Abstract
Artificial intelligence (AI)-based tissue diagnostics is entering a new phase driven by pathology foundation models: large-scale encoders and multimodal systems pretrained with self-supervised and vision-language objectives. Yet the evidence base has expanded faster than the methods used to evaluate and interpret claimed capabilities. The central clinical question is no longer whether these models can achieve strong retrospective performance, but what new capabilities they provide, under what conditions those capabilities are demonstrated, and whether they can translate into reliable diagnostics in clinical settings. This review makes three contributions. First, we define histopathology-centric foundation models and distill the technical factors that shape their behavior: pretraining data regime, learning objective, and downstream adaptation to clinical endpoints. Second, we introduce a practical capability framing that distinguishes general-purpose capabilities (coverage across tissues, scales, stains, scanners, and institutions) from functional breadth (task primitives, adaptability, and analysis level), while accounting for modality scope spanning vision, language, and genomics. Third, we synthesize reported results as an evidence map rather than a leaderboard, clarifying where capability is supported by reported evidence, where reproducibility is constrained by access or reporting limits, and where further validation is needed. We then analyze current benchmarking practice and identify common confounders, including heterogeneous adaptation protocols and pretraining-evaluation overlap, and propose deployment-aware recommendations built around broad-and-deep benchmarks, standardized adaptation "budgets," overlap auditing, and domain-agnostic evaluation. Finally, we review emerging clinical-utility evidence and argue that foundation models are most compelling when tied to workflow-defined endpoints, calibrated operating points, and measurable operational benefit. We conclude with actionable recommendations for converting capability demonstrations into clinically reliable data-driven diagnostic systems.