Letting the model choose what to label next. Instead of annotating randomly, you label the examples the model is least sure about, which reaches good performance with far fewer labels. It changes annotation from a fixed upfront cost into an ongoing loop.
Technical view
Query strategies include uncertainty sampling, query by committee using disagreement among an ensemble, and expected model change. Effective when labelling dominates cost, but it produces a biased, non-independent training set that is unrepresentative of the deployment distribution, which complicates evaluation. Batch selection also needs a diversity term, or the model requests many near-identical examples.
Related
Part of AI Terminology Index, section: Classical Machine Learning.