I think the logic actually goes "if you reject upfront all programs with a certain type of badness as well as some programs that falsely appear to have that type of badness, you will reduce the costs associated with bad programs by catching many of them sooner, at the cost of having to rewrite some already-good programs in less-simple ways to avoid the appearance of badness."

Let's put that on a bumper sticker :)

This article appears to me to failing in being entirely cogent; his example of the if that returns a number or a string seems nonsensical, and therefore "bad" by some definitions, which also points out that bad is subjective.

That being said, perhaps it would have been a better article if the author presented some arguments of "bad" programs that pass the type system, as type systems are no silver bullet and do not guarantee rejection of all bad programs.