It is repeatedly shown that the models are aligned towards tasks, not constraints. Which is also how the current market and companies operate, and use cases in general favour that. So I am not sure it is incidental, inevitable, "vague" or "just hard" as opposed to by design.
If you define a task and constraints to that task that contradict each other, the model is gonna try to solve the task against the constraints. It is perfectly aligned to what it is supposed to be aligned, because "alignment with human ethics" and other stuff is just a theatre or afterthought at best.
If you define a task and constraints to that task that contradict each other, the model is gonna try to solve the task against the constraints. It is perfectly aligned to what it is supposed to be aligned, because "alignment with human ethics" and other stuff is just a theatre or afterthought at best.