I think with the AI it's not only question about following the spec, but about the completness of the spec. Before the AI, if specification wasn't complete, then program would lack of a feature, won't start correctly or crash. Nowadays, agents will happily fill any missing specification parts with random stuff, often looking reasonable at the first glance. This random content can be different with every iteration of the LLM work.

To find the right position in this spectrum, I think we need to find a level of specification that we care about.