Learn AI /Model Quality and Evaluation /Accuracy
Accuracy
- Starter
- Model Quality and Evaluation
- Programme & data teams
- Managers & front offices
Back to the milestone you came from
In five seconds
What it is
The share of cases a system labels correctly out of all the cases it was tested on.
In plain words
Accuracy tells you what fraction of all answers were right.
An everyday example
A basket contains 90 dark socks and 10 light socks. A sorter calls every sock dark. It is correct 90 times out of 100, so its accuracy is 90%. That sounds respectable until someone notices that it failed on every light sock. The common group dominates the score.
An examination usually gives each question similar importance and a fixed answer key. Real AI evaluation may involve uneven class sizes, uncertain labels, and situations where one type of mistake matters much more than another, so the same percentage can hide serious weaknesses.
Remember thisAccuracy is the share of all tested cases labelled correctly. When one class is much more common, the score can look strong while performance on the rarer class remains poor.
Understand it
Understand it
Why it matters
A single percentage can strongly influence whether leaders trust a system, buy a product, or change a workflow. If that percentage hides important mistakes, people may approve a tool that looks good on paper but fails where the consequences matter most.
How accuracy is calculated
After viewing this, a reader should be able to explain that accuracy is the number of correct predictions divided by the total number of cases, and why that can still hide problems.
01
Start with labelled cases
Use a test set where the correct answer for each case is already known.
Draw 10 small cards in a grid, each with a true label badge such as A or B.
02
Compare predictions
For each case, place the system's predicted label next to the true label.
Add a second badge to each card for the prediction, aligned beside the true label.
03
Count matches
Mark each case as correct if prediction and true label match, otherwise incorrect.
Outline matching cards in green and non-matching cards in red; add a small tick or cross.
04
Compute the share
Divide the number of correct cases by the total number of cases.
Show a large equation: 8 correct / 10 total = 80%, with a small note under it about rare cases.
A four-step horizontal diagram. Each step is a rounded rectangle with a simple icon and short label. Thin arrows connect left to right. Step 1 shows a stack of 10 case cards labelled with true answers. Step 2 shows the model's predicted labels added to each card. Step 3 highlights cards in green for matches and red for mismatches. Step 4 shows the fraction 8 correct out of 10 total, then 80%. Below the final step, a warning note says: 'Useful summary, but check whether rare important cases were missed.'
Hovering or tapping a card reveals its true label, predicted label, and whether it counts as correct. A toggle switches to an imbalanced example where high accuracy still misses all rare cases.
Describe this without the diagram
This diagram explains accuracy in four steps. First, start with test cases that already have correct answers. Second, compare the system's prediction for each case with the known correct answer. Third, count how many predictions match and how many do not. Fourth, divide the number of correct predictions by the total number of cases to get a percentage, such as 8 out of 10 equals 80%. The diagram also warns that this percentage can be misleading if important cases are rare.
With reduced motion: A static four-panel diagram shows the same sequence without animation: labelled test cases, predictions beside labels, green and red match markings, and a final fraction converted to a percentage with a caution note about rare cases.
Accuracy is a simple score for checking a system’s predictions. You compare what the system predicted with the correct answers, then count how many times it got the answer right.
What goes in: a set of cases, the system’s predicted labels for those cases, and the true labels known from reliable checking. What happens: each case is marked as correct or incorrect. What comes out: one number, usually written as a percentage.
For example, if a system checks 100 messages and labels 92 of them correctly, its accuracy is 92%. This sounds straightforward, and sometimes it is useful.
But accuracy can mislead when one outcome is much more common than another. If 95 out of 100 messages are routine, a weak system could label everything as routine and still reach 95% accuracy while completely missing the unusual but important cases. That is why accuracy should be separated from precision, recall, and the pattern of false positives and false negatives.
How “Sorting dark and light socks” maps onto the real thing
| In the story | In the AI system |
|---|---|
| the student's answers | the system's predictions |
| the official answer sheet | the true labels used for evaluation |
| counting how many answers match | counting correct predictions |
| 80% correct on 50 questions | accuracy as correct cases divided by all cases |
A worked example
At a regional hub, an information management officer evaluates a classifier that identifies paragraphs containing numerical claims before monitoring reports are uploaded to the portal.
