When an expert witness types a prompt into a generative AI tool to filter, rank, or surface documents before forming an opinion, that prompt is part of the expert's methodology. And methodology, under the Federal Rules of Civil Procedure, is discoverable. That is the unmistakable message from Connecticut Magistrate Judge Thomas O. Farrish's May 18, 2026 ruling in Conservation Law Foundation v. Shell Oil, No. 3:21-cv-00933 (D. Conn.).
The decision is narrow in its facts but sweeping in its implications. It is the first federal court ruling to treat AI-assisted document culling as a component of expert methodology subject to Rule 26 disclosure, and it arrives at a moment when generative AI tools are quietly embedded in the workflows of consultants, scientists, and technical experts across virtually every area of complex litigation.
One procedural note bears on how much weight to give it: the magistrate judge's order has been stayed pending the district court's review of the plaintiff's Rule 72(a) objection, so it is not yet a final ruling. But its reasoning is already shaping how litigants approach expert AI use — and the underlying logic is unlikely to disappear regardless of how the objection is resolved.
What the Court Decided — and Why It Matters
The dispute arose when the Conservation Law Foundation sought production of AI prompts that the opposing party's expert had used to identify and winnow documents before preparing an opinion. The expert's proponents argued that the parties' Rule 29 stipulation — an agreement modifying certain discovery procedures — shielded those prompts from disclosure. Judge Farrish rejected that argument on two independent grounds.
First, the court found the Rule 29 agreement ambiguous with respect to AI-generated materials, and held that ambiguity in a discovery-limiting agreement cannot carry the weight of excluding otherwise discoverable material. The parties had not specifically addressed AI prompt logs when they negotiated the stipulation, and the court declined to read that silence as protection.
Second, and more fundamentally, the court reasoned that when an expert uses AI to cull or identify documents that will inform an opinion, those prompts are functionally inseparable from the expert's analytical process. Rule 26(a)(2)(B) requires disclosure of "the facts or data considered by the witness" and "any exhibits" to be used, and Rule 26(b)(4) governs the scope of expert discovery. Treating the AI prompts as mere technical logistics — rather than as inputs shaping which evidence the expert ever considered — would allow a significant methodological step to escape scrutiny entirely.
The Methodology Problem Generative AI Creates
Traditional expert discovery assumes that a human expert reviews a defined universe of materials and then applies professional judgment to reach conclusions. The adversary can probe that judgment through deposition and cross-examination. Generative AI disrupts that model in a way courts are only beginning to confront.
When an expert delegates initial document identification to a large language model, the prompt becomes the first filter between the full evidentiary record and the expert's actual analysis. A poorly constructed or subtly leading prompt can systematically exclude relevant documents without the expert — or opposing counsel — ever knowing. The Conservation Law Foundation ruling recognizes this dynamic. If opposing counsel cannot see the prompts, they cannot evaluate whether the AI's output was complete, unbiased, or fit for the expert's stated purpose.
This is not a theoretical concern. Prompt design choices — word selection, framing, inclusion or exclusion of context — materially affect what a large language model returns. An expert who prompts a model to find documents "supporting" a particular conclusion will receive a different document set than one who prompts it to find all documents "relevant to" that conclusion. That difference is methodologically significant, and Conservation Law Foundation says it is now discoverable.
Immediate Practice Implications for Counsel
The ruling creates concrete obligations that litigators must address before discovery disputes materialize, not after. Several steps warrant immediate attention.
Audit expert AI workflows now. Retaining counsel should ask testifying experts — at the time of engagement, not at the eve of disclosure — whether they are using any AI tool to assist with document review, research, or analysis. That question must be specific enough to capture large language models, AI-assisted review platforms, and embedding-based search tools, not just traditional keyword search.
Revisit ESI protocols and protective orders. Existing ESI stipulations and protective orders almost certainly do not address AI prompt logs as a category of discoverable material. Counsel should negotiate explicit language before the issue becomes contested. Where prompts contain legitimately sensitive information — such as attorney mental impressions embedded in a prompt drafted with counsel's input — a targeted Rule 26(b)(3) analysis and appropriate protective order language should be in place before production is demanded.
Document the AI tool's parameters. The version of the model used, the date of use, and any system-level instructions or retrieval-augmented generation configurations may all bear on the reliability of the expert's methodology. Preserving that metadata from the outset is far less burdensome than reconstructing it under a discovery order.
How Law & Forensics Helps
Law & Forensics works with litigation teams and in-house counsel to evaluate how AI tools are integrated into expert workflows, assess whether those integrations create discoverable artifacts under Rule 26, and develop ESI protocols and expert engagement procedures that address generative AI from day one. Our technical and legal professionals can help audit existing expert methodologies, provide independent expert testimony on AI tool reliability and prompt design, and assist in drafting protective order language tailored to the evidentiary and competitive sensitivities involved. When AI becomes part of the expert's method, it becomes part of the litigation risk — and we help you manage it before it becomes a problem the court has to solve.