Artificial Intelligence in Longevity Research: From Molecular Clocks to Interventional Geroscience

Artificial intelligence (AI) has expanded the scale, resolution, and modalities through which biological aging can be estimated, but evidence that these estimates are valid intervention targets or surrogate endpoints remains limited. We conducted a narrative review of PubMed-indexed literature identified through documented PubMed and Consensus searches in October 2026, plus a separate search of primary US Food and Drug Administration (FDA) and European Medicines Agency (EMA) documents. No protocol was registered, Embase, Scopus, and Web of Science were not searched, and screening was by a single reviewer. We distinguish four biomarker properties: chronological-age estimation, prognostic association, responsiveness to intervention, and validated surrogacy. Deep-learning clocks across epigenomic, proteomic, retinal, electrocardiographic, and magnetic resonance imaging (MRI) modalities report low chronological-age error, and several show prognostic associations with mortality, but calibration was rarely reported in the records reviewed and head-to-head comparisons with mortality-trained clocks in the same cohort were uncommon. Randomized trials show that responsiveness is clock-specific: in the CALERIE trial, caloric restriction slowed DunedinPACE but did not significantly change PhenoAge or GrimAge, and effect sizes were small. Machine learning has accelerated senolytic and mTOR-pathway candidate nomination, and a generative-AI-designed TNIK inhibitor reached phase 2a in idiopathic pulmonary fibrosis, an age-related disease indication rather than an aging indication. Coverage of the twelve hallmarks of aging by AI studies is uneven: epigenetic alteration and cellular senescence are best represented, whereas autophagy, proteostasis, genomic instability, and telomere attrition are thinly represented. Constraints include reliability, tissue specificity, confounding by ancestry and social factors, the gap between prediction and causation, and the absence, in the primary documents consulted, of a pathway that accepts a biological-age measure as a surrogate endpoint. We propose an evidence ladder and a translational chain for appraising AI-derived aging biomarkers and set out priorities for 2026-2030.