If a Walmart has ~100 people in it and wants to get rid of 4 shoplifters but really sucks at selecting them well then the likely result is 4 normal people are very upset while all of the shoplifters are still there.

In the same scenario, even if Walmart is right about who they ejected 75% of the time then they still have ~1 shoplifter remaining and ~1 very upset person.

Even in an ideal world where Walmart is right about ejection 100% of the time it doesn't mean they start receiving 0 new shoplifters either, it just means the number of people wrongly made upset is 0.

Discord's problem (on both ends) lies in lack of depth in investigating bans. It takes resources to review when someone shouldn't be banned and it takes resources to make sure you ban everybody. Putting too low of resources into banning just means that both sides of the scale manage to get tipped in the wring direction at the same time.