<p>Inferring fine-grained information from aggregated data is a fundamental challenge across science and policy. This review reframes this ill-posed problem through the lens of modern Artificial Intelligence (AI). We trace the methodological evolution from foundational statistical approaches and Bayesian hierarchical models, which address identifiability and uncertainty, to transformative AI paradigms. Specifically, we examine how deep learning and generative models leverage weak aggregate supervision to learn complex patterns and synthesize realistic microdata. A computational benchmark compares these paradigms, demonstrating AI’s capability to recover latent structures where classical methods often fail. We discuss the shift from explicit statistical modeling to flexible, data-driven inference, addressing key implications for validation and ethical governance. The review concludes by outlining a future centered on hybrid models that combine statistical rigor with the scalability of AI.</p>