presentsUNAIVERSE
The UN AI timeline

    Learn AI /Model Quality and Evaluation /False positives and false negatives

    False positives and false negatives

    Also called: false positive, false negative, true positive, true negative, type I error, type II error.

    Detail

    In five seconds

    What it is

    False positives are wrong yes predictions, while false negatives are wrong no predictions.

    In plain words

    There are four basic results when a system makes a yes-or-no call: correct yes, correct no, wrong yes and wrong no. The two wrong ones are false positives and false negatives, and they tell you what kind of mistakes the system makes.

    An everyday example

    Amina is editing meeting notes from Nairobi, Bogotá and Bangkok. One sentence says, “After the programme officers spoke to the finance team, they approved the request.” She must decide whether “they” refers to the programme officers. If she marks it as a match and that is correct, that is like a true positive; if she marks it as a match but the sentence really meant the finance team, that is like a false positive. If she marks it as not a match and she is right, that is like a true negative; if she marks it as not a match but “they” really did mean the programme officers, that is like a false negative.

    Where the story stops working

    This analogy is about human language ambiguity, where even people may disagree about the correct answer. In many evaluation tasks, there is a clearer reference label, although not always. Also, real AI systems usually process thousands of cases and produce summary rates, not one careful judgement at a time.

    Remember thisOverall accuracy can conceal the mistake that matters. Count false positives and false negatives separately, then assess their consequences and how the decision threshold shifts the balance.

    Understand it

    Understand it

    Why it matters

    A system can look good overall while still making the wrong kind of mistakes for your purpose. If you only hear that a tool is '90% accurate', you may miss that it wrongly flags many innocent cases, or misses many real ones. False positives and false negatives make those trade-offs visible. That helps non-specialists ask better questions: What harm comes from each kind of error? Who is affected? Are some groups hit more often? Is the threshold set for convenience, safety or fairness? Those questions matter in procurement, policy, programme design and day-to-day use.

    How changing the threshold changes false positives and false negatives

    After using this explainer, a reader should be able to say how moving a decision threshold changes true positives, false positives, true negatives and false negatives, and why that affects precision and recall.

    50
    • correctly flagged (true positive)
    • wrongly flagged (false positive)
    • missed (false negative)
    • correctly left alone

    The explainer shows a set of scored items split into actual positives and actual negatives, plus a movable threshold. As the threshold moves, items on one side are labelled predicted positive and items on the other side predicted negative. Counters update for true positives, false positives, true negatives and false negatives, along with linked precision and recall values.

    Users can drag the threshold, use left and right arrow keys on a linked slider, and switch between a few preset class balances to see how the same model behaviour can produce different counts and score trade-offs.

    Describe this without the diagram

    Screen-reader version: 'This simulation explains a yes-or-no classifier. Items have scores and a threshold decides which items count as predicted positive. If an actually positive item is above the threshold, it is a true positive. If an actually negative item is above the threshold, it is a false positive. If an actually negative item is below the threshold, it is a true negative. If an actually positive item is below the threshold, it is a false negative. Lowering the threshold usually increases recall by finding more actual positives, but it can also increase false positives and lower precision. Raising the threshold usually does the opposite. A text table reports the current four counts and the resulting precision and recall.'

    With reduced motion: Reduced-motion version: the same score axis and threshold appear without animated movement. When the threshold changes, the display updates instantly and a static table beneath it shows the current true positive, false positive, true negative and false negative counts, plus precision and recall.

    Many AI and non-AI systems make a yes-or-no prediction: spam or not spam, relevant or not relevant, risk flag or no flag. To judge that kind of system, you compare its prediction with what was actually true.

    That comparison gives four possible outcomes. A true positive means the system predicted “yes” and the real answer was yes. A true negative means it predicted “no” and the real answer was no. A false positive means it predicted “yes” but the real answer was no. A false negative means it predicted “no” but the real answer was yes.

    The words matter because the two kinds of mistakes can have very different consequences. A false positive may wrongly include, accuse or interrupt. A false negative may miss something important. Which mistake matters more depends on the context. In health screening, security checks, content moderation, search, or humanitarian triage, the safer balance is not always the same.

    You may also hear type I error and type II error. In many testing contexts, type I error corresponds to a false positive, and type II error corresponds to a false negative. But those labels come from statistics and can confuse people outside that setting, so it is often clearer to say false positive or false negative directly.

