New AI-Discovery Rulings: Risks for Experts

Two federal courts examining very similar facts — experts using generative AI (ChatGPT, in both cases) to help prepare their reports — have reached different conclusions about what has to be turned over in discovery. The law here is not settled, and an expert's AI workflow may be treated very differently depending on which court is asking.

Higher-risk AI uses (more likely to be ordered disclosed):

• Using AI to generate case summaries, legal standards, or factual material that experts consult in reaching their opinion — even just as a quick reference. One court held that this counts as "facts or data considered" and must be produced, full stop.
• Using AI to help decide which documents or data to review in the first place. One court treated this as part of an expert's methodology — comparable to disclosing the expert's dataset-selection criteria — and ordered it produced.

Lower-risk AI uses (more likely to be protected):

• The prompts and questions an expert typed, when used the way one uses a research-engine query. One court found prompts aren't themselves "facts or data" and would improperly expose an expert's mental impressions and case theory.
• AI-generated deposition or trial-prep outlines (e.g., "what might the other side ask me?") — protected as trial-preparation material revealing an expert's focus and concerns, not an expert's opinion.
• AI-assisted drafting experts incorporate into your own report — protected as a draft, similar to any other draft of a report.

Even where content is protected from discovery, it isn't necessarily off-limits at trial: courts have signaled that AI methodology — including the risk of hallucinated citations — is fair game for cross-examination.

The two cases

Villanueva v. Las Vegas Metropolitan Police Department (D. Nev., Sept. 29, 2026): In this §1983 civil-rights case, plaintiff's expert Tom Melton testified he'd used ChatGPT to research and draft his report. He hadn't disclosed this in response to an earlier subpoena, and once it came out, the defense sought his full 40-page ChatGPT history log. The magistrate judge initially ordered the whole log produced; on review, District Judge Traum split it three ways. AI-generated case summaries the expert consulted: produced, because the Ninth Circuit reads "facts or data considered" broadly to include anything an expert looked at with factual content, not just what he relied on. His prompts and questions, his AI-generated deposition outline, and AI material he drafted from: all protected, because they reveal his thinking and trial strategy rather than facts underlying his opinion. The court also held that using a generative AI tool doesn't waive privilege, reasoning that the AI platform is "not a person" but a research/drafting tool — and noted that methodology concerns can be addressed through impeachment at trial instead.

Conservation Law Foundation v. Shell Oil Co. (D. Conn., May 18, 2026): Defendants moved to compel production of the AI prompts that CLF's expert, Dr. Naomi Oreskes, used while preparing her report — specifically, the prompts she used to cull a large defendant document production down into a working subset. CLF resisted on three grounds: that AI prompts fall outside the scope of discovery altogether; that the parties' existing Rule 29 agreement (which protected “expert notes, drafts, or communications” made during report drafting) covered the prompts as “notes”; and that there was nothing left to produce, since the expert had used only “search terms,” not “prompts,” and had already turned those over. Magistrate Judge Thomas Farrish rejected all three. He held that an expert's methodology is a proper subject of discovery, and that using AI to narrow down which documents to work with is itself part of that methodology. He found the Rule 29 agreement's coverage of “notes” wasn't clear enough to bar discovery of AI prompts. And he found CLF's claim of nothing-to-produce undercut by its own expert's assistant, who referred to “prompts” in a sworn declaration. CLF was ordered to disclose any AI prompts or queries used by the expert or her team in producing the report, due June 1, 2026.

Legal analyses have framed the split this way: both federal courts agree that outputs an expert actually relies on to form an opinion are discoverable. They disagree on how to treat the prompts themselves — Nevada analogizes them to private research queries; Connecticut analogizes them to disclosable search methodology.

Why this matters beyond these two cases

This split isn't isolated — it reflects that rulemakers haven't settled on an approach. The federal Advisory Committee on Evidence Rules has been developing a proposed Rule of Evidence 707, aimed at AI-generated evidence that functions like expert opinion, since 2025. After a public comment period that closed in February 2026, the committee chose at its May 2026 meeting not to finalize or republish the rule — instead withdrawing the published draft to study a revised version, with further discussion expected at a fall 2026 mini-conference alongside a related proposed rule on deepfakes. That pushes any final rule out by roughly another year at minimum. For now, there's no uniform federal standard — just scattered, sometimes conflicting, district-court rulings like these two.

Until this settles, expert witnesses may assume their AI usage could be scrutinized differently depending on where a case is filed. Separating AI use by category (research vs. deposition prep vs. drafting) and keeping clear records of each may help experts and retaining counsel if privilege arguments are raised later, whatever a given court ultimately decides.

Status note: As of this writing, the Villanueva ruling isn't final. Defendants filed a motion for reconsideration of Judge Traum's September 29 order on October 20, 2026, with the plaintiff's response due November 3, 2026 — the same date CLF's objection is set for oral argument in Connecticut. Neither “final word” referenced in this summary is actually final yet. 

This article was prepared with the help of an AI assistant (Claude, by Anthropic) and reviewed and edited by LexVisio.

Sources:

1. Federal Court Splits On Disclosure Of Expert Witness AI Prompts (Mayer Brown)
2. Analysis: Proposed FRE 707 Stalls — But Is That Such a Bad Thing? (Bloomberg Law)
3. Villanueva v. Las Vegas Metropolitan Police Department (D. Nev)
4. Villanueva v. Las Vegas Metropolitan Police Department (D. Nev) — Docket Order
5. Conservation Law Foundation v. Shell Oil Co. (D. Conn)
6. Conservation Law Foundation v. Shell Oil Co. (D. Conn) — Docket Order
7. October 2026 Evidence Rules Committee Agenda Book (U.S. Courts)