Labeling a dataset for sentiment analysis

I was reading articles on sentiment analysis and NLP and there is something I cant quite understand. One of the methods to label a dataset is to use something like textblob with a polarity dictionary that would count words in a positive and negative dictionary and give a score based on it.
Then the dataset is used to train a classification algorithm. My question is, why do we bother with ML at all while we have a rule-based labeling method that we trust so much to the point we consider it as the ground truth?