I don’t see why this is so surprising; I see this as a downgrade from what would be possible with source to source translators and stepwise refinement as developed in the formal methods community.
Now, if you train the model yourself, perhaps much of the problems can be avoided. I remember a section in the CS classic Structure and Interpretation of Computer Programs (AKA SICP p. 161 in footnote) that most old Fortran programs had a pattern of map, filter and accumulate, which I suspect hasn’t changed that much.
But adding randomness via the LLM to this code conversion process can’t be a good thing.