Webinar Recap: Treating Generative AI Records as Discoverable ESI
By the Legal Cyber Academy editorial team ·
Free courses on this
Recorded lectures, free to watch, no account needed.
Generative AI tools now sit inside ordinary business workflows, and their records are starting to surface in document requests, preservation letters and motions to compel. Much of this material is unfamiliar, changeable and short-lived, yet the rules that govern it have not changed.
A Legal Cyber Academy webinar, Discovery of AI: Prompts, Training Data, and the New ESI, took up that problem. The panel included moderator Daniel B. Garrie, founder of Law & Forensics and a neutral at JAMS; Scot G. Doyen, a shareholder at Clausen Miller P.C.; Karen Silverman, CEO and founder of The Cantellus Group; Mari Grace, counsel at Debevoise & Plimpton LLP; and Patricia Antezana, counsel at Reed Smith LLP.
Generative Output Is Not a Fixed Record
The panel started with how the technology works. A generative model predicts likely next words from patterns in its training data rather than retrieving a verified answer. The same prompt can yield different results on successive runs, so outputs are not reproducible by default.
Enterprise deployments stack layers on top of a frontier model: vendor services, fine-tuning, department and user instructions, and controls on what the system can reach. Each layer shapes the result, which is why panelists stressed activity logs showing who used which tool with which settings.
Two Layers of Artifacts
Panelists sorted potentially discoverable material into two groups. Inference-layer artifacts reflect what users did: prompts and outputs, system prompts and configuration, tool-call and agent logs, retrieval records, and usage telemetry showing who did what and when. Model-layer artifacts describe the engine itself: descriptions of training and fine-tuning data and their provenance, model cards and evaluation reports, version history, validation and testing records, and governance materials. The panel expected testing and governance records to be thinner than requesting parties might hope.
Retention Is Short and Often Outside the Client's Control
Logs may be kept for days, and sometimes only a day or two. Corporate customers often lack contractual access to that data, and vendors can retire models at their discretion. The panel urged organizations to map where this material lives before any preservation duty arises. Panelists also flagged consumer AI accounts on employees' personal and mobile devices as the next round of the fights once waged over text messages, raising Rule 34 questions of possession, custody or control.
Precision Beats Breadth
The Federal Rules still govern, and in some recent cases judges have asked requesting parties to tie each item sought to a claim or defense. The panel favored early interrogatories to learn what exists and what the system calls it, followed by narrower document requests and deposition topics. Defined terms matter because engineers, vendors and lawyers use the same words differently. Demands for every weighting score in a model, or to freeze an entire model, invite burden objections.
Panelists presented the Rule 26(f) conference as a tool for cooperation, not advantage. Asking the other side to explain an unfamiliar system can also shift risk onto a party that refuses. For producing parties, the panel said burden arguments are weaker when a company profits from the same data, and pointed to sampling or iterative approaches.
Privilege Does Not Follow the Prompt
Typing a legal question into an AI tool does not make it privileged. The panel recommended enterprise tools, prompts that state the work is at counsel's direction, and a separate, logged work-product instance for experts, while noting that AI output an expert relies on can be discoverable.
The free course covers each of these points in more depth, including the recent decisions the panel reviewed and its discussion of autonomous AI agents.
Watch the full webinar course
This article draws on the Legal Cyber Academy course Discovery of AI: Prompts, Training Data, and the New ESI with Daniel B. Garrie, Scot G. Doyen, Karen Silverman, Mari Grace and Patricia Antezana. The course is free and includes the full recording, the slides, study notes and a final exam with a certificate of completion.
Go deeper — courses on this
Securities LawUser Guide: New SEC Cyber Rules and Balancing the Risk to Trade Secrets
The background and intent of the SEC's 2023 cybersecurity rules, reporting and board responsibilities…
Daniel B. Garrie, Michelle Tanney and 2 others · 1h 5m
RansomwareDecrypting the Threat: How to Protect Your Organization from Ransomware Risk
How ransomware works and how to manage the risk: tabletop exercises, OFAC guidance on ransom payments…
Daniel B. Garrie, Menachem Shafran and 2 others · 1h 3m
Cyber IncidentsNavigating Data Breaches: Vendor Risks
How third-party vendors create data breach risk, what to ask after a vendor breach, how to review…
Daniel B. Garrie, Stacy Harrison and 6 others · 2h 5m
Keep reading
- Webinar Recap: Turning Blockchain Records Into Admissible EvidenceA panel of litigators and neutrals covered how to collect, authenticate and attribute crypto evidence as digital assets reach everyday cases…
- Degree or certification: which screen are you trying to pass?Four different screens gate digital forensics hiring, and they reward different things. Work out which one stands between you and the job, t…
- Travel Rule Compliance for Crypto Firms: The Gap That Gets YouThe FATF Travel Rule requires virtual asset service providers to transmit originator and beneficiary data on transfers. Here is where compli…
Get the next one by email
Plain-English analysis of the law-and-technology developments that change how you advise. No more than monthly, and you can leave whenever you like.