We read a lot about bias in AI or even bias in education. "According to common usage, 'bias' is always a deviation that systematically tends toward a certain direction." Fair enough. But what is it exactly that bias is a deviation from? The best we can manage is something like "a pattern considered to be correct and desirable in some way." But that opens the door to a pressing need to consider more deeply something commonly perceived to be a problem. That's what this paper (22 page MS-Word) does. It offers a five-category taxonomy (mechanism, effect, content, stage, feature) that is "a taxonomy of bias individuation practices." This is extended to complete a much larger categorization of types of bias. These are then applied to a five-stage 'research pipeline' to identify where they occur. I think it's a start, but a taxonomy is no more than a way of describing a landscape, and should not be mistaken as an understanding of that landscape.
Today: Total: [Direct link] [Share]
