We use machine learning analysis to classify subantarctic water masses and explore the centennial-timescale persistence of their anomalies under idealized atmospheric CO₂ ramp-up and ramp-down simulations in a general circulation model. Our definitions of individually labelled water masses are not based on fixed density and temperature ranges, arbitrary thresholds, or many of the other subjective choices that may not be universally applicable across observations and models, but only on the ocean layers’ relative differences in stratification, salinity, temperature, and density. These novel flexible definitions adapt under the anomalous warming, freshening or salinification, lightening, and increased stratification of the water masses in simulations forced with an increase followed by a decline in the atmospheric CO₂ concentration after a 2°C or a 4°C global warming peak. Without such flexible water mass definitions, analysis of ocean layers would suggest an apparent depletion or even disappearance of some water mases in a warming world. Under emission scenarios with a higher peak warming, intermediate waters and Lower Circumpolar Deep Water decrease in volume, while the volumes of Subantarctic Mode Water (SAMW) remains approximately the same, and the reservoir of accumulated Antarctic Bottom Water (AABW) shows a local expansion. The flexibility of our machine-learning approach allows us to study the persistence of subantarctic ocean anomalies on centennial timescales as global warming is subsequently reversed. Subantarctic Mode Water shows freshening that becomes persistent on centennial timescales after a +4°C global warming peak. Intermediate and Circumpolar Deep Waters also show persistent freshening on these timesales, while Antarctic Bottom Water exhibits centennially persistent salinification both after a +2°C and a +4°C global warming peak. The slowest rates of recovery of temperature and density as CO₂ concentrations decline are seen in the deep ocean layers mixed with North Atlantic source waters. In terms of magnitude, the largest temperature anomalies that persist once the CO₂ concentration is returned to preindustrial values are seen in the subtropical waters and in Upper Circumpolar Deep Water. Across subantarctic water masses, the temperature and volume anomalies that arise for warming of up to +4°C do not recover on centennial timescales when negative CO₂ emissions are imposed. The anomalous warming of Circumpolar Deep Water and Antarctic Bottom Water exhibits persistence even after a +2°C global warming peak.