An AI-designed drug is a molecule whose structure was proposed or selected with the help of machine learning, then made and tested through conventional laboratory and clinical methods.
Design is one step, not the whole job
Machine learning is good at searching a very large space of possibilities and ranking candidates by predicted properties. That is the design step. Making the molecule, checking it in cells and animals, and testing it in people are separate steps that follow the same rules they always have.
A useful way to hold this: AI changes which molecule you start with. It does not change what that molecule must survive afterward.