Why does someone need to spot check the training data against the output? I'd assumed the algorithm is trying to maximize a metric, which is best completely aut

Why does someone need to spot check the training data against the output? I'd assumed the algorithm is trying to maximize a metric, which is best completely automated without human intervention to slow it down or provide biased opinions. Even fuzzier goals like 'not be racist' would get turned into metrics (eg. proportion of words used that are on this watchlist or similar scoring algorithm)