Brady, Discovery, and AI: From Assertions of Knowledge to Process Accountability

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Artificial intelligence is changing criminal-case workflows before courts have settled how existing disclosure doctrine applies to it. The resulting questions are often framed too broadly: Is AI discoverable? Does an AI system's output become Brady material? Does the prosecution "know" what an automated system can find? Those formulations obscure the operational facts on which legal analysis must depend. An agency-controlled AI system may be authorized to access information without retrieving it; may retrieve information without flagging it; may flag it without delivering it into an accountable workflow; and may deliver it without any authorized human reviewing it. These are distinct, contestable events. This article argues that AI-assisted criminal-discovery governance should therefore begin with process accountability: the ability to reconstruct what a system was authorized to access, what it actually did, what it surfaced, how findings were routed, and what human review followed. The article engages with Vince Dutto's call for a pre-case-law conversation about Brady and AI. Dutto's contribution is practical and appropriately cautious: AI changes evidence review, data aggregation, work-product questions, expert methodology, and the records created by professional analysis. This article builds on that contribution without attributing to it a categorical claim that current Brady doctrine automatically requires disclosure of every AI prompt, output, training-data artifact, or system-wide technical record. Instead, it argues that traditional disclosure doctrine is necessary but insufficient when organizations deploy systems that retrieve, filter, summarize, rank, and continuously flag case-related information. Meaningful downstream disclosure depends on upstream procurement, provenance, retention, review, and verification controls. The article proposes a bounded framework rather than an automated legal rule: (1) distinguish AI used as evidence, expert methodology, internal analytical support, and agentic evidence-review infrastructure; (2) separate machine-generated leads from human legal judgment; (3) record the five operational states of availability, retrieval, flagging, delivery, and review; and (4) preserve attributable, time-stamped, case-linked records subject to explicit retention, access-control, change-management, and verification procedures. The objective is not to declare that an AI system makes Brady determinations or that every machine artifact is discoverable. It is to ensure that courts, prosecutors, defense counsel, and oversight bodies can evaluate what the institution's systems actually did rather than rely on untestable assertions about what the institution knew.