Data labeling is the process of attaching the correct answers, or labels, to raw data so that a supervised machine-learning model has something to learn from: drawing boxes around objects in images, transcribing speech, tagging text with sentiment or categories, or rating model outputs. Because the quality and quantity of labels often determine how well a model performs, labeling is one of the most important and most expensive parts of applied machine learning.
A range of tooling and services exists to manage this work. Open-source platforms such as Label Studio provide interfaces for annotating many data types, support multiple projects and annotators, and can integrate with models to generate pre-labels or drive active-learning loops in which the model picks the most useful examples to label next. Commercial labeling providers like Scale AI offer large managed workforces and tooling. Weak-supervision approaches such as Snorkel reduce manual labeling by combining noisy programmatic labeling rules instead of hand-labeling every example.
Labeling quality, consistency between annotators, and clear guidelines are recurring practical concerns, since inconsistent labels cap how good a model can ever become.
For a business reader, data labeling is the often-underestimated operational cost behind any custom AI model, and improving it is frequently the cheapest way to improve results.