Magistrate Judge Orders Production Of AI Prompts Used By Expert Witness

Key Takeaways

This ruling may open the door to broader discovery of expert AI usage. Parties should consider addressing AI usage explicitly in their discovery protocols and agreements rather than relying on general provisions that courts may decline to extend to novel AI-assisted workflows. 

In an environmental enforcement action between the Conservation Law Foundation (“CLF”) and Shell Oil Co. and its affiliated entities (“Shell”), Magistrate Judge Thomas O. Farrish ordered CLF to produce the AI prompts its expert witness used in preparing her report. See Conservation Law Found., Inc. v. Shell Oil Co., No. 3:21-cv-00933, ECF No. 970 (D. Conn. May 18, 2026). The ruling appears to be the first federal court decision requiring an expert witness to disclose AI prompts.

CLF’s expert used OpenAI’s GPT-4o, accessed through a private Microsoft Azure server, to upload Shell’s documents produced in discovery and filter them with AI prompts into a relevant subset. Shell moved to compel production of those prompts. The magistrate judge granted the motion, rejecting each of CLF’s three objections.

First, CLF argued that the prompts fall outside the scope of discovery under Rule 26(b)(2). The court disagreed, finding that an expert witness’s methodology is “fair ground for discovery” and that the AI-assisted filtering of Shell’s documents was part of that methodology. Id.

Second, CLF argued the prompts qualified as “expert notes” excluded from discovery under the parties’ Rule 29 agreement. The court disagreed, finding the agreement insufficiently clear to cover AI prompts as “expert notes” and declining to extend it to exclude otherwise relevant discovery. Id.

Third, CLF argued it had already produced all responsive materials to this request because its expert used “search terms,” not AI “prompts.” The court rejected this argument because the expert’s own research assistant had used the word “prompts” in a sworn declaration, giving Shell an adequate basis to question CLF’s representation. Id.

The court ordered CLF to amend its discovery responses to disclose any AI prompts or queries used by the expert, or state in its sworn amended discovery responses that no such material exists, cautioning that Rule 37(b) sanctions could follow if any such representation later proved untrue.

CLF filed an emergency motion to stay the ruling and a Rule 72(a) objection, arguing the magistrate judge’s ruling was clearly erroneous or contrary to law, the standard required to overturn a magistrate judge’s discovery ruling. The district court stayed the order pending resolution of CLF’s objection.

In response, Shell argued that CLF had not met its burden to overturn Judge Farrish’s order. Shell contended that the magistrate judge straightforwardly applied established expert discovery rules in holding that the expert’s AI-assisted review constitutes discoverable methodology under Rule 26 and correctly concluded that the Rule 29 stipulation lacked sufficiently clear language to cover materials of this kind.

Parties and their counsel should be aware of these issues as they continue to arise and are ruled upon by courts across the country.