Federal Court Splits On Disclosure Of Expert Witness AI Prompts

In June, we reported on Conservation Law Foundation, Inc. v. Shell Oil Company, et al. (Case No. 3:21-cv-00933, D. Conn.), in which Magistrate Judge Thomas Farrish ordered the plaintiff to produce the generative AI prompts its expert witness used in preparing her expert report. That decision appeared to be the first federal ruling to treat an expert’s AI prompts as discoverable methodology.¹ A new decision from the District of Nevada goes the other way on prompts, holding that an expert’s questions and prompts to a Gen AI platform are protected from disclosure, but that AI-generated case summaries the expert considered must be produced.

The Nevada Decision

On September 29, 2026, District Judge Anne R. Traum issued an order in Villanueva v. Las Vegas Metropolitan Police Department, et al. (Case No. 2:24-cv-00125-ART-DJA, D. Nev.) (ECF No. 112). The case is a civil rights action brought under 42 U.S.C. § 1983 and state law, arising from the plaintiff’s detention at the Clark County Detention Center.

The plaintiff’s expert, Tom Melton, testified at his October 1, 2025 deposition that he used a Gen AI platform to help research and draft his expert report. The plaintiff had not disclosed this AI use in response to an earlier subpoena for all materials the expert considered in forming his opinions. After the deposition, the plaintiff refused to produce the Gen AI platform’s history log and listed it as attorney/expert work product on its privilege log.

Ruling from the bench, Magistrate Judge Albregts granted the defendants’ motion to compel production of the entire history log used with the Gen AI platform. He reasoned that it should have been disclosed under Federal Rules of Civil Procedure 26(a)(2)(B)(ii) and (iii) because the expert “considered it in forming and supporting his opinions.” He also found that it was neither work product nor privileged. The plaintiff described the 40-page log as containing three categories of material: (a) case summaries of controlling excessive-force decisions; (b) a Gen AI platform-generated deposition outline used to predict likely deposition questions; and (c) prompts to analyze issues and draft portions of the report, such as the opposing expert’s qualifications.

On the plaintiff’s objection, Judge Traum held that ordering production of the entire log was “clearly erroneous or contrary to law.” She granted the objection in part and overruled it in part.

AI-generated case summaries must be produced. The court applied the Ninth Circuit’s broad reading of “facts or data considered” in Republic of Ecuador v. Mackay, 742 F.3d 860, 869 (9th Cir. 2014). Under that reading, the rule covers “any material considered by the expert, from whatever source, that contains factual ingredients,” and not only material upon which the expert relied. Relying on Hirata v. S. Nevada Health Dist., 2016 WL 10706085 (D. Nev. 2016), which required disclosure of deposition summaries that an expert consulted only as a quick-reference guide, the court held that the expert “consulted” the Generative AI case summaries to reach his opinion. Those summaries were therefore discoverable “facts or data.” The court also noted that the log is the only place the defendants can look for alleged “hallucinated case citations…quotes…and references to broad legal standards” the expert relied on or considered.

Prompts and questions are protected. The court found that a question “is not something that an expert can ‘consider’ or ‘rely’ upon as ‘facts or data’ in reaching their opinion, just as an expert would not ‘consider’ or ‘rely’ upon their online research queries.” It further held that the prompts would reveal the expert’s “mental impressions or strategy” and theory of the case. Discovery into them would therefore be impermissible whether they are treated as draft-report material under Rule 26(b)(4)(B) or as trial preparation materials under Rule 26(b)(3).

The deposition outline is protected. The court held that the AI-generated deposition outline was not “facts or data,” because the expert used it to prepare for his deposition and for trial, not to form his opinion. Production was barred under Rule 26(b)(3)(B) in any event, because the outline would reveal the expert’s “focus, priorities, or areas of concern.”

AI-assisted drafts are protected. Rule 26(b)(4)(B) protects expert drafts “regardless of the form in which the draft is recorded.” Because the expert testified that he sometimes copied and pasted outputs from the Gen AI platform into his report, the court treated those outputs as protected draft material. The court added that, at minimum, the outputs were protected mental impressions.

The court made two further rulings. First, after its own in camera review, it held that the magistrate judge did not clearly err by relying on the plaintiff’s good-faith description of the log. Second, and significantly, it held that using Generative AI does not waive privilege, because it “is not a person, but was an application used in this instance as a research, drafting, and editing program.” The court also noted that many of the defendants’ concerns about the expert’s methodology and reliance on Generative AI can be addressed at trial through impeachment.

How Villanueva Differs from Conservation Law Foundation

The two decisions may be reconciled in part by how the experts used AI. In Conservation Law Foundation, the AI platform shaped which documents the expert reviewed, a process the court called “an aspect of that methodology.” In Villanueva, the court separated outputs that fed into the expert’s opinion from inputs and outputs that reflected the expert’s thinking or trial preparation. Even so, the courts disagree squarely on prompts. Villanueva compares prompts to online research queries, which are not discoverable. Conservation Law Foundation treats them like dataset selection criteria, a long-accepted subject of discovery.

Key Takeaways

• The law is unsettled. Two federal courts have now reached different conclusions on whether expert AI prompts are discoverable. Practitioners should expect jurisdiction-specific and fact-specific outcomes.
• Outputs that inform opinions are at the highest risk. Under Villanueva, AI-generated content an expert “consults” in forming an opinion, such as legal or factual summaries, is likely discoverable even if the expert only used it as a reference.
• Separate AI uses by category. The Villanueva court sorted the expert’s AI use into research, deposition preparation, and drafting. Experts and counsel should keep these uses distinct, and keep records that support privilege and draft-protection arguments.
• Expect impeachment. Even protected AI use may be explored at trial, including the risk of hallucinated citations.
• Our earlier recommendations still apply. Set protocols for expert AI use, address AI in retention agreements and Rule 29 stipulations, and raise AI discovery issues at the Rule 26(f) conference.

We will continue to monitor the pending objection in Conservation Law Foundation and other developments in this fast-moving area.

¹ The court is scheduled to hear oral argument on an objection to this order on November 3, 2026.