I'm trying to remember where I saw the idea developed. Possibly by Andreas Schou at Google+.
Basically: if you've got some discriminating test with Type 1 and Type 2 error (false positive, false negative, don't ask me which is which), then as the actual rate of incidence falls, no matter how sensitive your test, at some point it's mostly making false positive rather than true positive selections.
Cue a system which rewards based on, say, arrests, rather than on correct convictions, and doesn't penalise on false arrests or wrongful convictions.
You're going to see a lot of false/wrongful arrests and convictions.
Makes me think if part of what happens within immune systems isn't similar.
Bruce Schneier probably covers similar ground, now that I think of it.
