Learn AI Create the Future — with Pranay
Main site

AI glossary

Data labelling

In one sentence Data labelling is the human work of writing the correct answer onto each training example — slow, costly, and quality-critical.

By Updated

Data labelling is the process of humans attaching the correct answer — the label — to each raw example so a model can learn from it.

It is checking ten thousand exam papers so that, later, a machine can learn to check them itself. Someone must sit and mark each one, the marking must be consistent from paper to paper and person to person, and every careless tick becomes a lesson the model faithfully absorbs. Supervised-learning is powered by this quiet, enormous industry: annotators drawing boxes around cars, tagging toxic comments, transcribing audio, marking tumour boundaries.

The craft is harder than it sounds, because real examples resist clean answers. Is a sarcastic compliment a positive review? Half your annotators will say yes. The professional toolkit deals with exactly this: written labelling guidelines with worked edge cases, multiple annotators per example with disagreements measured (inter-annotator agreement) and adjudicated, and gold-standard checks salted into the queue to catch fatigue.

Cost drives the workarounds. Active learning has the model pick which examples are most worth labelling next, spending human time only where the model is confused. Weak supervision generates rough labels from rules. LLMs now pre-label at scale with humans reviewing — cheaper, though it imports the LLM's own biases. And synthetic-data arrives labelled by construction.

The uncomfortable rule of thumb: label quality caps model quality, and an afternoon auditing your labels is often worth more than a week of model tuning.

Where to go next