AI generated tests can "lock down" features pretty well, do the basic bounds checking etc. and make sure no unrelated change breaks the feature.

But the more complex bits require human instruction and/or intervention.