A false negative is a case that is actually positive but the model predicted as negative. The fraud went through. The disease was not caught. The harmful response passed the filter. It is one of the four cells of a confusion matrix, and in most production systems it is the expensive one.
Statisticians call it a Type II error. Biometrics calls it a false reject. Medicine calls it a miss. They all mean the same thing: the condition was there and the system said no.
The rate is straightforward:
false negative rate = FN / (TP + FN)That denominator is every actual positive in the data. The false negative rate is therefore 1 - recall, or equivalently 1 - sensitivity, depending on which vocabulary your team uses.
Key takeaways
- A false negative is a real positive that the model labeled negative. It is the FN cell of the confusion matrix and the same thing as a Type II error.
- False negative rate is
FN / (TP + FN), which is exactly1 - recall. Driving false negatives down and driving recall up are the same project. - False negatives and false positives almost never cost the same. In fraud, medical screening, and safety filtering the asymmetry can be orders of magnitude, and the decision threshold should reflect that.
- At a fixed model, you cannot reduce both error types at once. Lowering the threshold trades false negatives for false positives. Reducing both requires a better model, better features, or better retrieval.
- False negatives are systematically undercounted, because a miss produces no alert. You only learn about it from a delayed downstream signal such as a chargeback or a complaint.
Where the false negative sits
| Predicted positive | Predicted negative | |
|---|---|---|
| Actually positive | True positive | False negative |
| Actually negative | False positive | True negative |
Worked example. A model reviews 1,000 loan applications. Sixty of those borrowers will eventually default, and the model flags 45 of them as high risk. That leaves 15 false negatives, a false negative rate of 15 / 60 = 0.25, and a recall of 0.75. One quarter of the defaults were approved. Whether that is acceptable depends entirely on what a default costs relative to turning away a good borrower, which is a business question the metric cannot answer for you.
The asymmetry nobody should average away
Accuracy treats a false negative and a false positive as one mistake each. That is almost never true.
Fraud detection. A missed fraudulent transaction costs the chargeback, the goods, and the investigation. A blocked legitimate transaction costs a support ticket and some customer annoyance. The ratio is large and it runs one direction.
Medical screening. A missed diagnosis can mean a disease progresses untreated. A false positive usually means a follow-up test and a bad week. Screening programs are deliberately tuned toward high sensitivity for exactly this reason, then paired with a more specific confirmatory test.
Safety and content filtering. A false negative ships harmful output to a user. A false positive refuses a benign request. Both are real costs, and here the ratio is genuinely contested rather than obvious, which is why over-refusal has become its own failure mode.
Once you accept the asymmetry, the threshold follows from it. If a false negative costs C_FN and a false positive costs C_FP, the expected-cost-minimizing rule is to predict positive when the model’s probability clears C_FP / (C_FN + C_FP). Ten times more costly to miss than to over-flag gives a threshold near 0.09, not 0.5. Most teams never move off the default 0.5, which silently declares the two errors equally expensive.
Because that trade is continuous, the honest way to present a binary classification model is as a curve rather than a point. A precision recall curve shows the whole frontier, and the operating point becomes an explicit choice instead of a leftover default. The related view from the other column is precision, which counts false positives rather than false negatives.
Why false negatives hide
False positives announce themselves. Someone was blocked, someone escalated, a ticket got filed. The feedback arrives within hours and it arrives with a label attached.
False negatives are silent. Nothing happens, which is what a negative prediction looks like when it is correct too. You find out later and indirectly: a chargeback lands sixty days on, a customer reports something the filter should have caught, an auditor finds a pattern. By then the label is delayed, incomplete, and biased toward the misses that happened to be discovered.
Three practices reduce the blind spot.
- Sample and label the negatives. Pull a random sample of predicted-negative cases and have a human review them. It is the only unbiased estimate of the false negative rate you can get.
- Instrument the downstream outcome. Wire chargebacks, escalations, refunds, and complaints back to the prediction that preceded them, so a miss becomes traceable rather than anecdotal.
- Track the rate per slice. A stable overall false negative rate can hide one merchant category, one language, or one customer tier where the model has quietly stopped working. Comparing recall across groups has a name, recall parity, and it is the version of this check that holds up in a fairness review.
False negatives in LLM and agent systems
Every classifier inside an AI stack has a false negative rate, and most of them are unmonitored.
Retrieval. A retriever that fails to return the chunk containing the answer has produced a false negative. The model then answers from whatever it did get, and the failure surfaces as a confident wrong answer rather than as a retrieval error. Nothing throws, nothing retries, and the only way to count these is to ask questions whose answers you already know.
Evaluators. An LLM as a judge that flags hallucinations is a classifier, so it has false negatives of its own. A judge that misses half the hallucinations reports a clean dashboard while the system degrades. Any evaluator you rely on needs to be scored against a human-labeled set before you trust its output.
Tool selection and routing. An agent that should have called a tool and did not has made the same class of error. It is a false negative on the decision to act, and it tends to show up as an incomplete answer rather than an exception, which is part of why agent failures are hard to catch. The broader picture of how these components fit together is in the AI agent handbook, and the metric selection question sits with any classification model in the pipeline.
FAQ
What is the difference between a false negative and a false positive?
A false negative is an actual positive predicted as negative: the model missed something real. A false positive is an actual negative predicted as positive: the model raised an alarm on nothing. False negative is a Type II error, false positive is a Type I error. They sit in opposite cells of the confusion matrix and they respond in opposite directions to a threshold change.
How do you reduce false negatives?
Two paths. The cheap one is lowering the decision threshold so more cases get flagged, which reduces false negatives and increases false positives on the same model. The real one is improving the model: better features, more examples of the positive class, better retrieval, or a different architecture. Only the second reduces both error types at once, which is why threshold tuning should be treated as a policy choice rather than a fix.
Is the false negative rate the same as 1 minus recall?
Yes. Recall is TP / (TP + FN) and the false negative rate is FN / (TP + FN). The two share a denominator and sum to 1. Sensitivity and true positive rate are other names for recall, so FNR = 1 - sensitivity as well.
Why is accuracy a bad metric when false negatives matter?
Accuracy counts every mistake equally and is dominated by the majority class. On a dataset where 1% of cases are positive, a model that predicts negative for everything is 99% accurate and catches zero positives. Every one of those positives is a false negative. Recall, precision, and the false negative rate expose that; accuracy conceals it.
Example:
A model flags a credit card transaction as ‘not fraud’ when it is actually a fraudulent transaction.
