What Did Your Expert Type into ChatGPT? Courts Are Starting to Ask

Expert witnesses are increasingly using generative artificial intelligence tools to review documents, run analyses, and draft reports. That raises a novel question which courts are beginning to address: are an expert’s AI prompts discoverable? 

A first-of-its-kind ruling. In Conservation Law Foundation, Inc. v. Shell Oil Co., a federal magistrate judge in Connecticut ordered a plaintiff to produce the AI prompts its expert used in preparing her report. The expert, a climate-science historian, had used a commercial AI tool to filter a large set of the defendants' documents down to a workable subset. She disclosed the AI use and some search terms, but the defendants wanted the actual prompts. The court agreed they were entitled to them.

The court reasoned that the prompts were discoverable because an expert's methodology is discoverable. The prompts drove the process that decided which documents the expert reviewed, so the prompts were part of that methodology—no different in principle from the formulas or code an expert routinely must disclose. The court rejected the argument that a stipulation shielding "notes, drafts, or communications" covered the prompts. It appears to be the first federal decision squarely addressing whether an expert's AI inputs are discoverable. One caveat: the order was stayed while the district judge reviews it, so this is not yet the last word.

Why it matters. The ruling ties directly to reliability. Under Daubert and Rule 702, a party must be able to test how an expert reached an opinion. When AI narrows the universe of material an expert considers, the opposing side cannot probe – absent discovery of the expert’s prompts -- whether the tool introduced bias, excluded harmful documents, or hallucinated. Framed that way, the prompts are the audit trail for a methodological choice that shaped the opinion's foundation.

Two cautionary tales. Two recent cases show what is at stake. In Laake v. 3M Company, a Texas jury trial arising from a fatal gas explosion, a defense expert relied heavily on ChatGPT—one prompt reportedly asked the tool to show that 3M bore zero percent of the fault. The AI use surfaced at his deposition, and the prompt history became impeachment material at trial. The jury assigned 3M thirty percent of the fault and returned a verdict exceeding $61 million.

In Ledoux v. Outliers, Inc., a Washington federal court sanctioned a lawyer $3,000 under Rule 11 after she used AI to generate fabricated citations. She had passed an AI-generated citation table to two experts, and both incorporated the fictitious sources into their reports without catching the errors. The court found the conduct tantamount to bad faith.

Practical steps. Litigators should assume an expert's AI prompts and outputs may have to be produced, and plan accordingly. Preserve prompts, session logs, and outputs from the start—missing records can invite a spoliation fight. Address AI in the retainer agreement: require the expert to disclose whether and how AI was used, to preserve the artifacts, and to avoid result-driven prompts like the one that sank 3M's expert. On the other side, serve targeted requests for prompt histories and reference lists. And verify every citation independently before it reaches the court.

None of these decisions binds courts in Georgia or the Eleventh Circuit. They do, however, demonstrate how existing discovery and reliability rules can be invoked to reach AI-assisted expert work. AI is now on every litigator's radar, and cross-examination about it should be expected. An expert who treats an AI session as a disposable scratchpad may learn that the missing record hurts as much as a damaging one would have.

This post is for general information only and is not legal advice. For guidance on a specific situation, please contact our office.