3 kinds of bias in AI models -- and how we can address them