Every yes-or-no system makes two kinds of mistake, and they rarely cost the same. A smoke alarm that goes off when you burn toast is a false positive: annoying, cheap. A smoke alarm that stays silent during a real fire is a false negative: catastrophic. In medicine, a screening test that flags a healthy patient (false positive) leads to an extra test; one that misses a disease (false negative) can delay treatment. In statistics the same pair is called type I and type II errors.
Most classifiers output a score and a threshold turns it into yes or no. Moving the threshold trades one error for the other, as Google’s course shows: raising the score needed to call an email spam lets more spam through; lowering it sends more real mail to the spam folder. There is usually no setting that eliminates both. The counts sit in a confusion matrix, and precision and recall are the standard way to summarize how many alarms are false and how many real cases are missed.
Choosing the balance is a business and ethical decision, not just a technical one. Fraud teams, content moderators and security operations centers decide how many false alarms staff can review versus how many misses they can accept; too many false positives cause “alert fatigue,” where people stop trusting the system. The same framing applies to AI products: a guardrail that blocks harmless requests is producing false positives, and a hallucination detector that misses invented facts is producing false negatives. For rare events, see imbalanced datasets.
Sources
- Google Machine Learning Glossary: false positive (FP) — Defines a false positive as a model mistakenly predicting the positive class, such as classifying a non-spam email as spam; checked October 24, 2026.
- Google Machine Learning Glossary: false negative (FN) — Defines a false negative as a model mistakenly predicting the negative class, such as missing a spam email; checked October 24, 2026.
- Google Machine Learning Crash Course: Thresholds and the confusion matrix — Shows that raising the classification threshold lowers false positives and raises false negatives, and that the right balance depends on the costs of each error; checked October 24, 2026.
