Leaf margin characters are strong predictors of mean annual temperature (MAT) in modern plant communities and widely used tools for reconstructing paleoclimates from fossil floras. However the frequency of nonentire-margined species may vary dramatically between different habitats of the same forest. In this paper we explore the potential for this habitat variation to introduce error into temperature reconstructions based on field data from a modern lowland forest in Amazonian Ecuador. The data show that the provenance of leaves can influence temperature estimates to an important degree and in a consistent direction. Woody plants growing along lakes and rivers underestimated MAT by 2.5¬8‚Äì5¬8C while those in closed-canopy forest provided very accurate predictions. The high proportion of liana species with toothed leaves in lakeside and riverside samples appears to be responsible for a large part of the bias. Samples from closed-canopy forest that included both lianas and trees however were more accurate than tree-only or liana-only samples. We conclude that paleotemperature reconstructions based on leaf margin characters will be misleading to the extent that fossilization provides a better record of certain habitats than others. The preponderance of lake and river deposits in the angiosperm fossil record suggests that underestimation of mean annual paleotemperature may be common.