Learn AI /Model Quality and Evaluation /Recall
Recall
- Intermediate
- Model Quality and Evaluation
- Programme & data teams
- Human rights practitioners
Also called: sensitivity, true positive rate.
Back to the milestone you came from
In five seconds
What it is
The share of real positive cases that a system successfully flags.
In plain words
Recall tells you how good a system is at catching the cases you do not want to miss.
An everyday example
A city office receives 200 public feedback forms after a transport consultation. Staff want to find every form that mentions blocked access for older persons or persons with disabilities. They use a keyword tool to pull out likely matches for review. If 40 forms truly mention blocked access and the tool finds 32 of them, its recall is based on those 32 found out of the 40 that really mattered.
A keyword tool is much simpler than many AI systems. Real models may use patterns learned from data, may behave differently on new data, and may involve a threshold that changes recall, so the office example does not show the full mechanics.
Remember thisRecall asks: among all the cases that were truly positive, what proportion did the system flag at the chosen threshold?
Understand it
Understand it
Why it matters
If a system misses too many real cases, a good-looking overall score can hide a serious failure. Recall helps people judge whether the system is safe enough for tasks where missed cases have real consequences.
Precision–recall threshold simulator
After using this explainer, the reader should be able to say that recall is the proportion of real positive cases found, and that changing the decision threshold can raise recall while also affecting precision.
- correctly flagged (true positive)
- wrongly flagged (false positive)
- missed (false negative)
- correctly left alone
Show a horizontal threshold slider above a row or scatter of scored items. Each item has a true label known to the explainer: positive items in one shape or colour, negative items in another. As the slider moves, items on one side of the threshold become ‘flagged’. Counters update live for true positives, false positives and false negatives, and the displayed recall value changes accordingly. A companion precision value can also update so readers can compare the two measures without confusing them.
Readers drag a threshold slider left or right, or use keyboard arrow keys on the focused range control, to see how the set of flagged items changes and how recall and precision update together.
Describe this without the diagram
This interactive shows a set of items with known true labels and model scores. You control a threshold that decides which items are flagged as positive. When you lower the threshold, more items are flagged. This usually catches more of the truly positive items, so recall often increases, but it can also flag more items that are not truly positive. The explainer reports the current counts of true positives, false positives and false negatives, and states recall as the number of true positives divided by the total number of truly positive items. Its main lesson is that recall measures missed real cases, not false alarms.
With reduced motion: Readers see the same scored items, threshold control and metric values, but changes happen without animation. When the threshold moves, the highlighted items and the recall and precision numbers update instantly in a static layout.
Recall measures how well a system finds the cases you care about. A “positive case” means the kind of item you want the system to catch, such as harmful content, fraud, or consultations that mention access barriers.
You start with items that already have a correct label, often called the ground truth. The system checks those items and marks some as positive. Recall looks only at the items that truly are positive, then asks: how many of those did the system catch?
The calculation is true positives divided by all real positives. True positives are cases that really were positive and were correctly flagged. If there are many false negatives, meaning real positives the system missed, recall goes down.
High recall matters when missing a real case is costly. But recall alone does not tell you how many false alarms the system creates. That is why it is often discussed alongside precision, which asks a different question.
How “Finding all the forms about blocked access” maps onto the real thing
| In the story | In the AI system |
|---|---|
| all forms that truly mention blocked access | all real positive cases |
| forms the tool pulls out | cases predicted as positive |
| 32 true matches found | true positives |
| 8 relevant forms left in the pile | false negatives |
A worked example
The duty officer evaluates the classifier against 100 passages from a cleared test set. Desk officers have confirmed that 20 passages contain commitments requiring follow-up.
Starting from: 100 test passages, including 20 actual positive cases.
- The classifier flags 25 passages at the selected threshold.
- Reviewers confirm that 16 of the flagged passages contain follow-up commitments. These are true positives.
- Because the test set contains 20 actual positives, the classifier missed 4 of them. These are false negatives.
- Recall is TP / (TP + FN), or 16 / (16 + 4) = 0.80.
Result: Recall is 80%. At this threshold, the classifier catches 16 of the 20 passages that require follow-up and misses 4. The 9 irrelevant passages it also flagged affect precision, not recall.
In a United Nations workplace
A duty officer in a regional bureau is finalising a note for the file following a bilateral. Before close of business, an approved classifier flags passages that may record commitments requiring follow-up by desk officers.
Recall shows what share of the passages confirmed by reviewers as containing follow-up commitments the classifier actually flagged. Low recall means genuine commitments are being missed, even if every passage it did flag appears relevant. The result applies at the threshold tested.
Before you try thisBilateral notes may contain confidential or politically sensitive material. They must not be uploaded to a public or otherwise unapproved tool. Testing must use an environment approved for that material, or suitably cleared or synthetic text, and the classifier must not replace review by the responsible desk officer.
Where you may hear it
- An ICT governance board reviewing test results before approving a pilot
- A vendor demonstration to a procurement unit for a document-screening tool
- An evaluation office discussing a model validation report
- An inter-agency working group setting the threshold for a referral tool
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.
Precision
The share of flagged results that are actually correct positives.
F1 score
A single score that balances how many positive cases a system finds with how often its positive predictions are correct.
Side-by-side
Common misconceptions
Common misconception: High recall means the system is generally accurate.
Correction: Not necessarily. A system can catch most real positive cases and still make many false alarms. Recall looks at missed real cases, not overall correctness.
Common misconception: Recall and precision are just two words for the same score.
Correction: They answer different questions. Recall asks how many real positive cases were found, while precision asks how many flagged cases were actually positive.
Common misconception: If recall is 90%, the system is right 90% of the time.
Correction: No. A recall of 90% means it found 90% of the real positive cases. It says nothing by itself about negatives or about how many flagged cases were wrong.
Quick check
A system reviews 50 documents. Ten truly contain a problem that should be flagged. It correctly flags 8 of those 10. What is its recall?
Recall is true positives divided by all real positives. Here that is 8 divided by 10, so the recall is 80%. The total number of documents does not matter for this calculation.
Go deeper
Go deeper
The simple version
Recall tells you how good a system is at catching the cases you do not want to miss.
The technical version
Recall, also called sensitivity or true positive rate, is TP / (TP + FN): the proportion of actual positive instances that the classifier labels positive at a given threshold.
Reviewed external resources
Primary sourceAI Risk Management Framework (opens in a new tab)NIST
This link gives a risk-management perspective on evaluating AI systems and suits programme, governance and assurance readers who need to place metrics in a wider decision context.
Intermediatelink checked 2026-07-31
EncyclopaediaPrecision and recall (opens in a new tab)Wikipedia
This link gives a broad reference-style overview of precision and recall and suits readers who want definitions, formulas and related concepts in one place.
Intermediatelink checked 2026-07-31
CourseClassification: Accuracy, recall, precision (opens in a new tab)Google
This link gives a short teaching explanation with examples and suits readers who want a quick practical introduction to accuracy, recall and precision.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Recall (opens in a new tab)YouTube search
This search-results page gives multiple video explanations of recall and suits readers who prefer to compare several visual teaching styles before choosing one.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Recall (opens in a new tab)Reddit search
This search-results page gives community discussions about recall and suits readers who want informal explanations, examples and common points of confusion, with care about varying quality.
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
Recall asks: among all the cases that were truly positive, what proportion did the system flag at the chosen threshold?