want to agree on:
- The approved AI platform and deployment;
- The specific tasks the tool may perform;
- The information that may — and may not — be entered;
- Which members of the expert team may use it;
- How prompts, outputs, and related records will be preserved; and
- How the expert will review and verify the results.
A general authorization to use research assistants, software, or outside vendors may not necessarily be treated as permission to use any AI platform for any purpose. This is especially important if prompts contain privileged communications, attorney strategy, personal information, trade secrets, or documents designated confidential. Depending on the platform and its terms, entering that information may create privacy, privilege, work-product, cybersecurity, contractual, or confidentiality issues.
The named expert may wish to understand not only that AI was used, but how it was used. Counsel and the client should not learn about that workflow before the expert’s deposition.
A Protective Order Does Not Answer Every AI Question
The case also illustrates that two separate questions may arise:
- Was the AI tool used in a way that protected confidential information?
- Are the prompts or related records discoverable?
The parties had a standing protective order and later reported that they had resolved their dispute over whether the expert’s AI use violated it. The May 18 order therefore did not find a protective-order violation.
According to the plaintiff’s filing, the expert team used AI models through a private environment and reviewed the provider’s security terms. The plaintiff represented that prompts and outputs were not available to other customers or used to improve the models. Those were the plaintiff’s representations, not findings made in the May 18 order.
The parties also had a separate Rule 29 agreement limiting discovery of expert notes, drafts, and communications. But that agreement did not expressly address AI prompts, and the court concluded that simply calling prompts “notes” was not sufficiently clear to place them outside discovery.
Practitioners might not assume that a standard protective order or expert-discovery agreement covers AI merely because it addresses confidentiality, drafts, or communications.
Practitioners may wish to address AI expressly, including:
- Permitted platforms and security requirements;
- Data retention and model-training restrictions;
- Access to confidential or privileged material;
- Preservation of prompts, queries, and outputs;
- Whether specified AI materials are subject to expert discovery; and
- What happens when a tool does not retain — or cannot export — the relevant history.
If the parties intend to limit discovery of expert AI materials, they may wish to say so directly in the agreement. Whether a court will enforce that limitation would still depend on the circumstances and applicable law.
What Might Change Now?
Expert provider companies may consider maintaining an approved-tool list, requiring engagement-specific authorization, and applying those rules to employees, assistants, and subcontractors. They may also wish to understand whether the approved platform can preserve and export the relevant history.
Experts should understand and may wish to supervise any AI-assisted work performed by their teams. They might consider validating the results and using precise language when describing the process. “Search terms,” “prompts,” “queries,” and “outputs” may not be treated as interchangeable when discovery begins.
Retaining litigators may consider asking about AI before the engagement starts, confirming client approval, and addressing the issue in the engagement letter. They should also review the protective order, consider whether a more specific discovery agreement is appropriate, and set preservation expectations before the first AI session.
Litigators examining an opposing expert might keep the inquiry tied to the actual methodology and ask which tool and version were used, who operated it, what task it performed, what information it received, how the results were reviewed, and what records still exist. A focused inquiry may be more useful than a broad demand based only on the possibility that AI was involved.
The lesson is not that experts should avoid AI. AI can still be a valuable expert tool. But in expert work, it should be treated as an engagement decision — not simply a personal productivity choice.