ensuring dissenting opinions are present in the data set so that there is a proper diversity of opinions 

a method that adds noise to training data to obscure individual data points, thus protecting the privacy of individuals in the data set

teaching the AI to be polite and non-combative in its interactions with people, no matter the provocation

creating adversarial robustness

introducing made up examples into AI training sets in order to identify likely mistakes