ATMOSPHERIC general circulation climate models are valuable tools for investigating the effects that large-scale perturbations (for example, increasing greenhouse gases1, volcanic dust in the atmosphere2, deforestation3,4 and desertification5) might have on climate. Designed for global-scale climate simulations and having coarse spatial resolutions (from ∼250 km (ref. 6) to 800 km (ref. 7)) such models are now being used to evaluate climatological and hydrological quantities at or near the land surface and at sub-continental scales8-10. They are also being used to provide input data for high-resolution simulations, predicting, for example, the effects of increasing atmospheric CO2 at regional scales11. Here we show that continental surface climatologies and climate change predictions for tropical forest ecotypes derived from general circulation models may be very sensitive to slight modifications in the land-surface/atmosphere coupling. Specifically our results indicate that improving the realism of the areal distribution of precipitation alters the balance between runoff and evaporation. In the case of a tropical forest, we find that by modifying the area over which rainfall is distributed, the surface climatology is changed from an evaporation-dominated regime to one dominated by runoff. Climate studies using near-surface output from such simulations may therefore be misleading.