Public Safety AI Lab develops the science of governing, verifying, and securing AI — from evaluation to accountable deployment — in settings where a mistake is a wrongful outcome, not a bad recommendation.
Public safety is the hardest test of AI. Here a failure is not a bad recommendation — it can be a wrongful outcome, a due-process violation, an infringement of civil liberties. PSAIL builds the evaluation, governance, and assurance methods that can survive that scrutiny.
Research programs
National AI project
Synthetic personas
Papers in review or accepted
PSAIL organizes its work around research programs — not papers. Every publication, patent, and project connects back to one of these four.
Every research asset is a node in one connected graph. Trace a policy outcome back through the project, the patent, the paper, and the program that produced it — the full provenance of an idea.
From benchmarks to deployed systems — our projects turn evaluation science into tools institutions can actually use.
Releases, publications, talks, and milestones from the lab.
A field casebook on the eighteen months in which a 140,000-person public organization decided how it would adopt AI: seventeen decision points, what worked, what failed, and the organizational mechanism underneath each one.
AI did not change the kinds of crime, it changed their cost. The piece argues that defence has to make attacks expensive again — chokepoint by chokepoint — rather than betting on detection alone.
Hyundap answers modern questions from the standpoint of a historical figure by retrieving primary sources such as the Annals of the Joseon Dynasty, and shows the original record behind every answer. Fifteen works were selected from 602 entries by the Ministry of Culture, Sports and Tourism and five co-hosting bodies.
Work on dynamic DNN partitioning for public-safety edge networks was presented at KCC 2026.
Collaborate
We work with researchers, engineers, and public institutions who take reliability seriously. There is real work to do.