Data annotation is the work of turning raw material — photographs, recordings, documents, sensor sweeps — into labelled examples a model can learn from. A model cannot learn what a pedestrian looks like unless somebody has drawn boxes around thousands of pedestrians first.
The tasks split into families. Image labelling covers bounding boxes, polygons and key points. Text work covers entity tagging, classification and correction. Audio work is transcription and speaker turns. Video adds time: an object has to keep the same identity across hundreds of frames. Sensor work covers LiDAR cuboids for driving datasets.
None of it is unskilled, whatever the job adverts imply. The difficulty is consistency. Two reviewers labelling the same batch must reach the same answer, every time, and be able to explain why. That is why guidelines run to forty pages and why QA exists.
Pay tracks judgement. Drawing boxes is the entry point; evaluating model output — ranking two answers, spotting a hallucination, deciding whether a response broke policy — pays several times more because it requires reasoning in writing. That is the step most people skip, and it is the step that changes the rate.