Auto-labeling, also called pre-annotation, is when a machine learning model generates an initial set of labels on raw data automatically, which a human annotator then reviews and corrects rather than labeling everything entirely from scratch. It shifts the annotator's job from creating labels to verifying and fixing them.

‍

‍Why it matters

Labeling large datasets entirely by hand is slow and expensive, and a lot of that time goes toward relatively easy, repetitive decisions a model can often get right on its own. Auto-labeling exists to remove that repetitive work, letting human effort focus on the genuinely difficult or ambiguous cases the model gets wrong, rather than spreading equal attention across every single item regardless of how obvious it is.

‍

‍How it's used in practice

The typical workflow runs a model over the raw dataset first, generating draft labels for every item, then routes that output to human annotators who confirm correct labels quickly and fix the ones that are wrong. This is usually significantly faster than labeling from a blank slate, though it comes with a real risk, annotators can develop a bias toward accepting the model's suggestion even when it's subtly wrong, which is why some teams deliberately track how often pre-annotated labels get corrected versus accepted, as a check on whether reviewers are actually reviewing rather than rubber-stamping.