It seems that developers think the solution to many AI challenges is just to try to wall off results from problematic areas, rather than to really try to understand the ways the AIs are interpreting or failing to recognise issues humans find important. Kind of a Tienanmen Square erasure approach. Yet there are many examples of humans circumventing these controls, and presumably just because the results are not shown for images doesn't mean there are no results. What does training mean in that context?
I mean where is the return on investment here? Classifying gorillas isn’t a must have for users, and if you don’t have perfect precision you’re racist. Why would anyone bother?