Training a ML model to understand how "most sensible, considerate humans" would act is anything but straightforward. I'm not even sure most people would even consider cropping an athlete at the jersey number regardless of race or gender - it just doesn't make sense yet the machine seemed to do it at the same rate for male vs. female. Retrofitting discrimination onto this result only once you learn the label of the data isn't particularly useful. We want to know how to make good non-harmful predictions in the future.