Past climate variations may be known from reconstruction methods that use proxy data as predictors. Among them, borehole reconstructions is a well established technique to recover the long term past surface air temperature (SAT) evolution. It is based on the assumption that SAT changes are strongly coupled to ground surface temperature (GST) changes and transferred to the subsurface by thermal conduction. We evaluate the SAT-GST coupling during the last millennium (LM) using simulations from the Community Earth System Model LM Ensemble (CESM-LME). The validity of such premise is explored by analyzing the structure of the SAT-GST covariance during the LM and also by investigating the evolution of the long term SAT-GST relationship. The multiple and single-forcing simulations in the CESM-LME are used to analyze the SAT-GST relationship within different regions and spatial scales and derive the influence of the different forcing factors on producing feedbacks mechanisms that alter the energy balance at the surface. The results indicate that SAT-GST coupling is strong at global and above multi-decadal time scales in the CESM-LME however a relative small variation in the long term SAT-GST relationship is also represented. Although at global scale such variation does not impact significantly the SAT-GST coupling, at local to regional scales this relationship experiences considerable long term changes mostly after the end of the 19th century. Land use land cover (LULC) changes are the main driver for decoupling SAT and GST locally and regionally since they modify the land surface properties such as albedo, surface roughness and hydrology, and thus the energy fluxes at the surface. Snow cover feedbacks due to the influence of other external forcing are also important for corrupting the long term SAT-GST coupling. Our findings suggest that such local and regional SAT-GST decoupling processes may represent a source of bias for SAT reconstructions from borehole measurement since the thermal signature imprinted in the subsurface over the affected regions is not fully representative of the long term SAT variations.