It strongly depends on the problem set.

For teaching id broadly day either set it as a prerequisite to examine something it can't do or leave it until the last lecture to show the course material can be applied more efficiently.

Learning is as much getting used to using tools correctly as understanding what the tools do and how they do it.

I've worked with those who yes, have had their brain rot due to blindly trusting Mathematica or equivalent and they never spot first step errors from their blind spot. That being said I've worked with one person who used it really well and I think they were genuinely more efficient for it existing.