Artificial intelligence is changing discovery. The Federal Rules, so far, are not. That distinction may be the most important lesson for corporate counsel. Employees now use generative AI to analyze contracts, prepare reports, develop strategy, investigate problems, and communicate with customers. Experts use AI to locate and analyze evidence. Lawyers use it to organize facts and prepare litigation materials. Each interaction may create prompts, outputs, chat histories, logs, drafts, and metadata.
When litigation arrives, the central question will not be whether those materials are “AI.” Instead, it will ring very familiar:
Are they relevant, within a party’s possession, custody, or control, nonprivileged, proportional—and, if so, were they preserved?
The emerging cases suggest that courts will answer those questions principally through existing discovery and evidentiary rules.
Prompts May Be Evidence
Consider an executive investigating an internal compliance concern. She uploads emails to an enterprise AI system and asks: “Identify anything suggesting we knew these invoices were inaccurate.” The system identifies several communications. She then asks: “Give me the strongest arguments that these invoices nevertheless comply with the contract.” Months later, the company is sued for fraud. Opposing counsel requests the executive’s prompts, outputs, and AI history. Are they discoverable? Potentially.
Rule 26(b)(1) permits discovery of nonprivileged matter relevant to a claim or defense and proportional to the needs of the case. The Rule does not distinguish between email, text messages, collaboration platforms, or AI-generated information. Fed. R. Civ. P. 26(b)(1). Rule 26(f), moreover, expressly requires parties to discuss preservation and discovery of electronically stored information. And recent cases are beginning to apply those principles directly to AI.
In In re OpenAI, Inc. Copyright Infringement Litigation, the Southern District of New York required production of a sample of approximately 20 million de-identified ChatGPT output logs after concluding that the requested material was relevant and proportional; the court later denied reconsideration. In re OpenAI, Inc. Copyright Infringement Litig., No. 25-MD-3143 (SHS) (OTW), 2025 WL 3468036 (S.D.N.Y. Dec. 2, 2025), aff’d, 2026 WL 21676 (S.D.N.Y. Jan. 5, 2026).
Another early dispute arose in Tremblay v. OpenAI, Inc., where authors alleging copyright infringement relied on ChatGPT-generated summaries of their works as evidence supporting their claims. OpenAI sought discovery of the plaintiffs’ pre-suit testing, including the prompts and outputs that produced both favorable and unfavorable results. A magistrate judge initially compelled production, reasoning that the account information, negative test results, and testing process were necessary to evaluate the results plaintiffs had placed at issue. On review, however, the district court drew an important distinction: although certain factual information concerning the testing was discoverable, the prompts crafted by counsel reflected counsel’s mental impressions about how to interrogate ChatGPT and therefore constituted opinion work product entitled to heightened protection. Tremblay v. OpenAI, Inc., No. 23-cv-03223-AMO, 2024 WL 3748003, at *1–3 (N.D. Cal. Aug. 8, 2024).
A bit more on privilege below. But the lesson is not that every prompt is discoverable. It is that AI data receives no categorical exemption from Rule 26 and that the discoverability of a prompt or generative AI output logs may depend substantially on who created it, why it was created, what it reveals, and whether the party has relied on the resulting AI analysis to support its claims or defenses.
Rule 34: A Prompt May Still be ESI
Rule 34 permits requests for documents and electronically stored information within a party’s possession, custody, or control. Fed. R. Civ. P. 34(a). The rule is technologically neutral. If responsive prompts, outputs, logs, or AI-generated drafts exist and are within a party’s control, the fact that they reside inside an AI platform instead of an email server, a database, or a file cabinet, may not, by itself, remove them from Rule 34 consideration. Relatedly, a particularly noteworthy development involves experts. In Conservation Law Foundation, Inc. v. Shell Oil Co., a magistrate judge ordered revised Rule 34 responses concerning AI prompts or queries used by an expert and her team to reduce a large document production to the materials used in her analysis. The court regarded that process as part of the expert’s methodology and therefore potentially discoverable. Conservation Law Found., Inc. v. Shell Oil Co., No. 3:21-cv-00933 (VDO), ECF No. 970 (D. Conn. May 18, 2026). The order was subsequently stayed pending district-court review, so it should be treated as an important developing decision rather than settled law; but illuminating nonetheless.
The practical implication is immediate: counsel retaining experts should determine before the first prompt whether AI will be used, what the system retains, and whether prompts and outputs can later be collected.
Preservation: The Litigation Hold Has a New Data Source
The duty to preserve does not begin because information is discoverable. It arises when litigation is reasonably anticipated and extends to relevant information that should reasonably be preserved. Rule 37(e) addresses the consequences when ESI that should have been preserved is lost because reasonable preservation steps were not taken and the information cannot be restored or replaced. Fed. R. Civ. P. 37(e). Prejudice permits appropriately tailored curative measures; the most severe sanctions require intent to deprive.
DR Distributors, LLC v. 21 Century Smoking, Inc., 513 F. Supp. 3d 839 (N.D. Ill. 2021), remains an especially useful published treatment of these obligations. The court emphasized identification, preservation, collection, review, and production as fundamental components of ESI discovery and analyzed whether reasonable preservation measures had been taken. Courts continue to rely on its Rule 37(e) analysis.
AI complicates that familiar obligation. Some platforms retain complete histories; others permit deletion; enterprise settings may differ from consumer versions; and relevant information may exist simultaneously in the user’s account, vendor systems, exported documents, and downstream business records. The answer may not be to “preserve every prompt.” It may be, rather, to identify potentially relevant AI systems when a hold arises and make defensible preservation decisions about AI data within the firm’s possession, custody, or control.
