The problem is that improper input data can come from data scientists, or it can be a result of the same issues that cause society's bias. In that later case it isn't improper data so much as data properly collected from an improper society. That is super complicated to adjust for even if you decide to do so. Children of poverty that are malnourished grow up with deficits that might make them less valuable employees, extending the cycle. But the data and AI were not wrong to correlate childhood poverty with performance levels.

  Then there is the far trickier case of the AI picking up on real things that we choose not to accept or talk about. Perhaps it is safe to give an example using white men?  When they go bald very early in life(prince William for instance), it generally is due to high testosterone. Which in turn leads to all sorts of predictable behaviors and health issues, good and bad. AI would spot all that from a simple picture, judging people on their appearance can occasionally work really well. But that doesn't mean we should do it.  And again, it isn't improper data, just data we choose not to act on.