Reinforcement corrosion has been recognized as an influential factor in the seismic fragility, both demand and capacity models, of aging reinforced concrete (RC) bridges. For capacity models, accurate and applied prediction tools accounting for aging effects are yet to be well established. Current practices usually perform numerical analyses to obtain time-variant capacity models, which are time-consuming particularly when multi-source structural and environmental uncertainties are considered and sometimes even suffer computational non-convergence. To address these issues, this study leverages a rigorously optimized artificial neural network architecture to develop data-driven models for rapid estimates of probabilistic curvature capacity of corroded circular RC bridge columns with flexural failure modes. An extensive database of multi-level curvature limit states (i.e. slight, moderate, severe, and complete) is created through experimentally validated moment–curvature analyses. A new threshold for the moderate limit state is defined based on the strain of core concrete, rather than cover concrete, to account for the potential full erosion of the cover with drastic corrosion. The data-driven probabilistic capacity models are applied to aid the lifetime seismic fragility assessment of a typical highway bridge, where the spectral acceleration at 1.0 s (Sa-10), peak ground velocity (PGV), and Housner intensity (HI) are found consistently, for the first time, as optimal intensity measures for probabilistic demand modeling of RC columns with different extents of aging effects. For the ease of application, the database and code for the data-driven probabilistic capacity models are accessible at https://bit.ly/3uAa8EY .