PDA Whiteboard Paper #2: What the FDA Is Building with AI

George Kwiecinski recently presented an analysis of the FDA entries in the FY25 HHS AI Use Case Inventory. The two-page paper, What the FDA Is Building with AI, is now on the PDA Whiteboard for public reading and member ratings through approximately November 15, 2026.

What the inventory reports

HHS listed 447 AI use cases across its agencies. The FDA accounts for 67 entries. Of those, 42 are reported as deployed, 12 as pilots, and 13 as pre-deployment. The Center for Drug Evaluation and Research accounts for 40 FDA entries, including 30 reported as deployed.

The FDA entries cover different kinds of work. The inventory classifies 23 as generative AI, 22 as natural language processing, 18 as classical or predictive machine learning, two as computer vision, and two as other. Selected examples in the paper include tools that extract facility information from submissions, group quality signals in Field Alert Reports, and help reviewers examine adverse event reports.

What the numbers do not tell us

An inventory entry is a reported use case, not necessarily a separate system or a newly launched project. A "deployed" label does not establish how often a tool is used, how accurate it is, or whether it improved outcomes. The descriptions of the selected tools point to information extraction, signal detection, and support for human review. They do not show that regulatory decisions are automated.

That distinction matters when discussing AI adoption at FDA. The inventory offers a useful map of reported activity, but each project still needs its own evidence on intended use, validation, human oversight, results, and cost.

Read and respond

Read the paper on the PDA Whiteboard. PDA members can rate it during the 45-day window. Comments, corrections, and critical readings are welcome. They help sharpen the next round of research.

The two-page PDF includes the selected-project table, method, scope limits, and references.