If a neural network (or any other AI algorithm) includes race as one of its input variables, and through its training set optimization utilizes that trait in or

If a neural network (or any other AI algorithm) includes race as one of its input variables, and through its training set optimization utilizes that trait in order to deduce who may be guilty of a crime, and the cumulative error is indeed significantly improved by allowing that trait as an input -- then we have the need to rationally define exactly why this is bad, and what exactly the AI is missing which is needed to address that problem. I'd imagine the answer is something along the lines of the AI needing to incur a strong penalty when it falsely accuses or falsely casts suspicion, or misses a crime it would've found without that trait distracting it, etc. But the point IMO is that the only way to stop this problem is to understand it without being afraid of talking about it. You talk of the AI as being stupid for this problem -- but the stupidity lies in the axioms we're feeding it, and that only gets fixed by finding out where WE are going wrong in a rigorous and scientific way.