Before Discovery Begins, Map the AI Records
Reprinted with permission from the September 22, 2026 edition of the “New York Law Journal” © 2026 ALM Global Properties, LLC. All rights reserved. Further duplication without permission is prohibited, contact 877-256-2472 or asset-and-logo-licensing@alm.com.
New York already has the preservation and discovery framework. The practical task is identifying consequential AI records, their custodians and their retention settings early enough to use it.
By Pouya Shafabakhsh
The next AI discovery problem may begin before anyone serves a discovery request. At matter opening, litigators routinely identify custodians, email systems, messaging platforms, shared drives and other sources of electronically stored information. In 2026, one more question belongs at the front of that exercise: which AI systems materially touched the facts, decisions, communications or litigation work, and what records do those systems still retain?
That question is deliberately narrower than asking whether anyone used AI. A company may use AI hundreds of times a day, while only a handful of interactions have any connection to a dispute. A law firm may use an approved research or drafting tool without creating evidence relevant to the merits. An expert, client employee or deal team may use another system in a way that does affect a disputed decision. The practical challenge is to separate consequential uses from background use before routine deletion, account settings or vendor architecture make the distinction impossible to reconstruct.
New York’s existing framework is capable of handling that task. 22 NYCRR Part 161, effective June 1, 2026, governs AI used in preparing papers submitted to courts of the Unified Court System and expressly excludes material constituting or proffered as evidence from its definition of a court “paper.”
Three days after Part 161 took effect, Nassau County Supreme Court decided Assini v. Hayward, 2026 N.Y. Slip Op. 26086, and quashed subpoenas directed to OpenAI for a self-represented defendant’s prompts, uploads and outputs used in litigation preparation. The decision did not create a general rule that AI records are either discoverable or protected. It showed something more useful for day-to-day practice: AI interaction records can become the subject of a discovery dispute, and their treatment depends on why they were created, what they contain and the protections that apply.
New York’s preservation framework is equally established. In VOOM HD Holdings LLC v. EchoStar Satellite L.L.C., 93 A.D.3d 33 (1st Dep’t 2012), the First Department applied the reasonable-anticipation standard to electronically stored information and emphasized the need to suspend routine destruction when the preservation duty arises. VOOM long predates generative AI. That is precisely why it remains useful. The first preservation question is not whether a record came from a chatbot. It is whether potentially relevant information exists, whether there is a duty to preserve it, and whether reasonable steps are needed to keep it available.
Start With the Use, Not the Tool
A defensible AI record map begins with the matter, not with a list of software. Counsel should identify the decisions, communications and workstreams that could matter to the claims or defenses, then ask whether an AI system materially influenced any of them. Examples may include a business recommendation later challenged in litigation, a contract comparison relied on during a transaction, an internal investigation summary, an expert’s analytical workflow, a client communication, a document-classification process or litigation preparation.
This approach avoids two opposite errors. One is ignoring AI records because the final work product looks conventional. The other is issuing a sweeping instruction to preserve every prompt across an organization. Neither is a useful substitute for matter-specific judgment. The objective is to identify AI interactions with a meaningful connection to the dispute and document why they were included—or excluded—from the preservation scope.
Separate Source Evidence From AI Interaction Records
An AI output should not displace the underlying evidence from which it was derived. If a system summarized 400 emails, compared 60 contracts or ranked thousands of documents, the source collection remains a different evidentiary object from the prompt, output or model-generated synthesis. A useful intake review therefore maps both layers: the source records and the AI interaction that transformed, summarized, classified or interpreted them.
That separation matters later. The source documents may be independently relevant even if the AI interaction is not. Conversely, the interaction may matter because it shows what information was presented to the system, what instruction was given, what result was returned, or how a human decision-maker used that result. Treating the two layers separately also reduces the risk that a generated summary becomes a substitute for a native record that should have been preserved in its original form.
Find the Record Before Setting the Hold
The visible chat window is not a record map. For each consequential use, counsel should identify the user, account or workspace, the system involved, the date range, the source material provided to the system, the categories of records created, the current retention setting, the available export method and whether additional records are held by a third-party provider. The answer may differ between a personal account, a law-firm enterprise workspace, a client-controlled platform, an e-discovery application and an expert’s system.
This is where an early technical conversation can save legal work later. A preservation decision is difficult to implement if no one knows whether the relevant item is a saved conversation, an administrator log, an exported file, a model-generated artifact or a record that exists only within a vendor environment. The goal is not forensic collection of every system at intake. It is enough to know what may exist, where it is likely to reside, who can preserve it and whether there is a time-sensitive retention issue that needs attention.
