Is a false positive always more costly than a false negative?

No. Relative costs depend on the task and the action taken.

Does a 1% false positive rate mean 99% of alerts are correct?

No. It concerns actual negatives; alert correctness is measured by precision and also depends on prevalence.

Can human review prevent harm from a false positive?

It can reduce consequences, but the review still consumes time and may introduce delay or its own errors.

AI Glossary / Data science and analytics / Beginner

False positive

Definition

A false positive occurs when a classifier predicts the positive class for a case whose reference label is negative. It is an incorrect positive detection or false alarm relative to the task's class definition. For a system that flags damaged items, an undamaged item flagged as damaged is a false positive. Positive names the target class; it does not mean that the outcome is beneficial.

What it means

A false alarm can consume review time, interrupt a legitimate activity, or cause an unnecessary follow-up. Its importance depends on what the system does after the prediction. A flag sent to a reviewer has different consequences from an automatic rejection, even if the underlying classification is identical. Evaluation therefore needs both a clear reference label and an account of the action triggered by the positive prediction.

How it works

Compare predictions with reference labels and count actual negatives predicted positive as FP. Count actual negatives correctly predicted negative as TN. The false positive rate is FP divided by FP plus TN. Precision is TP divided by TP plus FP and answers a different question: how many positive predictions are correct? Raising a threshold for predicting positive generally reduces or leaves unchanged the number of false positives on the same scored dataset, but it can also increase misses. The right operating point depends on the task and available evidence.

Three practical examples

Email handling: a legitimate message is classified as unwanted mail and diverted from the recipient's main inbox.
Quality inspection: an acceptable product is flagged for a defect that reference inspection does not find.
Content review: a permitted message is flagged as violating a defined policy and sent for unnecessary review.

Think of it this way

Think of a smoke alarm sounding when no fire condition it is meant to detect is present. The alert is positive, while the relevant underlying condition is negative. The analogy illustrates a false alarm. Real classification tasks can have disputed or delayed labels, and an alarm's usefulness also depends on how quickly and appropriately people respond.

Worked example

A fictional classifier checks 10,000 records for a rare issue. Reference review identifies 100 actual positives and 9,900 actual negatives. The model correctly flags 90 positives and falsely flags 99 negatives. Its false positive rate is 99/9,900 = 1%. There are 189 positive predictions in total, so precision is 90/189, approximately 47.6%.
Thus a seemingly small 1% false positive rate can coexist with more false alarms than true alerts when the negative class is much larger. The proportion of alerts that are false is 99/189, approximately 52.4%, which is not the false positive rate. The model also misses 10 positives, a separate error type. These counts make the review workload concrete: staff face 189 alerts, of which 99 are unnecessary under the reference definition. The invented example illustrates prevalence and denominators, not the performance of a real product.

Illustrative case study

In a fictional document platform, an automated check flags records with unusual formatting. Reviewers find that many flags concern valid templates from a particular partner. The team checks the label definition, evaluates performance by template type, and investigates whether training examples adequately represent those formats. It considers changes to the model and review process while monitoring missed genuine issues. Simply suppressing the partner's alerts would hide possible real problems, so any adjustment receives a suitable evaluation. The case illustrates investigation of false alarms without claiming an actual reduction.

Common misconceptions and mistakes

False positive rate is not the fraction of alerts that are wrong; those metrics use different denominators. A positive score is not necessarily a calibrated probability, and an alert is not automatically proof of the condition. Eliminating every false positive by never predicting positive would also eliminate useful detections. Low false-alarm rates measured on an unrepresentative test set may not hold in deployment. Labels can be incomplete or inconsistent, so apparent model errors sometimes reveal a reference-review problem that also needs investigation.

Business use

Businesses should measure the cost and volume of unnecessary actions, not only a headline rate. Review capacity can be overwhelmed when a rare target is sought in a large stream, even with a low false positive rate. Track precision, recall, and the total alert volume alongside the false positive rate. Evaluate how often users experience repeated false alarms and whether those alarms delay legitimate work. Threshold or workflow changes should be compared against both missed cases and unnecessary interventions, using data representative of the intended operation.

Check your understanding

A model has 99 false alerts among 9,900 actual negatives and 189 alerts overall. What is its false positive rate?

Answer: 99/9,900 = 1%. The fraction of alerts that are false is 99/189, about 52.4%, a different measure.

Related terms

Learn next

Build a confusion matrix for a rare target and calculate both false positive rate and precision. Change the class prevalence while holding rates fixed to see why alert quality changes. Then study threshold selection and workflow costs so you can evaluate a detector in the context of the decisions it triggers.

Sources

Google Machine Learning Crash Course, classification metrics: https://developers.google.com/machine-learning/crash-course/classification/accuracy-precision-recall?hl=en . Google, thresholds and confusion matrix: https://developers.google.com/machine-learning/crash-course/classification/thresholding . Google, ROC and AUC: https://developers.google.com/machine-learning/crash-course/classification/roc-and-auc . Primary educational excerpts were consulted for error definitions, denominators, and threshold tradeoffs. Examples are original illustrations; editorial review remains pending.

Previous: False negative rate Next: False positive rate