“... bias in training materials is inevitable and there are better ways to deal with adverse effects than to scrub them out of existence. Efforts to eliminate bias would simply create new kinds of bias and distort the validity of the original datasets. Bias can be unjust by some standards but serves a useful purpose. ... the way to deal with it is to use other systems developed by those not involved in the original code, assisted by subject matter experts who could spot damaging bias in AI output and mitigate it in a commonsense manner. The way to deal with bias is not to eliminate it (you can't) but to identify bias and take it into account. Again, education and accountability are better than censorship and new biases.”