The Emerging Framework on AI Prompts in Federal Discovery (2026)

On May 18, 2026, Magistrate Judge Thomas O. Farrish of the U.S. District Court for the District of Connecticut granted defendants' motion to compel reliance materials in Conservation Law Foundation, Inc. v. Shell Oil Co., No. 3:21-cv-00933 (VDO) (Judge Vernon D. Oliver presiding; Farrish, M.J.), ECF No. 970. The operative order required CLF to revise its Rule 33 and Rule 34 responses to call for "artificial intelligence prompts and/or queries" used by the plaintiff's expert Dr. Naomi Oreskes and her team in producing her expert report. The order is reported as the first of its kind in federal litigation. As of June 3, 2026, the ruling is stayed pending CLF's Rule 72(a) objection to the district judge.

The case itself is environmental. Conservation Law Foundation sued Shell for alleged Clean Water Act and climate-resilience violations at a Connecticut petroleum terminal. The expert in question is the historian Naomi Oreskes. None of that is patent-specific. What made the order important for patent practitioners is the legal framing: the magistrate treated AI prompts as part of an expert's methodology under Federal Rule of Civil Procedure 26(b), not as protected work product and not as the equivalent of search terms. That framing, if it survives Rule 72(a) review, applies the same way to a patent expert who uses generative AI to filter prior art, summarize a technical reference, or build a noninfringement opinion.

Beyond CLF v. Shell, an emerging framework is taking shape across federal courts on when AI prompts are discoverable, when they are shielded as work product, and how protective orders should be structured. Arnold & Porter's June 2026 survey ("The Emerging Framework on AI Prompts, Privilege, and Discovery") and adjacent practitioner commentary draw distinctions that matter more for patent practice than any single magistrate's order. This page covers both: the CLF order itself, and the framework it sits inside.

This article presents published case reporting and law-firm analysis for educational purposes. It is not legal advice. Consult a registered patent attorney or trial counsel for litigation decisions.

What happened, in short

Dr. Oreskes (with the assistance of Dr. Alexander Kaurov) used OpenAI's ChatGPT, hosted on a private Azure server, to filter and categorize Shell's document production. The objective was to surface the materials most relevant to her expert opinion. Shell moved to compel reliance materials and pressed for the prompts and queries Dr. Oreskes fed to the model, along with related technical details about how the server was configured.

CLF opposed, framing the prompts as the rough equivalent of search terms or attorney work product. The magistrate rejected both arguments. Judge Farrish ruled that the prompts were part of the methodology by which the expert formed her opinion and fell within the scope of expert discovery under Rule 26(b). A pre-existing Rule 29 stipulation between the parties limiting discovery of "expert notes, drafts, or communications" was held to be "not quite clear" enough to cover AI prompts. The order required CLF to revise its Rule 33 and Rule 34 responses to call for "artificial intelligence prompts and/or queries" used by the expert team.

What the order is, more precisely: a directive to amend interrogatory and document-request responses so that they call for the AI prompts the expert used. It is not, on its face, a blanket order to produce all outputs, all transcripts, or all server configuration details. Reporting from major firms varies on how broadly to read the implications for production of model outputs and configuration; the conservative read is that production of prompts is what the order required, and that the broader exposure flows from the amended responses CLF must serve.

The magistrate did not engage at length with the work-product doctrine, because the Rule 29 stipulation was the vehicle plaintiff had relied on. That leaves open the question of how work-product analysis would apply to AI prompts in a case without a similar stipulation. Practitioners should not assume work product will provide automatic shelter on the next motion to compel.

Why "prompts are search terms" failed

Plaintiff's strongest argument was analogical: prompts to a large language model are functionally similar to search terms in an electronic discovery search. Courts have generally protected the iteration of search terms as work product, on the reasoning that they reflect attorney thought processes about what is relevant.

The court rejected the analogy for two reasons that practitioners should internalize:

1. The model is not a Boolean filter. A search term retrieves documents that match. A prompt instructs a model to interpret, summarize, categorize, or rank. The prompt embeds judgment, framing, and instructional choices that materially shape the output. That is closer to methodology than to retrieval.
2. The output is not deterministic in the way search results are. Two runs of the same prompt against the same model can produce different outputs. The prompt is therefore not just the question; it is part of the experimental setup that produced the expert's findings.

