When an expert uses generative AI: verification, prompt records, and discovery risk

Generative AI can help an expert search, sort, classify, summarize, or identify material for closer review. That does not relieve the expert of the obligation to exercise independent judgment. The litigation question is usually not whether a tool appeared in the workflow. It is whether the opinion remains traceable, verified, and explainable on the record.

Where AI use enters an expert workflow

AI-assisted work may occur before an expert reaches a final opinion. A tool might narrow a document collection, organize source material, identify recurring themes, draft code, suggest research leads, or produce a provisional summary. Each use has a different evidentiary consequence. The more the tool shapes source selection, analysis, or a stated conclusion, the more likely counsel and an opposing expert will examine the workflow.

The practical record includes the system and version used, the material supplied to it, meaningful prompt or query instructions, outputs that informed later work, the expert’s verification steps, and the limits the expert recognized. Not every transient interaction will be material. A disciplined workflow identifies what actually informed the opinion and preserves what is needed to explain it.

The verification lesson from Kohls

In Kohls v. Ellison, the District of Minnesota declined to consider an expert declaration at the preliminary-injunction stage after the declaration included nonexistent academic citations generated by GPT-4o and not verified by the expert. The court distinguished AI use for research from an expert’s failure to check the work presented to the court. The order is a direct reminder that citation verification and independent professional judgment remain central to credibility.¹

The case should not be read as a categorical bar on AI-assisted expert work. It shows why an expert’s final opinion must rest on reviewed sources, articulated reasoning, and a method the expert can defend without asking the court to trust an unverified output.

Prompt records and methodology discovery

A 2026 discovery order in Conservation Law Foundation, Inc. v. Shell Oil Co. treated prompts used by an expert to cull document productions as part of the expert’s methodology. The order was stayed while an objection remained pending, so it is not a final nationwide rule. It is nevertheless a useful signal that prompts, search terms, outputs, and AI-assisted selection criteria may become discovery issues when they inform an expert’s work.²

Engagement terms, litigation-hold instructions, discovery protocols, and expert-report preparation should address AI use with specificity. A generic reference to notes or drafts may not resolve disputes about tool interactions, system instructions, or query history.

Rule 707 remains a proposal

Proposed Federal Rule of Evidence 707 has focused attention on how courts might assess machine-generated evidence. It is not an operative federal evidence rule. The Judiciary’s August 2026 public-comment package lists proposed Evidence Rules 104 and 902, rather than Rule 707, and the current framework remains the existing rules governing expert testimony, authentication, discovery, and case-specific reliability.³

What a defensible workflow looks like

A defensible workflow does not require an expert to preserve every exploratory thought. It does require clarity about the process that produced the opinion. The expert should be able to identify the relevant source materials, describe the tool’s role, explain the validation performed, and state whether an AI output was merely a lead, a convenience, or part of the analysis itself.

That distinction matters when counsel evaluates admissibility, prepares an expert for deposition, drafts discovery requests, or assesses whether a technical issue requires a specialist in machine learning, digital forensics, or evidence authentication.