Overview of Discovery of AI: Prompts, Training Data, and the New ESI
This class examines how discovery practice applies to generative AI systems, including the prompts, outputs, and underlying model artifacts that may become relevant evidence in litigation. Faculty explain, in plain terms, what generative AI is and how it differs from traditional document retrieval, why outputs are mutable even when the same prompt is run twice, and how layered technology stacks (system prompts, fine-tuning, retrieval context, harnesses) affect what can later be reconstructed and produced.
The discussion distinguishes inference-layer artifacts (prompts, outputs, system prompts and configuration, tool and agent logs, retrieval records, and usage telemetry) from model-layer artifacts (model cards, evaluation reports, version history, validation and testing records, and governance materials), and explains why some of these are far less complete than requesting parties expect. Panelists discuss why AI-related evidence can disappear quickly, including short vendor retention windows, contracts that do not give corporate customers access to underlying logs, model deprecation practices, and the growing use of consumer AI accounts for work-related tasks. The program also addresses possession, custody, and control questions under Rule 34 when AI tools or agents are involved, the practical difficulties of "deposing" an autonomous agent, and the importance of understanding an agent's memory and audit trail limitations.
Practical guidance is offered on using interrogatories and Rule 26(f) conferences to narrow discovery requests, using the vendor's own terminology for artifacts, and working cooperatively with opposing counsel to focus requests given the technical complexity involved. Finally, the panel discusses privilege and work product considerations specific to AI tool use, including the risks of commingled expert use of AI tools and the practice of creating separate, logged tool instances for work done at counsel's direction, along with recent decisions addressing burden, proportionality, and the production of AI logs and expert prompts.