The 5 Questions I Ask Before Any High-Stakes AI Model Ships
The one thing I want people to remember is: the highest-stakes AI mistake I've seen wasn't a bad algorithm. Instead, it was a proxy variable nobody thought to question. These five questions are how I catch it now. Background: Every year, more AI models get deployed into decisions that can genuinely harm people — bail, hiring, lending, medical triage. Criminal justice has been the highest-stakes proving ground for this for over a decade, and it's taught me five questions that now go through my head before any consequential model reaches production, regardless of industry. I'll walk through one real case from criminal justice risk assessment where skipping question 3 quietly baked bias into a "neutral" model, and show how the same blind spot shows up in fraud detection, hiring algorithms, and credit scoring. Five minutes, one framework, one story, one slide the room can screenshot and actually use Monday morning.
