Learn AI /How Machines Learn /Label

Label

Starter
How Machines Learn
Technical readers
Programme & data teams

Also called: target, labels.

Detail

In five seconds

What it is

The expected answer attached to an example so a model can learn what output should match that input.

In plain words

A label is the answer attached to an example so the system can learn what output it should produce.

An everyday example

A maintenance team reviews daily printouts from a water pump. On each sheet, they circle the final judgement: “normal”, “needs inspection”, or “urgent fault”. A trainee studies many old sheets, looking at the readings and the circled judgement together. Later, when shown a new sheet without a circle, the trainee tries to give the same kind of judgement.

Where the story stops working

A human trainee can use outside knowledge, ask questions, and notice context not written on the sheet. A machine learning model does not do that by default. It only learns statistical patterns from the training data it receives, and the labels may themselves contain mistakes or disagreement.

Remember thisA label is the expected answer attached to a training example. It is distinct from the input features and from the prediction a trained model later produces.

Understand it

Why it matters

If you work with any machine learning system, labels affect what the system can learn, how well it performs, and where it may fail. Poor labels can teach the wrong pattern. Missing labels can limit which methods are possible. Inconsistent labels can make results unstable. For managers and programme staff, asking 'Where did the labels come from, who assigned them, and how reliable are they?' is often more useful than asking only which algorithm was used. Labels also help you judge whether a system’s output can be trusted for the task, or whether human review remains essential.

How a label guides learning

After viewing, the reader should be able to say that a label is the expected output paired with an input and used during training to measure error.

01
1. Start with an example
Show one training example split into input details and expected answer.
Draw a card labelled 'Training example'. On the left list input features such as 'vibration 8.1, temperature 74, flow 42'. On the right place a bold tag 'Label: needs inspection'.
02
2. Compare output with label
The model produces a trial output and checks how far it is from the label.
Add a box labelled 'Model in training'. Arrow from the example card into the box. From the box, draw an output tag 'trial output: normal'. Below, show a comparison badge 'does not match label'.
03
3. Adjust the model
The mismatch is used to update internal settings so similar cases are handled better next time.
Draw a feedback arrow from the comparison badge back into the model box with the text 'update settings'.
04
4. Use the trained model
For a new input, the model receives no label and produces a prediction.
On the right, draw a new card labelled 'New example' with input features only. Arrow into a box labelled 'Trained model'. Output tag says 'Prediction: needs inspection'. Add a small note beneath: 'No label available yet'.

A left-to-right flow with one example moving through training and then a new example moving through use. The diagram highlights that labels are present during training but absent when the trained model later makes a prediction on new data.

Hover or tap each stage to highlight whether a label is present there: present in steps 1 and 2, absent in step 4.

Describe this without the diagram

The diagram shows a four-stage flow. First, a training example contains input features and a label, which is the expected answer. Second, a model in training produces a trial output and compares it with the label. Third, the mismatch causes the model to update its internal settings. Fourth, a new example with no label is sent to the trained model, which produces a prediction. The key idea is that labels guide learning during training, but they are not usually available at the moment of prediction.

With reduced motion: A static left-to-right diagram with four labelled boxes and arrows. Labels appear only in the training part, while the final prediction stage explicitly notes that no label is provided for the new example.

In learning from examples, each input is paired with an expected output. That expected output is the label. If the input is a sensor reading, the label might be “normal” or “fault”. If the input is a house record, the label might be the sale price.

What goes in is a set of examples. Each example usually has features, meaning the measurable details the model uses, and a label, meaning the answer it should learn to produce. During training, the system compares its own output with the label and adjusts its internal settings to reduce the mismatch.

What comes out later is a prediction for new inputs. In supervised learning, labels are the teaching signal that lets the model connect patterns in the features to desired outputs. Without labels, the system may still find structure, but that is a different approach, such as clustering in unsupervised learning.

A label is often confused with a feature, a prediction, or ground truth. A feature is part of the input. A prediction is the model’s output after learning. Ground truth usually means the best available reference answer, often used as a label, but the two terms are not always identical because some labels are noisy, incomplete, or based on human judgement.