Preserve First; Classify Carefully
Preservation should not be confused with production. Assini is a useful reminder of that distinction. The court quashed the vendor subpoenas after applying New York’s litigation-preparation framework and finding persuasive authority protecting certain AI-assisted litigation preparation. But the decision is a trial-level ruling tied to its facts; it should not be read as a categorical privilege for AI prompts. CPLR 3101 separately addresses privileged matter, attorney work product and trial-preparation material, and CPLR 3103 gives courts tools to limit or condition disclosure where appropriate.
For practice purposes, the sequence matters. If an AI interaction may be relevant and subject to a preservation duty, keeping it available does not concede that it must later be produced. Counsel can preserve material while separately evaluating privilege, work-product protection, relevance, proportionality, confidentiality and the scope of any request. That approach protects both sides of the analysis: potentially relevant information is not lost merely because its legal classification is unresolved, and preservation is not treated as a waiver of available objections or protections.
Bring the Map Into Discovery Planning
New York’s Commercial Division already provides a useful model for making technology visible early. Rule 1(b) requires counsel appearing at the preliminary conference to be sufficiently versed in their clients’ technological systems to discuss electronic discovery competently. Rule 11-c calls for counsel to confer before the preliminary conference about electronic discovery topics, and Rule 11-c(h) states that a party should take reasonable steps to preserve ESI that it has a duty to preserve. The ESI Guidelines remain advisory, but the structure is practical: understand the systems, identify the issues early and address disagreements before they become avoidable discovery motion practice.
An AI record map can make that conversation more concrete. It need not be elaborate. For each material AI use, a litigation team can record the custodian, system or workspace, matter-related purpose, source data, record types, retention setting, export path, person or entity controlling the record, preservation action and any preliminary protection concern. That single page can expose gaps quickly: an important workspace with no export route, an expert using an unidentified tool, a client account set to delete history, or a vendor holding records that the client cannot directly access.
What a Proportionate Protocol Avoids
A disciplined protocol should reduce discovery, not expand it. It avoids a generic “preserve all AI” instruction that no custodian can apply consistently. It avoids assuming that every prompt is relevant merely because AI was used. It avoids waiting for an adversary’s request to learn that a system no longer retains the information. And it avoids treating a subpoena to a technology provider as the first step in understanding records that may be better identified through the party, custodian or system owner.
It also supports narrower, better-informed discovery requests and objections. Once counsel understands the record types and architecture, the parties can discuss date ranges, custodians, accounts, subject matter, export formats and proportionality with greater precision. Where protected material is implicated, the same map helps isolate it rather than commingling unrelated matters, personal information or litigation strategy with potentially responsive business records.
The Practical Point Is Visibility
The emergence of AI does not require New York courts or litigators to abandon familiar preservation and discovery principles. Part 161 carefully defines the scope of its court-paper policy. Assini illustrates how AI interactions can become the subject of a vendor-directed discovery dispute. VOOM supplies a durable preservation principle. The Commercial Division’s ESI rules show how technological systems can be brought into case planning in an orderly way.
The practical adjustment is earlier visibility. When a matter is reasonably expected to involve AI-touched facts or litigation work, counsel should know which uses mattered, what source evidence sits behind them, where the related records reside, how long they are retained and who can preserve them. From there, New York’s existing rules can do what they are designed to do: separate what should be preserved from what should ultimately be produced, protected, limited or excluded. The objective is not more AI discovery. It is a sufficiently complete record to make the right discovery decision when the question actually arrives.
——————————————————
Pouya Shafabakhsh is a Judicial Forensic AI Auditor and author of North American Litigation AI GRC Methodologies. He is co-founder and chief AI governance officer of Radsam Academy of AI Sovereign Governance. He is not an attorney. This article is analysis and opinion, not legal advice.
Reprinted with permission from the September 22, 2026 edition of the “New York Law Journal” © 2026 ALM Global Properties, LLC. All rights reserved. Further duplication without permission is prohibited, contact 877-256-2472 or asset-and-logo-licensing@alm.com.
Would you be opposed to finding your firm's AI gaps before opposing counsel or insurance companies do? I invite you to assess your firm using our AI Risks Calculator. Safely, securely, privately, and confidentially—with no need to enter any information, not even your email address—you can get a comprehensive AI Risks and Liabilities report in just five minutes.
If you have a sensitive file and want to protect it against AI hallucinations, bias, algorithmic discrimination, and AI-assisted risks through an Air-Gapped Shadow AI Audit before court, or if you are involved in a legal dispute related to AI and need a Judicial Forensic AI Audit, at Radsam, we are willing to assist you whether through our human-based services or by providing our in-house solution, the Sovereign Sanctuary Vault. In these cases, I invite you to fill out our Pre-Qualifying Assessment Form. We will review your information, qualifications, and capacity, and will contact you within two business days.
Comments