AI that writes code must be tested by its own output's test suite.

Clever prompting doesn't prove a program is correct. Requiring AI-generated code to pass the same tests as human code keeps everyone honest.