Why a Medical AI That Finds Only 30% of Patients Keeps Saving More Lives

In Pennsylvania, a hospital system runs an AI tool to screen for colorectal cancer. Every week, it scans through hundreds of overdue patients and flags only 27.7% of those who actually have the disease. In other words, it skips seven out of ten. Over the years it's been running, the people it does find have seen their two-year mortality drop by 43%. The AI was never designed to find everyone. It answers a single question: given that nurses have limited time and colonoscopy slots are finite, who should we call first? Geisinger Health System deals with around 450 overdue patients every week. The AI runs through their age, sex, and recent blood work, then flags the eleven highest-risk individuals. Nurses make eleven calls that week. Eleven is a designed number. Flag a hundred people, and nurses can't get through to all of them, and the colonoscopy schedule won't hold. Flag precisely, so every person who makes the list actually gets followed up and booked in. That four-month head start is often what separates early-stage from late-stage. Among those the AI flags for a colonoscopy, 6 in 100 turn out to have cancer. Among those it doesn't flag, fewer than 1 in 100 do. One thing the research documented: when a hospital tried using an automated voice call to notify patients about screening, the follow-through rate was nearly identical to doing nothing at all. A real person on the phone is what gets people to actually show up. The nurse makes the eleven calls matter. We tend to evaluate AI by the numbers it produces: how much it finds, how often it's wrong, how often it hallucinates. Today's leading models, with access to tools, hallucinate around 3% of the time. When uncertain, they're more likely to say "I can't verify that" than to generate a plausible-sounding answer. Most people's mental image of AI reliability is stuck a few years back. The 27.7% figure and the 3% hallucination rate point to the same issue: the number alone doesn't tell you what framework it's measured in. A metric that looks terrible in one context might be exactly right in another. Before the number means anything, you have to know what question it was answering. Next time you see "AI accuracy" or "how much AI found," ask first: what is it measuring, and under what conditions? Is the thing being measured actually the problem you're trying to solve?