Starting from: The reviewed test set contains 200 paragraphs: 180 without numerical claims and 20 with them. The classifier labels 190 paragraphs correctly, but misses 10 of the 20 paragraphs containing numerical claims.
- Compare the classifier's label for each paragraph with the label assigned during human review.
- Count 190 correct labels out of 200 evaluated paragraphs.
- Calculate accuracy: 190 divided by 200 equals 95%.
- Check performance by class. The classifier found only 10 of the 20 paragraphs containing numerical claims, despite its 95% overall accuracy.
Result: The classifier has 95% accuracy, but it misses half of the less common numerical-claim cases. The headline score is therefore insufficient for deciding whether it can support the pre-upload check.
In a United Nations workplace
An internal auditor in a country office tests a classifier that marks claims in a draft statement as either supported or unsupported before the statement goes for clearance at close of business.
Accuracy shows the share of claims the classifier labels correctly against a set already reviewed by staff. It is a useful headline result, but a high score may simply reflect that most claims are supported. The auditor also needs to check how often the tool misses the less common unsupported claims.
Before you try thisDraft statements may contain sensitive or uncleared material. Do not enter the text into a tool that has not been approved for that material. Every flagged or unflagged claim still requires human review before clearance.
Where you may hear it
- A vendor demo in a procurement meeting
- An ICT governance board reviewing a pilot
- A results meeting on a model evaluation report
- An internal audit discussion on automated controls
Avoid confusion
Avoid confusion
Often confused with
Precision
The share of flagged results that are actually correct positives.
Recall
The share of real positive cases that a system successfully flags.
F1 score
A single score that balances how many positive cases a system finds with how often its positive predictions are correct.
Class imbalance
A dataset has this problem when one class appears far more often than another, which can make overall accuracy look strong while a model misses rare cases.
Side-by-side
Common misconceptions
Common misconception: High accuracy means the system works well in every important situation.
Correction: Accuracy averages all cases together. If some cases are rare or more important, the average can stay high even while performance on those cases is poor.
Common misconception: Accuracy and precision mean the same thing.
Correction: Accuracy asks how many total predictions were correct. Precision asks, among the items predicted as positive, how many really were positive.
Common misconception: An accuracy score from one test set proves future performance.
Correction: A score only reflects the data and conditions used for testing. Performance may change on new data, in another country, or after the real-world process changes.
Quick check
A fraud detector is tested on 1,000 transactions. Only 20 are actually fraudulent. The system labels every transaction as not fraudulent and gets 980 labels correct. What is the best conclusion?
If fraud is rare, a system can achieve high accuracy by always choosing the common class. Here it misses all 20 fraudulent transactions, so accuracy alone gives a misleading picture.
Go deeper
Go deeper
The simple version
Accuracy tells you what fraction of all answers were right.
The technical version
In classification, accuracy is the number of correct predictions divided by the total number of evaluated cases, often written as (true positives plus true negatives) over all cases. It is sensitive to class imbalance because common classes dominate the score.
Reviewed external resources
Primary sourceAI Risk Management Framework (opens in a new tab)NIST
This official framework helps readers place performance measures within wider AI risk management and suits programme and leadership audiences.
Intermediatelink checked 2026-07-31
EncyclopaediaAccuracy and precision (opens in a new tab)Wikipedia
This link gives a broad introductory explanation of accuracy and precision and suits readers who want a quick overview with general background.
Intermediatelink checked 2026-07-31
CourseClassification: Accuracy, recall, precision (opens in a new tab)Google
This course page explains how accuracy relates to recall and precision and suits beginners comparing common classification metrics.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Accuracy (opens in a new tab)YouTube search
This is a search results page for video explanations of accuracy, useful for readers who prefer different teaching styles and want to compare several presenters.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Accuracy (opens in a new tab)Reddit search
This is a search results page for community discussion of accuracy, useful for readers who want informal questions, examples and debates from practitioners.
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
Accuracy is the share of all tested cases labelled correctly. When one class is much more common, the score can look strong while performance on the rarer class remains poor.