1. Home
  2. Glossary
  3. Data labeling
AI glossary · Models & training

Data labeling

Data labeling: Data labeling is the work of attaching correct answers to raw data, such as tagging emails as spam, drawing boxes around pallets in photos, or rating chatbot replies, so a machine learning model can learn from those examples.

Most useful machine learning is supervised, meaning the model learns from examples that already carry the right answer. Someone has to supply those answers. That is labeling: reading ten thousand support tickets and tagging each with a category, marking which transactions were fraud, transcribing audio, or ranking two AI responses to say which is better. The labeled set becomes the model's ground truth.

Quality matters more than volume. If two labelers would tag the same ticket differently, the model learns the disagreement. Good labeling projects start with written guidelines and edge-case examples, measure how often labelers agree, and route ambiguous items to an expert. Labels also go stale: a category that made sense last year may not match how customers write today.

Labeling is a real industry, with vendors and crowdsourcing platforms supplying workers, and it is where much of the human effort behind AI actually happens. The human feedback used to tune chat assistants is labeling too. And every time you click thumbs up or thumbs down on an AI reply at work, you are producing a label the vendor may use to improve its model, so check what your company's terms allow.

Example at work

A customer support manager wants an AI tool to route incoming tickets. Before anything is built, her team labels 1,500 past tickets with one of eight categories using a one-page guide. The first pass shows two categories overlap so often that agents cannot agree, so she merges them. The routing model trained on the cleaned labels is far more accurate than one trained on the original set.

Why it matters

If your team is asked to help train or evaluate an AI system, labeling is the job. Knowing that the quality of those labels sets the ceiling for the model helps you push for clear guidelines instead of rushing through it.

Related terms