In plain English
Data labelling is the process of tagging raw data, such as images, text, or audio, with the meaningful labels a model needs to learn during supervised training. It is often slow, manual, and surprisingly influential: the care and consistency of labelling sets a ceiling on how good the resulting model can be.
Why it matters
Models are only as good as the labels they learn from. Inconsistent or careless labelling quietly bakes errors and bias into everything the model does afterwards.
A worked example
To train an animal classifier, people first draw boxes around cats and dogs in thousands of photos and tag each one, giving the model the answers to learn from.
Common confusion
Labelling is not a one-off chore. Real systems need ongoing labelling to handle new cases and to keep quality high as the data changes.