Encoding biases into machine learning models, and in general into the constructs we refer to as AI, is nearly inescapable — but we can sure do better than we have in past years. IBM is hoping that a new database of a million faces more reflective of those in the real world will help.