Rule 37: AI Does Not Outsource Counsel’s Judgment
AI also potentially affects discovery conduct itself. For example, in White v. Walmart, Inc., a federal magistrate judge confronted counsel who uploaded an opponent’s discovery responses into AI, asked the system to identify deficiencies, and copied the AI-generated criticisms into correspondence with opposing counsel and the court. The court held that lawyers must exercise independent judgment in discovery and could not rely exclusively on AI-generated objections. White v. Walmart, Inc., No. 1:25-cv-01120-RLY-TAB, ECF No. 37 (S.D. Ind. Apr. 14, 2026).
Similarly, Tijerino v. Spotify USA Inc. addressed counsel’s use of AI during the discovery meet-and-confer process and reinforced that Rule 37’s good-faith conferral obligation cannot simply be outsourced to an AI system. Tijerino v. Spotify USA Inc., No. 2:24-cv-02290, ECF No. 78 (E.D. La. July 7, 2025). AI can assist discovery. It does not replace counsel’s Rule 26(g) certification or Rule 37 responsibilities.
Privilege and Work Product: Who Asked the AI—and Why?
Not every responsive AI interaction may be discoverable. Rule 26(b)(3) protects qualifying materials prepared in anticipation of litigation. Courts are already confronting whether AI interactions fall within that protection. AI use may also implicate counsel’s independent confidentiality obligations. Model Rule 1.6(c) requires lawyers to make reasonable efforts to prevent unauthorized disclosure of, or access to, client information, while the duty of technological competence requires counsel to understand the benefits and risks associated with relevant technology. Those obligations may require counsel to consider not only a platform’s security and retention practices, but also whether particularly sensitive information should be redacted, anonymized, or otherwise protected before it is submitted to an AI system. What constitutes reasonable protection will necessarily depend on the sensitivity of the information, the capabilities and safeguards of the particular platform, and the circumstances of its use. In Morgan v. V2X, Inc., the District of Colorado held that a pro se litigant’s AI-assisted litigation work could receive Rule 26(b)(3) work-product protection, while concluding that the identity of the AI platform itself was not protected. The court also imposed AI-specific safeguards governing the use of confidential discovery material. Morgan v. V2X, Inc., No. 25-CV-01991-SKC-MDB, 2026 WL 864223, at *3–7 (D. Colo. Mar. 30, 2026).
The emerging lesson is highly fact dependent: who created the prompt, why it was created, whether counsel directed the work, what information was disclosed, and how the platform handled it may matter much more than the label “AI.”
Proof: Is the AI Output What You Say It Is?
Discovery is only half the problem. Eventually, someone may want to introduce AI-related material into evidence. Rule 901(a) on authentication requires sufficient evidence to support a finding that an item “is what the proponent claims it is.” Fed. R. Evid. 901(a). Rule 901(b)(9) specifically contemplates evidence describing a process or system and showing that it produces an accurate result. Existing machine-generated-evidence cases provide a useful starting point. In United States v. Lizarraga-Tirado, the Ninth Circuit distinguished human-entered information from automatically generated Google Earth information and explained that machine-generated information could be authenticated by evidence showing that the process or system produces an accurate result. 789 F.3d 1107, 1109–10 (9th Cir. 2015). That distinction may become increasingly important with AI.
Counsel confronting an AI-generated image, summary, analysis, or purported recording should ask: Who created it? What inputs produced it? Was it altered? Can the process be reproduced? What metadata exists? And where? Is the proponent offering a human assertion communicated through AI, or information generated by a machine process? Answers to each or any of these queries may impact what steps need to be taken for authentication.
AI also creates the opposite problem: authentic evidence may be attacked as synthetic. Rule 901’s relatively modest threshold remains applicable, but deepfakes and generative editing will likely make provenance, metadata, witnesses with knowledge, and expert testimony increasingly important. And where AI-generated analysis is offered through an expert, Rules 702 and 703 remain additional gates. The fact that an expert used AI does not relieve the proponent of establishing that the expert’s methodology is reliable and that the opinion rests on an adequate factual basis. Calling an analytical process “AI” does not establish reliability.
The Biggest Question: What Should Companies Preserve Now?
In most circumstances corporate counsel should not respond to these considerations by preserving every employee’s AI history indefinitely. That would often be unnecessary, expensive, and inconsistent with the basic concepts of proportional discovery. The better approach is familiar from traditional ESI practice:
• identify approved AI platforms and understand their retention settings;
• determine which business functions use AI for legally significant decisions;
• add relevant AI sources to litigation-hold questionnaires;
• determine whether prompts, outputs, and logs can be exported or preserved;
• address AI use expressly with experts and litigation vendors;
• include AI in Rule 26(f), ESI-protocol, and protective-order discussions when relevant; and
• document reasonable preservation decisions.
The emerging cases suggest a simple rule. Do not necessarily ask simply whether AI is discoverable. Ask whether the information created through an AI tool would be discoverable if it had been created by email, text message, spreadsheet, or any other electronic system. From that standpoint, the Federal Rules already supply much of the answer.
Bottom Line
AI is creating new sources of evidence, not a new law of discovery. Prompts may reveal intent. Outputs may document decision-making. Logs may establish a chronology. Expert prompts may reveal methodology. AI-assisted litigation work may implicate privilege or work product. And AI-generated evidence will test familiar rules concerning authentication, reliability, and proof.
For corporate counsel, the challenge is therefore not predicting a future set of AI discovery rules. It is ensuring that existing preservation, discovery, privilege, and evidentiary practices recognize that AI has become another place where legally significant information is created. The organizations best prepared for AI discovery will be those that make that adjustment before the litigation hold arrives.
Related Attorneys
Counsel
Counsel
