AIs are very good at finding a (locally) optimal solution without understanding the context of what that entails. When is the biggest jump in engagement? When an ad ends and the show resumes. Users tend to turn up the volume and return wandering eyes back to the screen. Therefore, every time an ad break ends, the system starts an ad break. This approach also maximizes the amount of ad time per hour of content.
(I am joking, if it is not obvious)