The realism of the Hadley Centre's coupled climate model (HadCM2) is evaluated in terms of its simulation of the winter North Atlantic Oscillation (NAO), a major natural mode of the Northern Hemisphere atmosphere that is currently the subject of considerable scientific interest. During 1400 y of a control integration with present-day radiative forcing levels, HadCM2 exhibits a realistic NAO associated with spatial patterns of sea level pressure, synoptic activity, temperature and precipitation anomalies that are very similar to those observed. Spatially, the main model deficiency is that the simulated NAO has a teleconnection with the North Pacific that is stronger than observed. In a temporal sense the simulation is compatible with the observations if the recent observed trend (from low values in the 1960s to high values in the early 1990s) in the winter NAO index (the pressure difference between Gibraltar and Iceland) is ignored. This recent trend is, however, outside the range of variability simulated by the control integration of HadCM2, implying that either the model is deficient or that external forcing is responsible for the variation. It is shown, by analysing two ensembles, each of four HadCM2 integrations that were forced with historic and possible future changes in greenhouse gas and sulphate aerosol concentrations, that a small part of the recent observed variation may be a result of anthropogenic forcing. If so, then the HadCM2 experiments indicate that the anthropogenic effect should reverse early next century, weakening the winter pressure gradient between Gibraltar and Iceland. Even combining this anthropogenic forcing and internal variability cannot explain all of the recent observed variations, indicating either some model deficiency or that some other external forcing is partly responsible.