Analyses of simulations of variations in global and large-regional land surface air temperature (LSAT) for 1872–1998 using the HadAM3 atmospheric general circulation model are reported. The analyses are designed to test the accuracy of bias corrections to sea-surface temperature (SST) used in the Hadley Centre's global sea ice and SST (GISST3.1) data set, the more recent Hadley Centre sea ice and SST (HadISST) data set, and in the underlying Met Office historical SST (MOHSST and HadSST1) data sets. The tests are important because SST corrections considerably affect estimates of the magnitude of global warming since the late 19th century. Two ensembles of simulations were created using GISST3.1 as the lower boundary condition. The first ensemble, of six integrations, was forced using GISST with bias-corrections applied from 1871 until 1941, and was continued with no bias corrections to 1998. A second ensemble of four integrations, for 1871 to 1941, was forced with uncorrected GISST data. Simulations with uncorrected GISST show a substantial and often highly significant cold bias in simulated global and large-regional annual mean LSAT changes before 1942 relative to a 1946–65 reference period. By contrast, corrected SST data led to simulations of LSAT changes that are generally insignificantly different from those of observed LSAT in most regions before 1942. Tests on extratropical hemispheric scales generally validate the seasonal variation of the bias corrections, though less clearly before 1890 in some seasons. Issues about the quality of the LSAT data are raised by the results in a couple of regions. Over Australia, the model may have reconstructed LSAT changes using bias-corrected GISST with greater accuracy than the observations before about 1910. © Crown Copyright 2005. Reproduced with the permission of Her Majesty's Stationery Office. Published by John Wiley & Sons, Ltd.