Abstract During the last ice age, temperatures in Greenland frequently increased and decreased by up to 10° ${}^{circ}$C. Detailed climate‐model studies suggest that these abrupt temperature changes were caused by collapses and recoveries of the Atlantic Meridional Overturning Circulation (AMOC). The causes of these AMOC transitions are debated. Here we describe the development of a neural‐network‐based surrogate model of the AMOC. It is trained on 32,000 years of climate model integrations to build a set of stochastic differential equations that emulate the climate model’s AMOC behavior. In particular it reproduces the spectra and the asymmetry in the times it takes for the AMOC to recover and collapse, which makes it more realistic than previously published sets of coupled differential equations to study past AMOC transitions. Monte Carlo simulations suggest more deterministic and predictable collapses but stochastically forced recoveries.