We recommend treating the calculations as a general guidepost rather than a source of absolute truth. One common source of positivity training data is movie reviews - words that appear in reviews for poorly-rated movies tend to be more negative! Because there's no 'objective' set of data that can translate how positive, polite, or subjective a set of words is, the machine learning techniques look for similarities between your writing and writing samples that come with an approximate numerical score for positivity, politeness, or subjectivity. Politeness, positivity, and subjectivity rely on machine learning techniques that compare your writing to other samples of text that were used as training data. In those cases, trust your judgment rather than the calculation. In any individual email, the machine learning can arrive at answers that do not make sense. You've discovered one of the limitations of artificial intelligence in its current form.