[bastijn]

and you would be right. Say I have a broken phone which is an Android. A prejudice would be that androids are likely to break. Mathematically we could represent as

  P(broken|android) =~ P(android|broken)

We see from Bayes rule that this is false, since

  P(android) >> P(broken)

[braythwayt]

Even worse, what actually happens with prejudice is that if I think that Androids are shitty phones, when I see a broken Android, I add it to my collection of “Shitty Android” data points.

But when my iPhone breaks, I take it into Apple for service and add it to my collection of “Apple gives great service” data points.

Saying that prejudice is just applied statistics completely mischaracterizes prejudice.