Viral '98% of truck drivers drive dangerously' stat is a model score, not a real rate

A widely shared claim that 98% of truck drivers drive dangerously originated from a PLOS ONE study built on a Kaggle dataset of 120,000 trip records collected in California in January 2023. The 98% figure is actually the machine learning model's classification accuracy — how often it correctly distinguished 'anomalous' driving rows from normal ones within the dataset. Model accuracy does not indicate how prevalent dangerous driving is in the real-world population, especially when class labels are imbalanced. For example, a model that labels every row as 'normal' can still achieve 95% accuracy if only 5% of rows are flagged as anomalous, while detecting nothing meaningful. The episode highlights a common misreading of AI metrics: accuracy, precision, recall, and AUC each measure different things, and none directly translates into a population-level behavioral rate.