That’s a fair point. But it greatly limits the scope of human-introduced error. I think already for FLT, the surface area for error in the kernel and in axiom translation is substantially smaller than the entirety of the literature which Wiles’s proof recursively depends on.