AI Glossary / AI and machine learning / Beginner
Label
Definition
A label is the known answer attached to an example: the category, number or outcome the model is supposed to learn to predict, such as spam, churned or 42000.
What it means
Supervised learning is learning from answered questions. Every training example pairs the inputs with the answer, and that answer is the label: what the outcome actually was for this case.
Labels come in two broad kinds. Categories, fraud or not, high-medium-low, drive classification; numbers, next month's sales, the claim amount, drive regression. The kind of label decides the kind of model and the kind of error metric.
Labels are made, not found. Somebody or something must establish each answer: an investigator's verdict, a sensor reading, the observed fact that the customer did cancel. That production process is expensive, slow and error-prone, and it sets the ceiling on everything learned from it.
Label scarcity shapes strategy. Unlabeled data is usually cheap and abundant; labeled data is the expensive bottleneck, which is why whole industries exist around annotation, and why methods that squeeze more from fewer labels keep mattering.
For a business, the label is where the model's idea of truth is defined. Who decides the answer, by what rules, with what error rate: those choices are the foundation every later accuracy claim quietly stands on.
How it works
Define the label precisely: what outcome, measured when, by whom, with what tie-breaking and what minimum evidence.
Collect labels through a process with known quality: expert review, adjudication of disagreements, or trusted outcome feeds.
Store labels beside the features with timestamps, so training can match each input to the answer known at the right time.
Audit labels on a sample basis: measured label error belongs in every model evaluation as context for the accuracy figures.
Protect the label from contamination: nothing derived from the label may leak into the features, or the model learns to read the answer key.
Three practical examples
- A fraud team where confirmed investigator outcomes become the labels for the next model generation.
- A retailer labeling 20,000 support emails by intent, double-checked on a sample, to train the routing model.
- A demand planner using actualised sales as numeric labels, lagged so each training row pairs inputs with the outcome they truly preceded.
Think of it this way
Think of exam papers with the mark scheme stapled on. The questions are the features, the mark scheme is the labels, and a student learns only as well as the scheme is right. A misprinted scheme teaches confident error to every student who uses it.
The limit: the scheme answers yesterday's paper. Labels describe the outcomes of past cases under past rules, and when the world's marking habits change, yesterday's correct answers become today's bias.
Worked example
the misprinted scheme
Suppose a telecom samples two hundred labeled churn cases for audit. The figures are illustrative.
Fourteen per cent are wrong: customers marked churned who merely changed tariff, because the labeling rule conflated the two events.
Retraining on corrected labels lifts live precision by nine points, the largest improvement of the year, from no new model at all.
The audit cost a week of analyst time and outperformed a quarter of algorithm tuning, because the scheme had been misprinted, not the student.
Illustrative case study
A fictional lender let its collections system auto-generate labels from case closures, unaware that one team's closure code covered both repaid and written-off debts.
The model learned that written-off borrowers looked oddly creditworthy, priced a tranche of loans accordingly, and the loss review took months to find the label rule at the bottom of it.
The invented lesson: ask how the label is made before asking how the model performs. The answer key is part of the product.
Common misconceptions and mistakes
- Labels are objective facts. They are the outputs of a human or mechanical process with rules, incentives and error rates, and they deserve the same skepticism as any other data.
- More labels fix bad labels. Volume entrenches systematic error; only better labeling rules and audits fix it.
- Labeling is a one-off cost. Outcomes keep arriving, rules keep changing, and the labeling pipeline needs an owner for the life of the model.
Business use
Write the label definition down and get the business to sign it: it is the model's constitution.
Budget for label quality audits as a standing line, not a launch luxury.
Track label timeliness: stale labels teach the model yesterday's rules for tomorrow's decisions.
Review the label definition whenever the underlying business rule changes, because the model will keep learning the old rule long after the policy has moved on.
People also ask
What is the difference between a label and a feature?
Features are the inputs the model sees; the label is the answer it learns to predict from them.
What are classification and regression labels?
Classification labels are categories; regression labels are numbers, and the choice decides the model type and metric.
How accurate are typical labels?
It varies widely; audits regularly find five to fifteen per cent error in human-labeled sets, which is why sampling audits matter.
Check your understanding
Why does label quality cap model quality?
Answer: Because training teaches the model to reproduce the labels it is shown, so every systematic label error becomes a systematic model error, no matter how good the algorithm is.
Related terms
Learn next
Be Productive Using AI builds the habit of asking where the answers behind an AI system actually came from: https://aiforbusinessandlife.com/academy/be-productive-using-ai. [C1]
More in AI and machine learning
- L 1 regularization - L1 regularization is a penalty added to training that charges the model for the size of its weights, with the distinctive effect of pushing unneeded weights...
- L 2 regularization - L2 regularization is a penalty added to training that charges the model for the square of its weights, shrinking all weights toward zero to keep any single...
- L1-based feature selection - L1-based feature selection uses a model penalized by the sum of absolute coefficient values to identify a smaller set of input features.
- Label noise - Label noise is error or unreliability in the target labels used to train or evaluate a supervised learning system.
- Label Propagation - Label Propagation is a semi-supervised method that spreads labels from the few cases that have them through a similarity network, so unlabeled cases inherit...
- Lagged features - Lagged features are inputs built from the past values of a series: yesterday's sales, last week's balance, the same customer a month ago, giving a model...
Sources
[S1] Russell and Norvig, AIMA glossary; [S2] Google, Machine Learning Glossary; [S3] NIST AI Resource Center glossary; [S12] Aalto University, Dictionary of Machine Learning. Full URLs are in the source ledger. Verified 11 October 2026.
Previous: L1-norm principal component analysis Next: Label bias problem