Machine learning uses patterns in data to make predictions or classifications. A retailer might use it to forecast demand from past sales and customer behavior. Its value lies in a well-defined task—not in understanding every situation the way a person can. section 10.3 section 10.1 section 7.2
The distinction matters in healthcare. Labeling chest images “infected” or “not infected” made machine learning seem like a quick route to COVID-19 diagnosis. Yet a review of 2,212 studies found no identified model suitable for clinical use because of methodological flaws or underlying biases. section 6.2.1
When a pattern gives the wrong answer
A pneumonia triage model classified patients with asthma as low risk. In its training data, those patients had lower mortality because they received aggressive care—not because asthma made pneumonia safer. The model learned a pattern without recognizing what caused it. Without prompt hospitalization, these patients faced substantially higher risk. section 10.3
Use machine learning for bounded predictions, then test whether its data and outputs reflect the real problem. Domain experts can spot missing context and dangerous conclusions. Pattern recognition alone does not supply common-sense reasoning or the causal understanding needed to replace professional judgment. section 10.3 section 10.4.5 section 7.2 section 8.0
