May 18, 2026. District of Connecticut.
Nobody was watching.
No blockchain headlines. No nine-figure sanctions. No viral defendant. Just a quiet Monday in Hartford, and a federal magistrate judge signing a discovery order that landed on exactly zero front pages, because the people who understand why it matters are the ones still reading case dockets at 11 p.m.
The case: an environmental nonprofit suing an oil company over a fuel terminal. The detail that changes everything: the nonprofit’s expert, a Harvard historian and one of the most decorated scientists in her field, had used AI to process Shell’s documents. And opposing counsel wanted the prompts she typed.
Magistrate Judge Thomas O. Farrish said they could have them.
That ruling, ECF No. 970 in Conservation Law Foundation, Inc. v. Shell Oil Co., No. 3:21-cv-00933 (D. Conn.), is believed to be the first decision in the United States requiring an expert witness to produce her AI prompts in discovery.
Sit with that for a moment. The questions an expert typed into a chatbot, the private intellectual scaffolding of how she built her opinion, are now, in at least one federal court’s view, the other side’s business.
The legal system spent years arguing about whether a screenshot of a text message is proper evidence. Now we are arguing about who owns the questions you asked a machine. Progress. (At this rate, by 2035 we will have settled the admissibility of vibes.)
The expert, the tool, and the “secure server”
To understand why this order lands the way it does, you need to know who Dr. Naomi Oreskes is.
She is the Henry Charles Lea Professor of the History of Science at Harvard and an affiliated professor of Earth and Planetary Sciences. She holds a BSc in mining geology from the Royal School of Mines at Imperial College London (1981) and a PhD from Stanford in geological research and history of science (1990). She wrote Merchants of Doubt (2010) and seven other books. She is, by any measure, a geologist and historian of science, not a computer specialist.
That last detail is not a criticism. It is the entire point.
To make her document review workable, Dr. Oreskes and her assistant Dr. Alexander Kaurov used ChatGPT on what the filing describes as a “secure server” to sift through a large volume of Shell’s production documents and identify a relevant subset. The exact number of documents processed is not disclosed in public filings.
Here is the sentence that should make any litigator’s stomach drop: ChatGPT is a cloud-hosted service, meaning confidential litigation materials were transmitted to OpenAI’s infrastructure.
The filing called it a “secure server.” The architecture tells a different story. (Calling a cloud-hosted consumer AI endpoint a “secure server” is a bit like calling a postcard a “sealed envelope.” Technically both are written communications.) We will come back to that.
Three arguments, three rejections
CLF did not hand over the prompts quietly. They pushed back on three grounds. Judge Farrish was unmoved. All three failed.
First, CLF argued that AI prompts fall outside the scope of discovery under Rule 26(b). The court disagreed: prompts are part of the expert’s methodology, and methodology is precisely what Rule 26 reaches.
Second, CLF invoked a Rule 29 agreement it claimed shielded the prompts as drafts or notes. The court rejected this too, holding that such an agreement “must be quite clear” to have that effect, and this one was not.
Third, and this is where the argument collapses in the most avoidable way, CLF contended that Dr. Kaurov had used only “search terms,” not “prompts,” and that search terms had already been produced. The court looked at Dr. Kaurov’s own declaration. He had used the word “prompts.” That gave the court an evidence-backed reason to doubt the distinction CLF was now trying to draw.
(Three arguments walked in. The third one was defeated by the client’s own paperwork. A classic.)
The lesson here: your expert’s declaration should not undermine your expert’s discovery objection. The filing and the argument cannot point in opposite directions.
June: the order is stayed, the fight continues
Compliance was due June 1, 2026.
Two days later, on June 3, CLF filed a Rule 72(a) objection characterizing the ruling as “clearly erroneous,” reprising the same three arguments and adding that no prompts had actually been preserved in the first place. The district court stayed the order pending its ruling on that objection.
As of mid-June 2026, the district judge has not yet ruled.
That ruling, not the magistrate’s, will be the one practitioners remember. A district judge affirming prompt discoverability would carry significantly greater precedential weight. What we have now is a strong signal. What comes next could be binding direction.
Why every tool-dependent expert should be paying attention
This order promotes the computer-engineering expert quietly, almost incidentally, from a peripheral technical witness to a gatekeeper of methodology for every expert who touches AI. Legal commentators have noted the ruling but haven’t fully traced that implication.
Five shifts worth sitting with.
1. Your instrumentation layer is now testimony. A software or computer-engineering expert is more tool-dependent than a forensic accountant or a medical expert. The scripts, SQL queries, grep patterns, forensic suites, filtering criteria, and now LLM prompts are all part of how you reached your conclusion. That instrumentation layer is discoverable methodology. The surface area of what opposing counsel can probe just expanded to include every parameter you set.
2. You are the natural authority on whether the other side’s AI use was sound. Dr. Oreskes is a geologist and historian. She is not the person a court should ask to explain logging, non-determinism, context-window truncation, or retrieval misses. A computer-engineering expert can. A new sub-specialty is taking shape: auditing another expert’s AI methodology, the way a financial expert audits another expert’s valuation model.
3. Reproducibility cuts both ways. Deterministic software is reproducible: same input, same output, provably. Generative AI often is not. Temperature settings, model version updates, and non-determinism mean the same prompt can return different results on different days. You defend the reproducibility of your own methods and attack an opponent’s AI-derived findings on exactly those grounds. This connects directly to the Daubert / Rule 702 reliability requirement: a methodology that cannot be replicated cannot be tested by cross-examination.
4. Confidentiality of evidence is now an engineering question. The phrase “secure server” in CLF’s filings does meaningful work, or tries to. But ChatGPT is cloud-hosted. Whether the instance was a private deployment, an enterprise no-training endpoint, an air-gapped local model, or a consumer web interface changes the privilege and trade-secret analysis entirely. Chain-of-custody now includes data-residency analysis, model-training opt-out settings, retention logs, and whether the model provider’s terms of service allow the data to be used for training. A computer-engineering expert can answer those questions. An environmental historian cannot.
5. AI hygiene is now a preservation duty. If prompts, outputs, and interaction logs are discoverable, they are also subject to litigation-hold obligations. An expert who used AI carelessly and preserved nothing is not merely embarrassed; she is exposed to sanctions, preclusion motions, and adverse-inference instructions. The discipline of reproducible research, versioned inputs, timestamped outputs, model version and settings recorded, applies to litigation work now just as it does to scientific publication.
The Israeli parallel
Israeli court-appointed experts operate under a different procedural framework. The court expert in Israel is appointed by and reports to the court, not retained by a party. But the underlying methodological question is the same. When an Israeli expert uses AI to process a large dataset and the opposing side challenges the methodology, the expert’s documented process is what survives scrutiny. The Israeli framework’s emphasis on independence and accountability to the court makes undocumented AI use even more vulnerable to attack: there is no party-adversary dynamic to absorb the hit. The expert stands alone.
The CLF ruling is a preview of that conversation arriving in Israeli courtrooms.
The takeaway
A magistrate judge in Connecticut ordered an expert to produce her AI prompts, and the three-part argument against it failed completely. The district court’s ruling on the objection is the one to watch, but the direction of travel is already clear.
Document your tools. Preserve your prompts. Record your model version and settings. Treat your instrumentation layer as discoverable from day one. And when the other side’s expert used AI to reach her conclusions, ask for the prompts.
You now have standing.
The above is general information only and does not constitute legal advice. Specific facts of the case are drawn from the sources cited.