I am pretty sure that the problem is with the trainers and what they want to explicitly train the AI to answer. I assume the coding, itself, may be free of bias. Perhaps this works the same with children, too.