Learn AI /Model Quality and Evaluation /Precision
Precision
- Intermediate
- Model Quality and Evaluation
- Programme & data teams
- Human rights practitioners
Also called: positive predictive value.
Back to the milestone you came from
In five seconds
What it is
The share of flagged results that are actually correct positives.
In plain words
When the system says ‘this one is a match’, precision tells you how often that positive signal is correct.
An everyday example
A phone places 10 photographs in a folder labelled “blurry”. On inspection, 8 are genuinely blurry and 2 are clear photographs taken in low light. The phone’s precision for that folder is 8 out of 10, or 80%. This says nothing about blurry photographs left elsewhere in the gallery.
A librarian may use judgement, context and follow-up questions, while a model applies learned patterns from data. Also, in a real search system the threshold for what counts as “positive” can be adjusted numerically, which the trolley story does not show well.
Remember thisPrecision is the share of positive flags that are correct. It does not measure how many real positives were missed, and it can change with the threshold or data being assessed.
Understand it
Understand it
Why it matters
A system can look useful while still producing many false alarms. Precision helps you judge whether a positive result is trustworthy enough for your task, especially when acting on a wrong alert has costs.
Precision and false alarms
After using this explainer, a reader should be able to say that precision measures how many flagged positives are actually correct, and how it changes when false positives rise or fall.
- correctly flagged (true positive)
- wrongly flagged (false positive)
- missed (false negative)
- correctly left alone
An interactive panel shows a grid of cases split into true positives, false positives, true negatives and false negatives. Users adjust a threshold and class balance, then watch the predicted-positive group change and see precision update as a fraction and percentage.
Users can move a threshold slider, switch class balance presets, and compare stricter versus looser settings while the counts and precision update immediately.
Describe this without the diagram
Screen-reader version: You are given a set of cases with real labels and system predictions. Precision is calculated from only the cases predicted positive. The explainer reports the current counts of true positives and false positives, then reads the formula precision equals true positives divided by true positives plus false positives. When the threshold becomes stricter, fewer cases are flagged positive. If false positives drop faster than true positives, precision rises. If many incorrect cases are flagged, precision falls.
With reduced motion: For reduced motion, show two static states instead of animation: one looser threshold and one stricter threshold. Each state lists counts for true positives, false positives, true negatives and false negatives, with the precision formula written out and the predicted-positive group clearly outlined.
Precision looks only at the items a system marked as positive. It asks: out of those flagged items, how many were truly the thing we were looking for?
The input is a set of predictions and the real outcomes. For each case, the system either flags it as positive or does not. Then we compare those flags with reality.
A true positive means the system flagged something and it really was positive. A false positive means the system flagged something but it was not actually positive. Precision uses these two counts: true positives divided by all flagged positives.
So precision tells you how trustworthy a positive result is. High precision means few false alarms among the positives. It does not tell you how many real positives were missed; that is recall, a different measure.
How “The blurry-photo folder” maps onto the real thing
| In the story | In the AI system |
|---|---|
| Documents on the trolley marked relevant | Items the system predicted as positive |
| 8 documents that truly fit the request | True positives |
| 2 documents that do not actually fit | False positives |
| Checking only the selected stack | Calculating precision from predicted positives only |
A worked example
The procurement officer manually checks the product sheets flagged by the classifier as confirming both live captioning and keyboard navigation.
Starting from: The classifier flags 12 product sheets. Nine explicitly confirm both features. Three contain general accessibility language but do not confirm both requirements.
- Count all flagged product sheets: 12 predicted positives.
- Count the flagged sheets that confirm both required features: 9 true positives.
- Count the flagged sheets that do not confirm both features: 3 false positives.
- Calculate precision as TP divided by TP plus FP: 9 divided by 12, or 0.75.
Result: Precision is 75%. In other words, three quarters of the classifier's positive flags were correct. The calculation does not reveal how many product sheets confirming both features were left unflagged.
In a United Nations workplace
At a UN Country Team working group, a procurement officer is finalising the accessibility requirements in the concept note for a side event in the margins of the session. With clearance due by close of business, the officer tests a classifier that screens event-platform product sheets for explicit confirmation of live captioning and keyboard navigation, rather than a general claim that the platform is accessible.
High precision means that most product sheets flagged for follow-up actually confirm both required features. It does not show how many suitable platforms the classifier missed. The figure applies to the decision threshold and document mix tested, and may change when either changes.
Before you try thisDo not upload bids, pricing, evaluation comments or other procurement-sensitive material to a tool that has not been approved for it. This could expose confidential commercial information and compromise the integrity of the procurement process.
Where you may hear it
- During a vendor demo to a procurement unit, when staff ask how many automated flags survive manual review
- At an evaluation office meeting reviewing the results of a document classifier
- During a release-readiness review for a search tool
- At an ICT governance board meeting considering a change to a classifier's decision threshold
Avoid confusion
Avoid confusion
Often confused with
Accuracy
The share of cases a system labels correctly out of all the cases it was tested on.
Recall
The share of real positive cases that a system successfully flags.
Side-by-side
Common misconceptions
Common misconception: High precision means the system finds most of the real positives.
Correction: Not necessarily. A system can be very selective, flag only a few cases, and get most of those right. That gives high precision but may miss many true cases, which is low recall.
Common misconception: Precision and accuracy are basically the same.
Correction: Accuracy looks at all predictions, both positive and negative. Precision looks only at predicted positives, so it is often more informative when positive flags trigger action.
Common misconception: If precision is 90%, then 90% of all items are positive.
Correction: Precision is about the quality of the flagged positive set, not the share of positives in the whole dataset.
Quick check
A system flags 20 messages as urgent. Human review shows that 15 really are urgent and 5 are not. What is the precision?
Precision uses only flagged positives: true positives divided by all predicted positives. Here that is 15 divided by 20, which equals 75%. The unflagged messages do not enter this calculation.
Go deeper
Go deeper
The simple version
When the system says ‘this one is a match’, precision tells you how often that positive signal is correct.
The technical version
For binary classification, precision, also called positive predictive value, equals TP divided by TP plus FP. It measures the proportion of predicted positives that are true positives, and it depends on the decision threshold and the data distribution.
Reviewed external resources
Primary sourceAI Risk Management Framework (opens in a new tab)NIST
This framework helps readers connect measurement choices to risk management and governance, and suits practitioners evaluating AI systems in real organisations.
Intermediatelink checked 2026-07-31
EncyclopaediaPrecision and recall (opens in a new tab)Wikipedia
This page gives a broad written overview of precision and recall, suitable for readers who want definitions, formulas and examples in one place.
Intermediatelink checked 2026-07-31
CourseClassification: Accuracy, recall, precision (opens in a new tab)Google
This course-style resource explains precision alongside accuracy and recall with practical teaching examples, and suits beginners who want a guided introduction.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Precision (opens in a new tab)YouTube search
This is a search results page for videos about precision, useful for readers who prefer visual explanations and want to choose a teaching style that fits them.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Precision (opens in a new tab)Reddit search
This is a search results page for community discussions about precision, useful for readers who want to see common questions, confusions and informal explanations.
search results, not one pagelink checked 2026-07-31
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
Precision is the share of positive flags that are correct. It does not measure how many real positives were missed, and it can change with the threshold or data being assessed.