Beyond uncertainty in modern active learning for trustworthy AI

Active learning (AL) is a central response to the annotation bottleneck in modern artificial intelligence: when labels are expensive, a learner should query for the most useful forms of supervision rather than indiscriminately acquiring labels. However, contemporary AL is no longer a unified field organized around a small set of stable query principles. It is fragmented across acquisition strategies, supervision granularities, operational regimes, and evaluation protocols, making reported gains difficult to compare and, in some cases, to trust. This study offers a critical review and synthesis of modern AL, with particular attention to deep learning and deployment-oriented applications across medical imaging, computer vision, natural language processing, systematic review automation, recommender systems, anomaly detection, and structured prediction. The review makes three contributions. First, it proposes a four-axis taxonomy organized around acquisition logic, supervision granularity, operational regime, and evaluation realism. Second, it compares major acquisition families, including uncertainty-based, disagreement-based, expected-improvement, representativeness-based, diversity-aware, cost-aware, and shift-aware approaches, highlighting their assumptions, strengths, computational trade-offs, and recurrent failure modes. Third, it distills design principles and an actionable research agenda for trustworthy AL, emphasizing annotation cost, redundancy control, robustness under distribution shift, fairness, human oversight, and workflow-grounded evaluation. The central argument is that the main challenge for AL has shifted from identifying informative samples to designing supervision-allocation pipelines whose gains remain reliable across realistic annotation workflows, heterogeneous human effort, and deployment constraints.