I think it case of coding it may not be as bad, because new training data (AI generated code) is always empirically validated by tooling and by consumer. It may

I think it case of coding it may not be as bad, because new training data (AI generated code) is always empirically validated by tooling and by consumer. It may not be good but it mostly works, otherwise it is discarded or patched, so it has a bottom bar of "it works".