data labeling

basicwidespread

Data labeling is the work of adding the correct answer to images, texts, or audio recordings so that artificial intelligence can learn from it.

In plain terms

Imagine you are teaching a small child about animals from a picture book: you point to the picture and say, 'this is a dog', 'this is a cat'. Data labeling is the same, only for machines and at the scale of millions. Raw data (a photo, a sentence, an X-ray) is mute on its own; the label says what it shows or what the correct answer is.

A label can take many forms: a single word ('spam' or 'not spam'), a box around a pedestrian in an image, a transcript of an audio recording, or an evaluation of which of two chatbot responses is better. The model filters out the rule from thousands of examples, which it then applies to new, unseen data.

This is the least spectacular yet one of the most important work phases of artificial intelligence. The quality of the model depends largely on the quality of the labels: it can tolerate a few random errors, but if the training material is systematically incorrect or biased, the machine learns the error and the bias.

How it works

They collect raw data (images, texts, audio or video recordings) and select the portion worth labeling.
Written guidelines are prepared on what and how to label: what categories exist and what to do with borderline cases.
Labelers (humans, increasingly with the help of a pre-labeling model) tag data with labels in dedicated software.
Quality control follows: the same data is labeled by multiple people, discrepancies are compared, and disputed cases are decided by an experienced expert.
The model is trained and tested on a prepared, labeled dataset, and based on its errors, new data are labeled or old ones are corrected.

An everyday example

When a website asks you to select all the squares showing a pedestrian crossing or a traffic light, it is primarily checking that you are human, but your answers also label images as a side effect. The same happens when you mark a message as spam in your email: the filter learns from your signal too.

Who is affected

everyoneThe quality of the translators, search engines, spam filters, and chatbots used daily depends significantly on how carefully humans labeled and evaluated the data used for their training.

entrepreneur, executiveIf you want a model tailored to your company's own data, labeling is often one of the most time- and money-consuming parts of the project, although for large pre-trained models, far fewer labeled examples may suffice today.

developerThe labeling guide and quality control often improve the model more than refining the algorithm. → Code

patient, doctorDiagnostic AI systems learn from images labeled by doctors, so the accuracy of the labels directly affects the reliability of the findings. → Health

A common misconception

MythArtificial intelligence learns on its own, without human work.

In realityBehind most systems lies extensive human labeling work performed by masses of people worldwide, often for low wages. While large language models acquire their basic knowledge from unlabeled text, human-written sample responses and evaluations of the model's answers are still used for their refinement and verification.

Limits and risks

Labelers make mistakes and do not always agree with each other, and faulty labels reduce model accuracy.
Labels may embed the labeler's bias or cultural perspective, which the model adopts and can even amplify.
Labeling requiring expertise (such as medical or legal) is expensive and slow because only a trained person can perform it reliably.
Part of the work is done under poor conditions, and labeling disturbing content is mentally taxing, raising ethical questions for the industry.

Similar, but not the same

data cleaning – Data cleaning fixes erroneous, missing, or duplicate data, while labeling adds new information, the correct answer, to the data.
unsupervised learning – There, the model searches for patterns in data without labels, meaning labeling is omitted.
content moderation – The moderator decides whether content can remain on a platform, while the labeler prepares training material for a model, although both often work on the same content and with similar tools.

Related concepts

Broader

data preparation, machine learning, training data

Narrower

image annotation, text annotation, community tagging, automatic labeling

On the edge

supervised learning, adathalmaz, synthetic data, bias, reinforcement learning from human feedback

A brief history

Labeled training data has been a basic requirement since the beginning of supervised machine learning. The era of mass labeling was opened by the ImageNet image database introduced in 2009, whose millions of images were labeled by people recruited online via the Amazon Mechanical Turk platform. The breakthrough built on this, the AlexNet neural network, showed in 2012 how much machine image recognition improves when deep neural networks learn from large amounts of carefully labeled data and with high computing capacity.

Where things stand

By the end of 2026, data labeling is an independent, globally established industry in which models themselves perform an increasing share of routine tasks via pre-labeling. The focus of human work is shifting toward expert verification and evaluation of language model responses.

Status: 2026 November

In the news

Appears in 2 Agent Age stories; the curve shows weekly occurrence in the news magazine.

More than $137 million has been spent on AI ads in US midterm campaignsOct 19, 2026 · Marketing
California AI labeling law and platform rules: what creators need to discloseOct 19, 2026 · Marketing

Also known as

data labeling · labeling · data annotation · annotation · data labeling · data labelling · data annotation · data tagging

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.