Human in the Loop Is Not a Safety System
Putting a person in front of an approval button has become the default answer to AI risk. But if that person can't catch the error in the time they're given, nothing is being overseen. Here's how to design review that actually works.



