Rule 26 Expert Discovery: When an Expert Witness Uses AI, Is It Discoverable?

Case law following the 2010 amendments to Federal Rule of Civil Procedure 26 offers guidance on how courts are likely to address whether an expert witness’s AI records are subject to discovery.

Key Takeaways

• The 2010 amendments set the boundaries of Rule 26 expert discovery. Draft reports and most attorney-expert communications are protected, but the “facts or data considered” by the expert are not, and courts read that phrase broadly.
• An expert’s AI prompts, uploads, and responses can be discoverable. Each can contain “factual ingredients,” and a prompt used to filter a production can function like a search protocol, as the Conservation Law Foundation court treated it.
• An AI-generated report may still be a protected draft. Under McCoy, Taction Technology and Manning, the question is whether the expert substantially participated, or whether AI supplied the facts, analysis, or conclusion the expert later adopted.

Rule 26 Expert Discovery After the 2010 Amendments

This article addresses the question courts and lawyers have faced in recent cases: what happens when an expert witness discloses that they used AI in their reports, as discussed in Mr. Seelinger’s earlier Firm Insight, How Expert Witness AI Use Ends Up in Discovery.

Any answer about discovery obligations and protections for an expert witness must start with the 2010 amendments to Federal Rule of Civil Procedure 26. Per the Advisory Committee’s notes to those amendments: “routine discovery into attorney-expert communications and draft reports has had undesirable effects.” Namely, it found that the prior rule drove costs up and drove communications underground. Three of those amendments are particularly relevant here:

• Rule 26(a)(2)(B)(ii) was amended to narrow the required disclosure of the “facts or data considered” (previously it was “data or other information”) in a testifying expert’s report;
• Rule 26(b)(4)(B) was added, which protects “drafts of any report or disclosure required under Rule 26(a)(2), regardless of the form”; and
• Rule 26(b)(4)(C) separately protects most attorney-expert communications, subject to exceptions for compensation, facts or data counsel supplied and the expert considered, and assumptions counsel supplied and the expert relied upon.

As the Tenth Circuit observed in The Republic of Ecuador v. Bjorkman, “[o]ther comments make clear that the protections of Rules 26(b)(4)(B) and (C) are the exclusive protections afforded to expert trial-preparation materials.” The Eleventh Circuit also made that point in Republic of Ecuador v. Hinchee: Rule 26(b)(3)’s general work-product provision does not protect a testifying expert’s own notes and non-attorney communications simply because they were prepared for litigation. Other circuits and districts have echoed the Advisory Committee’s note; the 2010 protections do not block discovery into the development, foundation, or basis of an expert’s opinions.

To determine whether an expert’s AI use is discoverable, those first two amendments take on heightened significance. Each is discussed in turn.

1. “Facts or Data Considered” Is Defined Broadly

The 2010 amendment to Rule 26(a)(2)(B)(ii) narrowed the old phrase “data or other information” to “facts or data considered,” primarily to protect counsel’s theories and mental impressions. It did not narrow the expert’s factual disclosure obligation to materials ultimately cited or consciously relied upon. The Advisory Committee instructed courts to read “facts or data” broadly to include material from any source containing “factual ingredients,” so long as the expert considered it.

That distinction is often dispositive. Wenk v. O’Reilly held that an expert’s independent notes were not protected merely because they recorded the expert’s observations and analysis; those notes contained factual ingredients. More recent cases likewise apply an objective understanding of “considered,” commonly reaching material received, reviewed, read, or authored in connection with forming the opinion, even if the expert later says it was not relied upon. The party resisting disclosure generally bears the burden of showing the material was not considered.

An AI exchange can fit that rule in several ways. Uploaded records are factual inputs. A response that summarizes evidence, identifies technical literature, performs calculations, or supplies a factual proposition may itself contain factual ingredients. A prompt may embed selected facts or assumptions. And a prompt used to filter or rank a production may function like a search protocol or model parameter: it shapes the universe of facts presented to the expert. That was the basis of the Conservation Law Foundation order mentioned in Mr. Seelinger’s earlier article.

Accordingly, lawyers seeking production of an opposing expert’s prompts, responses, logs, and session history on AI platforms should have a sophisticated understanding of large language models and the various environments and workflows in which they are deployed. The difference between those that understand AI and those that don’t will play out in countless depositions in the coming years. Having a lawyer who can intelligently probe an expert’s use of AI, elicit key admissions, articulate the practical significance of those admissions, and offer a suitably tailored production demand for that expert’s AI use will all be decisive factors in obtaining that sort of essential discovery into the “facts or data” considered in their report.

