AI Data Lab Wiki

AI Data Lab Wiki › Glossary › Data labeling

What Is Data Labeling?

By AI Data Lab Wiki editors · checked November 15, 2026

Data labeling is the process of attaching tags to raw data, such as photos, text or audio, so a machine learning model can learn what each item shows. AWS describes it as identifying raw data and adding one or more informative labels, and IBM uses "data annotation" as another name for the same step. 1 2

What Is Data Labeling

Data labeling is the point where people tell a model what its training examples mean. A label might say whether a photo shows a bird or a car, which words a speaker said, or whether an x-ray shows a tumor. 1

IBM places labeling in the preprocessing stage of building a model, before training starts. It lists computer vision and natural language processing as the main uses. 2

How Data Labeling Works

Data labeling works by asking humans to make a judgment about unlabeled data, and AWS says the judgment can be as rough as a yes or no or as fine as picking out specific pixels. 1

AWS lists these practices that make the work more accurate:

  • Task screens designed to lower the effort needed to switch between items.
  • Labeler consensus, which sends each item to several people and merges their answers into one label.
  • Label auditing, which checks labels and updates them.
  • Active learning, which uses a model to pick the data that most needs a human label.

A set labeled this way, used as the standard for training and testing, is often called ground truth. AWS adds that a model can label some data automatically and pass the items it is less sure of to humans, a form of human-in-the-loop work.

Labeled Data vs Unlabeled Data

Labeled data differs from unlabeled data in how a model uses it: IBM says labeled data is used in supervised learning and unlabeled data in unsupervised learning. 2

IBM also says labeled data is harder and more expensive to acquire and store, while unlabeled data is easy to collect but more limited in use. The topic is part of the wider subject of AI training data.

Data Labeling vs Data Annotation

Data labeling and data annotation are the same activity in some sources and different in scope in others. IBM writes "data labeling, or data annotation" as one phrase. 2

Toloka draws a line: labeling assigns a predefined tag from a fixed set to each data point, while data annotation is the broader category that also covers spatial markup, entity tagging and preference ranking. CleverX describes the same split and gives an example of marking an email as spam (labeling) against drawing a box around each car in a street scene (annotating). 3 4

Data Labeling vs Annotation in Practice

Data labeling versus annotation in practice is a signal of how complex the task is, according to Toloka, which says the distinction has mattered more since 2023. Work on language models, such as writing instruction and response pairs or ranking model outputs, does not look like traditional labeling. 3

CleverX says no standards body owns the distinction, and that most teams use the words interchangeably. Both explanations come from companies that sell data work, so this wiki treats the split as common usage and not as a formal definition.

Who Does Data Labeling

Data labeling is done by people employed by data companies, by contractors on worker platforms, or by internal teams. Amazon's SageMaker Ground Truth documentation says customers could use workers from a vendor company or a private workforce, and notes the service is no longer open to new customers. 5

For language models, the InstructGPT paper describes human labelers who wrote demonstrations of desired behavior and ranked model outputs. See data annotator for the role and AI data labs for the companies that supply this work. 6

Questions About Data labeling

These are the questions about Data labeling that people search most.

What does data labeling mean?

Data labeling means adding tags to raw data, such as images or text, so a machine learning model can learn from it. IBM and AWS both describe it as part of preparing data before a model is trained.

Is data labeling the same as data annotation?

Data labeling is the same as data annotation in IBM's wording, and a subset of it in Toloka's. Toloka and CleverX say labeling attaches one tag from a fixed set, while annotation also covers boxes, entities and rankings.

What is labeled data used for?

Labeled data is used in supervised learning, where a model learns from examples that already carry the correct answer. IBM says unlabeled data goes to unsupervised learning, which looks for structure without those answers.

Who does data labeling?

Data labeling is done by human labelers working for vendors, on worker platforms or in private teams, sometimes with software help. AWS says labeling can also be partly automated by a model that labels some data.

More AI Data Terms

More AI data terms explained in this wiki:

  • RLHFTraining a model on human rankings of its answers.
  • AI red teamingStructured testing that tries to make an AI model fail.
  • LLM evaluationTesting a language model's outputs against set criteria.
  • Synthetic dataData generated by algorithms rather than collected from real events.

Sources for Data labeling

The sources for Data labeling are listed in the order the page cites them, read November 15, 2026.

  1. aws.amazon.com/what-is/data-labeling
  2. ibm.com/think/topics/data-labeling
  3. toloka.ai/blog/annotation-vs-labeling
  4. cleverx.com/blog/data-annotation-vs-data-labeling
  5. docs.aws.amazon.com/sagemaker/latest/dg/sms.html
  6. arxiv.org/abs/2203.02155