Where does a false positive sit among the four outcomes?
A binary classification produces four outcomes, and “positive” means the class the system is looking for, not a good result.
| Actually positive | Actually negative | |
|---|---|---|
| Predicted positive | True positive (TP): real fraud flagged | False positive (FP): legitimate payment flagged |
| Predicted negative | False negative (FN): real fraud missed | True negative (TN): legitimate payment passed |
A false positive is an error of commission: the system raised a flag it should not have. A false negative is the mirror error of omission. Every metric an evaluation reports, from accuracy to recall, is a ratio of these four cells.
What does the false positive rate mean?
The false positive rate is FP / (FP + TN): the share of real negatives the system wrongly flags (Google Machine Learning Crash Course, last updated 12 January 2026). It is easy to misread when positives are rare. Take a hypothetical 100,000 transactions with 1 percent fraud and a 2 percent false positive rate: 99,000 legitimate payments yield 1,980 false alarms, against at most 1,000 real frauds. A low false positive rate can still fill a review queue mostly with noise when positives are rare. Reviewers feel the share of flags that are wrong, one minus precision, as in anomaly detection.
False positive vs false negative: which costs more?
It depends on who pays for each error, and the decision threshold sets the trade between them.
| Use case | Cost of a false positive | Cost of a false negative |
|---|---|---|
| Fraud alerts | Analyst time, a declined customer | Lost money, chargebacks |
| Document review | A reviewer re-reads a clean file | A risky clause gets signed |
| Guardrails | A refused legitimate request, users route around the tool | Harmful or leaked output reaches a user |
| Security alerting | Alert fatigue, real alerts ignored | An intrusion goes unnoticed |
Lowering the threshold trades false negatives for false positives, never removes both. Set it with the people who own each cost, and route the uncertain middle band to a person through confidence gating rather than forcing a yes or no.
How do you measure false positives in an AI system?
Put labeled negatives in the eval dataset. Many detection test sets hold only cases the system should catch, so they cannot measure false alarms. Add benign cases that sit near the boundary: legitimate prompts that look adversarial, clean contracts with unusual wording. A test set with no negatives cannot show a false positive problem, so over-blocking surfaces first as user complaints. The same applies to an LLM as a judge: a judge that fails a correct answer is producing false positives, so validate it against human labels on both passing and failing examples.