Stanford study: doctors got worse after AI assistance was switched off

cen6wkf · reddit · 2026-09-27

A Stanford AI-med-ed finding shows trained colonoscopists performed measurably worse after an AI-assisted period than before they ever touched the tool — skills quietly rot when handed off.

The author ties this to an IBM Institute for Business Value report: skill erosion is invisible by design, and nearly half of surveyed organizations have no one accountable for the human side of AI strategy. Software already has tooling (LinearB, Waydev, Athenian) tracking machine-touched vs. human output, but no one has ported it to structural/civil engineering, where a P.E. stamp carries legal liability.

The second half is effectively a startup blueprint: add a blind-call-before-reveal step to plan checkers, build per-engineer drift dashboards, match juniors to seniors by override history, and report to leadership and insurers — with an estimated 18-30 months before a mid-size firm moves past pilot. Key assumption: firms will tolerate logging individual divergence, testable via ten structured QA conversations.

Original post →

More from AGI Musings

AGI Musings channel →