We demonstrate that the temperature signal in the planktonic foraminifera assemblage data from the North Atlantic typically does not originate from near-surface waters and argue that this has the potential to bias sea surface temperature reconstructions using transfer functions calibrated against near-surface temperatures if the thermal structure of the upper few hundred metres of ocean changes over time. CMIP5 climate models indicate that ocean thermal structure in the North Atlantic changed between the Last Glacial Maximum (LGM) and the pre-industrial (PI), with some regions, mainly in the tropics, of the LGM ocean lacking good thermal analogues in the PI. Transfer functions calibrated against different depths reconstruct a marked subsurface cooling in parts of the tropical North Atlantic during the last glacial, in contrast to previous studies that reconstruct only a modest cooling. These possible biases in temperature reconstructions may affect estimates of climate sensitivity based on the difference between LGM and pre-industrial climate. Quantifying these biases has the potential to alter our understanding of LGM climate and improve estimates of climate sensitivity.