For example, many people have racist tendencies but rationalize them as something else.
"There is a big group of male magentas dressed the same over there, I should avoid them because they are almost certainly a gang"
is rationalized as:
"I don't avoid the group of male magentas dressed the same because I'm racist and assume that most magentas are in a gang, but because I've been taught common sense. Common sense dictates people that look like that must be gangsters."
Perhaps if we interviewed people without being able to see them, their name, or anything about them and using voice obfuscation we could eliminate a lot of bias?