Not the OP, but using comparisons of 1995 software to try and make a point about the current software situation is some indication that people making that argum

Not the OP, but using comparisons of 1995 software to try and make a point about the current software situation is some indication that people making that argument are several decades behind. The software that we use in 2020 and need to be reliable in 2020 is quite different than 25 years ago, heck, the platforms running the software are quite different - in 1995 you'd be using some program running on an OS on your computer (or through a remote terminal on some other computer); in 2020 you'd be using some program running on a distributed set of cloud servers through a browser sandbox - and the way how the reliability of that remote system is maintained is simply different from how you'd ensure the reliability of a local utility program. That 1995 study is perhaps interesting enough to suggest doing some experiments again, but currently it's outdated enough to be simply irrelevant.