Under Daubert, an expert's methodology has to be testable. The court reasoned, in effect, that if the prompts are not produced, opposing counsel cannot cross-examine the methodology. That cuts to the heart of expert discovery.

What the order actually requires

Reading the order narrowly:

• Amended Rule 33 and Rule 34 responses calling for the AI prompts and queries the expert team used in producing the report. CLF must revise its existing interrogatory and document-request responses to no longer treat AI prompts as outside the scope of the requests.
• The prompts themselves are within the scope of what the amended responses must call for. Shell can then seek production through ordinary discovery mechanisms once the responses are revised.

Reading what the order does not clearly require:

• Blanket production of all model outputs and transcripts. The opinion-reliance materials Dr. Oreskes used remain governed by Rule 26(a)(2). The order's primary mechanism is amending the discovery responses, not directly compelling output production.
• Production of training data, model weights, or other proprietary information about the underlying AI system, since OpenAI was not a party.
• A broad rule that all uses of AI by experts must be disclosed in all cases. The order is rooted in the specific Rule 29 stipulation's silence on AI and the specific Rule 26(b) reasoning about methodology.

Read narrowly, this is an expert-discovery dispute decided on the language of one stipulation, requiring an amendment to discovery responses. Read broadly, it is the first federal magistrate to treat AI prompts as discoverable methodology and reject the analogy to work-product-protected search terms. Both readings are defensible. Patent litigators should plan for the broader reading, while acknowledging that the order itself is narrower and is stayed pending Rule 72(a) review.

The emerging framework: not all prompts are treated alike

The most important practitioner-side reading of CLF v. Shell is that it is one piece of a larger picture, not a standalone rule. Arnold & Porter's June 2026 survey, "The Emerging Framework on AI Prompts, Privilege, and Discovery," synthesizes what federal courts are doing across multiple matters. The synthesis matters because the framework distinguishes between categories of AI prompts that have been treated differently:

Distinctions drawn from Arnold & Porter, "The Emerging Framework on AI Prompts, Privilege, and Discovery," eData Edge (June 2026), with cross-reference to Hogan Lovells and Kirkland & Ellis treatments of the same emerging framework.

The practitioner takeaway is that "AI prompts are discoverable" is not a uniform rule. The doctrinal analysis turns on who used the prompt, in what role, and for what purpose. Patent practice intersects this framework at multiple points (litigation counsel, prosecution counsel, testifying experts, in-house teams), and the exposure profile differs at each point.

How the reasoning ports to patent litigation

Patent experts use generative AI in workflows that look superficially different from filtering document productions, but functionally identical for Rule 26(b) purposes. A few examples:

• Prior-art search and triage. A patent expert using a generative AI tool to summarize a candidate reference, compare it against the asserted claims, or rank a corpus of references by relevance to a specific limitation. The prompts that produced the summaries, comparisons, or rankings are methodology under the CLF v. Shell framing.
• Claim construction analysis. An expert who prompts a model to identify how a term of art was used in the relevant field at the time of invention. The prompt frames the question and the corpus, and shapes the answer. It is discoverable methodology.
• Damages model construction. An economist or accounting expert using generative AI to synthesize industry comparables, parse SEC filings, or build the underlying inputs for a reasonable royalty analysis. The prompts shape what data the model retrieved and how it was summarized.
• Noninfringement opinion drafting. An expert who uses generative AI to draft sections of a noninfringement analysis or to compare accused product behavior against asserted claim limitations. The prompts that produced the drafts shape the opinion.

In each scenario, the same Rule 26(b)(4) reasoning applies: the prompt is part of the methodology that produced the opinion. Opposing counsel can therefore plausibly compel production. Patent litigators should expect to see motions to compel AI prompts in expert discovery over the next 12 to 24 months, and to defend their own experts' use of these tools.

What patent practitioners should be doing now

The order is not a rule of decision binding any other court. It is, however, a credible signal of how federal courts will treat AI prompts in expert discovery. The practical response cleaves into two layers: how experts use AI, and how the underlying patent record was built in the first place.