    These terms are related to accuracy, precision and recall, but they are not the same. Accuracy is an overall score. Precision looks at how many predicted positives were actually positive. Recall looks at how many real positives were successfully found. The counts of true and false positives and negatives are the building blocks used to calculate those scores.

    How “Guessing who “they” refers to” maps onto the real thing

    In the storyIn the AI system
    Amina marks 'they' as referring to the programme officersThe system predicts a positive
    The sentence really does or does not refer to the programme officersThe ground truth is positive or negative
    A correct match or correct rejectionA true positive or true negative
    A wrong match or a missed matchA false positive or false negative

    A worked example

    At Headquarters, an information management officer evaluates a classifier that identifies service desk tickets reporting a system outage. The test uses 100 previously labelled, de-identified tickets in an approved environment.

    Starting from: Twenty tickets actually concern outages. The classifier flags 25 tickets, of which 15 are genuine outage reports and 10 concern other issues.

    1. Compare the classifier's label for each ticket with the reference label assigned during review.
    2. The 15 outage tickets correctly flagged are true positives. The 10 other tickets wrongly flagged as outages are false positives.
    3. Five of the 20 actual outage tickets were not flagged, so they are false negatives.
    4. The remaining 70 tickets concern other issues and were correctly left unflagged, so they are true negatives.

    Result: The confusion matrix contains 15 true positives, 10 false positives, 5 false negatives and 70 true negatives. The classifier finds 15 of 20 actual outages, but two out of every five tickets it flags are not outages.

    In a United Nations workplace

    During a records retention review, a national officer in a procurement unit tests a classifier that flags closed procurement files as eligible for disposal. The officer compares each prediction with the retention decision recorded on the file.

    A false positive is a file wrongly flagged for disposal, potentially removing a record that must be retained. A false negative is an eligible file left unflagged, prolonging storage and later review. Lowering the decision threshold may reduce missed eligible files while increasing incorrect disposal flags.

    Before you try thisProcurement files may contain commercially sensitive or personal information. They must not be uploaded to a tool that is not approved for that material. No file should be destroyed solely because of a model prediction. Apply the authorised retention schedule and require the designated records review and clearance.

    Where you may hear it

    • A vendor demo during a procurement meeting
    • An ICT governance board reviewing test results
    • A first reading of an evaluation report
    • A programme officer discussing a model's decision threshold with the evaluation office

    Avoid confusion

    Avoid confusion

    Often confused with

    Side-by-side

    Common misconceptions

    • A high accuracy score means false positives and false negatives are not a problem.

      Not necessarily. If the positive cases are rare, a system can get high accuracy by mostly saying 'no'. It may still miss many real positives, or it may create too many false alarms. You need to inspect the four outcome counts, not just one overall score.

    • False positives are always less serious than false negatives.

      The impact depends on the setting. A false positive in an airport bag check may mean an extra manual review. A false positive in a benefit, migration or safety-related system could wrongly interrupt services or trigger harmful scrutiny. A false negative can also be serious because it misses a real issue. Context decides the risk.

    • Type I error and type II error are different ideas from false positives and false negatives.

      They are closely related labels from statistics. In many common uses, type I error means false positive and type II error means false negative. The caution is that the statistical framing comes from hypothesis testing, so using the plain terms is often clearer for everyday system evaluation.

    Quick check

    A document filter marks 40 files as relevant. Later review shows that 12 of those 40 were actually not relevant. What are those 12 cases?

    Go deeper

    Go deeper

    The simple version

    There are four basic results when a system makes a yes-or-no call: correct yes, correct no, wrong yes and wrong no. The two wrong ones are false positives and false negatives, and they tell you what kind of mistakes the system makes.

    The technical version

    For a binary classification task, each case has a predicted label and a reference label. Comparing them yields four cells: true positive, false positive, true negative and false negative. These counts are often arranged in a confusion matrix. They are the basis for evaluation metrics such as accuracy, precision and recall: accuracy uses all four cells, precision focuses on predicted positives, and recall focuses on actual positives. The decision threshold often controls the trade-off: lowering it tends to increase true positives and false positives together, while raising it tends to increase true negatives and false negatives together. In imbalanced datasets, the raw counts and the cost of each error type may matter more than a single headline score.

    Reviewed external resources

    Where this came from

    Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.

    The one thing to remember

    Overall accuracy can conceal the mistake that matters. Count false positives and false negatives separately, then assess their consequences and how the decision threshold shifts the balance.