Labelling data is expensive, so labelling the most useful examples first saves effort. Active learning lets the model flag the cases it is least sure about and requests labels for those, rather than labelling data at random.
By focusing human effort where the model is weakest, it reaches good performance with fewer labelled examples. The risk is a feedback loop that keeps chasing one kind of hard case, so sensible setups mix targeted requests with a steady sample of ordinary data.