2. An AI-Generated Report Could Be a Protected Draft

Rule 26 does not define “draft,” and courts acknowledge that the boundary remains underdeveloped. The better-reasoned decisions look to function and circumstances: was the material created as a preliminary version of all or part of the report from which the final was made?

In Deangelis v. Corzine, for example, a consultant’s write-up and chart were protected where the evidence showed they were requested for, inserted into, and revised as part of the report. Other courts have protected draft worksheets and rough analyses prepared by an expert or the expert’s assistants for use in a report. The protection is not limited to a complete, expert-typed manuscript.

That reasoning can accommodate AI-assisted drafting. McCoy v. GEICO Indemnity Co. is a recent District of New Jersey case with analysis that could guide future decisions on draft protection for an expert’s AI-generated versions of a report. There, the court reiterated that Rule 26’s “prepared and signed” requirement does not demand that the expert personally put fingers to keyboard; counsel may reduce the expert’s substantive views to writing if the report genuinely reflects those views. The reasoning follows that if someone else can write the report based upon the expert’s substantive views, then an AI-generated report should qualify as a draft too under appropriate circumstances.

That focus on “substantive views” tracks the decision in Taction Technology, Inc. v. Apple Inc., which required “substantial participation” by the expert for a report to qualify as a protected draft. That language stems from a Northern District of Illinois decision that inspired others on this issue; although that case is decades old, its analysis is more relevant than ever in the AI era:

In contrast, preparing the expert’s opinion from whole cloth and then asking the expert to sign it if they wish to adopt it conflicts with Rule 26(a)(2)(B)’s requirement that the expert “prepare” the report. Preparation implies involvement other than perusing a report drafted by someone else and signing one’s name at the bottom to signify agreement.

Manning v. Crockett. Under the rationale of these cases, an AI-generated report or section of a report that is created from the expert’s own analysis and substantial participation has a reasonable claim to protection from discovery as a draft report.

Regardless, this issue will remain exceptionally sensitive to how the expert chose to use AI. Exploratory questions, factual summaries, document-selection queries, test results, calculations, and notes that might or might not later be used are relatively less likely to qualify for protection under Rule 26(b)(4)(B). That protection also does not negate the expert’s independent obligation to disclose the facts or data considered. The distinction turns on whether AI only served as a scrivener for an opinion the expert had already developed or supplied the facts, analysis, or conclusion the expert later adopted.

Frequently Asked Questions

Are an expert witness’s AI prompts and chat logs discoverable under Rule 26?

They can be. Rule 26(a)(2)(B)(ii) requires disclosure of the “facts or data considered” by the expert, which the Advisory Committee instructed courts to read broadly to reach material from any source that contains “factual ingredients.” Uploaded records, AI responses that summarize evidence or perform calculations, and prompts that embed facts or filter a production can all fit that rule, as the Conservation Law Foundation order shows.

Is an AI-generated expert report protected as a draft?

It may be. Rule 26(b)(4)(B) protects drafts “regardless of the form,” and courts such as McCoy v. GEICO have held that the expert need not personally type the report so long as it genuinely reflects the expert’s substantive views. Under Taction Technology and Manning v. Crockett, the expert must have substantially participated; a report prepared “from whole cloth” and signed does not qualify.

What did the 2010 amendments to Rule 26 change for expert discovery?

Three changes matter here: Rule 26(a)(2)(B)(ii) narrowed the required disclosure from “data or other information” to “facts or data considered”; Rule 26(b)(4)(B) added protection for draft reports; and Rule 26(b)(4)(C) protected most attorney-expert communications, with exceptions for compensation and for facts, data, and assumptions counsel supplied. The Tenth and Eleventh Circuits have held those are the exclusive protections for expert trial-preparation materials.

Secure the Expert’s AI Workflow Before the First Prompt

Understanding and securing your expert’s AI workflow should begin before the first prompt, not after a motion to compel. Likewise, uncovering discoverable AI evidence from an opposing expert benefits from understanding how large language models are deployed. Lawyers who understand both the legal and technical aspects of expert work can build a defensible workflow, protect legitimate work product, and test an adversary’s AI use with precision.

If Rule 26 expert discovery is on the horizon in your case, Chris Seelinger and Scarinci Hollenbeck’s Litigation attorneys can help. Schedule a consultation to discuss how your expert’s AI use, or your adversary’s, could affect your case.