If we would use classification accuracy as the main metric to optimize, the classifier could achieve high accuracy because it would classify all users as non-churned. If there are 90% of non-churned users in the dataset, it would achieve 90% classification accuracy.