Litigation-side: expert engagement and document holds

• Litigation holds now have to cover AI prompts and outputs. Standard litigation-hold language addressed to experts should explicitly require preservation of prompts, queries, system prompts, model identifiers, dates, and (where retained) outputs. A hold notice silent on AI may not survive a spoliation challenge.
• Engagement letters with testifying experts should specify AI use. Whether the expert may use generative AI, on what platform, with what data-retention policy, with what logging, and with what disclosure obligations to retaining counsel. Some firms are moving toward prohibiting consumer-grade GAI use entirely for testifying experts.
• Rule 29 stipulations should address AI prompts explicitly. The CLF case turned on a stipulation that was silent on AI. Future stipulations will be drafted with the question front of mind. Counsel for the patent owner and for the accused infringer have different incentives here, and the negotiation matters.
• Audit trails matter more than they did 12 months ago. Enterprise GAI platforms that log every prompt, every output, and every system configuration setting are substantially easier to defend in discovery than consumer-grade tools where prompt history is local, ephemeral, or shared with the model provider.

Prosecution-side: what the patent record looks like at filing

The CLF v. Shell order is an expert-discovery order, but the underlying tension it surfaces extends backward to the prosecution stage. If generative AI was used to draft the application, build the specification, or surface the prior art that informed the claims, those workflow choices become part of the broader litigation record. They feed into validity arguments, inventorship challenges, and (per Baker Donelson and Hogan Lovells analyses) potentially privilege waivers.

The practical takeaway: the more the patent record can stand on its own technical foundation, anchored to specific code, documents, or measured behavior in the inventor's actual work product, the less the AI-assisted prosecution workflow becomes the story in litigation. See our AI Patent Discovery & Privilege guide for how structured-disclosure tools with deterministic audit trails differ from generic GAI prompt logs.

The work-product question this order did not decide

Federal Rule of Civil Procedure 26(b)(3) protects documents prepared in anticipation of litigation. Opinion work product (mental impressions, conclusions, legal theories) enjoys near-absolute protection; fact work product can be discovered on a showing of substantial need.

AI prompts arguably fit both categories. The phrasing of a prompt reflects counsel's or the expert's mental impressions about what is relevant. The output is fact work product. The CLF v. Shell order did not engage with that doctrinal analysis at length because the Rule 29 stipulation was the threshold question and the stipulation did not cover AI. A future case without a comparable stipulation, where the privilege holder asserts work-product protection cleanly, will be the test.

Until that case is decided, the prudent posture is to assume that prompts may not be shielded as work product, and to design expert workflows accordingly.

The structured-disclosure alternative

One reason this order matters for patent practitioners is that it sharpens a distinction patent-tech vendors have been making informally for the past year: the difference between generic generative AI (where prompts are the methodology, retention is ad hoc, and the output is non-deterministic) and structured disclosure tooling (where the workflow is deterministic, the artifact is the codebase or the technical record itself, and the audit trail is built in by design).

For patent prosecution: if the technical evidence in the specification was identified by scanning an inventor's actual codebase (not by prompting a model to imagine what an invention might be), then the discovery question becomes "what files in the codebase were scanned, at what commit?" That is a categorically different audit trail than "what prompts did counsel feed to ChatGPT?". The first question has a deterministic answer rooted in artifacts the inventor controls. The second question has the audit-trail problem CLF v. Shell just put on the litigation map.

ObviouslyNot's scanner is built on the first model, not the second. See how the scanner works or our discovery and privilege guide for how the audit trails differ.

What we are not claiming

• This is not a patent case. The order arose in environmental litigation. The reasoning is portable to patent expert discovery, but no patent court has yet adopted it.
• This is not binding precedent. A magistrate's discovery order in one case is not a rule of decision for any other case, and this particular order is stayed pending Rule 72(a) review. It is a signal of how federal courts are likely to reason, not a guarantee, and the signal could change if the district judge sustains CLF's objection.
• AI prompts are not uniformly discoverable. The Arnold & Porter framework above distinguishes between attorney litigation prompts (often opinion work product), expert prompts (methodology under Rule 26(b)), party-side prompts (fact-specific), and consumer-AI use under protective orders (emerging dispute area). Practitioners should not read CLF v. Shell as a categorical rule.
• This does not mean AI use is prohibited. Experts can and should use modern tools. The implication is that they should use them with discovery in mind, not that they should stop using them.
• Work-product protection is not foreclosed. The magistrate did not reach the doctrine. A future case might extend more protection than this order suggests. Until then, plan conservatively.

Sources

Related from the resources