How “Marked answer sheets for pump checks” maps onto the real thing

In the story	In the AI system
pump reading sheet	input example
circled judgement on the sheet	label
trainee studying many past sheets	training
judgement on a new uncircled sheet	prediction

A worked example

A political affairs officer documents a text-classification pilot in a handover note. The training set contains paragraphs from cleared notes for the file, each reviewed and assigned an action category.

Starting from: Paragraph: “The delegation asked the office to circulate revised talking points by close of business.” Reviewed category: “follow-up required”.

Treat the paragraph as the input example and the reviewed category “follow-up required” as its label.
During training, the model uses patterns in the paragraph as features and produces an initial output, perhaps “for information”.
The loss function measures the difference between that output and the attached label. Training adjusts the model to reduce such differences across many examples.
For a new paragraph, no reviewed label is supplied. The model produces a prediction, which an authorised officer may confirm or correct.

Result: “Follow-up required” is the label for the training example. It is the expected answer used to teach the model, while the category produced for a new paragraph is a prediction.

In a United Nations workplace

Before completing a handover note for an incoming colleague, a political affairs officer in a country office records how an approved pilot classifies paragraphs from cleared notes for the file. Authorised reviewers have attached one category to each past paragraph, such as “follow-up required”, “for information” or “include in the next briefing note”. Those categories are the labels used during training.

The paragraph is the input and its reviewed category is the label. During training, the model compares its output with that category. When it later processes a new paragraph, its proposed category is a prediction, not a label, however confidently it appears on screen.

Before you try thisConfidential political reporting, names of contacts and security-sensitive details must not be entered into a tool that has not been approved for such material, because disclosure could expose individuals or compromise the office’s work. Even in an approved system, apply access and retention controls, and keep authorised staff responsible for the final assessment.

Where you may hear it

A review of an annotated dataset with the information management officer
A vendor demo during a procurement meeting
A model evaluation meeting before an internal pilot goes live
A handover briefing on a document-classification pilot

Avoid confusion

Often confused with

Common misconceptions

Common misconception: A label is just another input field in the data table.
Correction: Not quite. Input fields used to make the prediction are usually called features. The label is the output the model is trying to learn to predict. In a table they may look similar, but they play different roles in training.
Common misconception: Labels are always objective facts.
Correction: Some are factual, like a measured price. Others depend on human judgement, policy definitions, or limited information, such as whether text is 'harmful' or 'aligned'. That means labels can contain disagreement, bias, or error.
Common misconception: If data has labels, the model’s prediction will match them exactly.
Correction: Labels guide learning, but they do not guarantee perfect results. The model may see ambiguous cases, too little data, noisy labels, or examples unlike the training set. A prediction is an estimate, not the label itself.

Quick check

In a dataset for predicting pump faults, which item is the label?

Go deeper

The simple version

A label is the answer attached to an example so the system can learn what output it should produce.

The technical version

In supervised learning, a label is the target variable associated with an input example. The model maps features to this target by minimising a loss function, which measures the difference between the model output and the label. Labels may be categorical, as in classification, or numeric, as in regression. They are often created through annotation or derived from recorded outcomes. A label is not the same as a feature, and it is not the same as a prediction produced at inference time. 'Ground truth' is often used for labels when they represent the best available reference answer, but in practice training labels can be noisy, delayed, weak, or only approximate.

Reviewed external resources

EncyclopaediaLabeled data (opens in a new tab)Wikipedia
This link gives a general overview of labelled data and suits readers who want a quick reference-style introduction.
Intermediatelink checked 2026-07-31
CourseMachine Learning Crash Course (opens in a new tab)Google
This course link suits readers who want a practical beginner-friendly explanation of how labels are used in machine learning workflows.
Intermediatelink checked 2026-07-31
VideoStatQuest (opens in a new tab)StatQuest with Josh Starmer
This video source suits readers who learn well from step-by-step spoken explanations and visual teaching.
Starterlink checked 2026-07-31
VideoVideo explanations: Label (opens in a new tab)YouTube search
This search results page helps readers find multiple video explanations of labels and suits those comparing different teaching styles.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Label (opens in a new tab)Reddit search
This search results page helps readers see how people discuss labels in practice and suits those who want informal examples and common questions.
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

A label is the expected answer attached to a training example. It is distinct from the input features and from the prediction a trained model later produces.