A superset of this: you can worry a lot less about how you are perceived when dealing with anyone who has a very large N, because you'd have to be at the at like p = 0.995 of their distribution to be odd enough to trigger Serious Negative Consequences, or they'd do nothing else.
This is useful when e.g. dealing with government or banks and feeling like "Oh shoot, have I just said something which suggests I might be a criminal?"
Contingent on being early in funnel, if that classifier would select you, it would select a large number of confused people.
Contingent on being early in funnel, if that classifier would select you, it would select a large number of confused people.
e.g. "I'm calling about my account." "What's your name?" "$NAME." "That's not the name on the account." "Oh I'm so sorry it's my wife's account."
You don't have to be sorry or worry here. This is at least a 5% occurrence for them.
You don't have to be sorry or worry here. This is at least a 5% occurrence for them.
(In that example you might or might not be authorized to talk about it, either formally or by their understanding of their own policies, but "Called about an account one didn't have fully authority to talk about" is at least a 10% occurrence for them.)
The subtlety here:
The Bayesian likelihood of you being a criminal contingent on you saying something which suggests criminality *after they suspect you of being a criminal* may very, very well be high enough to cause consequences.
The Bayesian likelihood of you being a criminal contingent on you saying something which suggests criminality *after they suspect you of being a criminal* may very, very well be high enough to cause consequences.
(This is yet another reason to not answer, even in jest, "Yes, this is a bomb" if you are ever asked whether something is a bomb, even though a commanding majority of people saying "Yes" to that question probably aren't accurately identifying a bomb.)
