Description: Climate models used to produce global warming scenarios exhibit widely diverging responses of the thermohaline circulation (THC). To investigate the mechanisms responsible for this variability, a regional Atlantic Ocean model driven with forcing diagnosed from two coupled greenhouse gas simulations has been employed. One of the coupled models (MPI) shows an almost constant THC, the other (GFDL) shows a declining THC in the twenty-first century. The THC evolution in the regional model corresponds rather closely to that of the respective coupled simulation, that is, it remains constant when driven with the forcing from the MPI model, and declines when driven with the GFDL forcing. These findings indicate that a detailed representation of ocean processes in the region covered by the Atlantic model may not be critical for the simulation of the overall THC changes in a global warming scenario, and specifically that the coupled model’s rather coarse representation of water mass formation processes in the subpolar North Atlantic is unlikely to be the primary cause for the large differences in the THC evolution. Sensitivity experiments have confirmed that a main parameter governing the THC response to global warming is the density of the intermediate waters in the Greenland–Iceland–Norwegian Seas, which in turn influences the density of the North Atlantic Deep Water, whereas changes in the air–sea heat and freshwater fluxes over the subpolar North Atlantic are only of moderate importance, and mainly influence the interannual–decadal variability of THC. Finally, as a consequence of changing surface fluxes, the Labrador Sea convection ceases by about 2030 under both forcings (i.e., even in a situation where the overall THC is stable) indicating that the eventual breakdown of the convection is likely but need not coincide with substantial THC changes.