Future projections of evapotranspiration (ET) are of critical importance for agricultural and freshwater management and for predicting land–atmosphere feedbacks on the climate system. However, ET from phase 5 of the Coupled Model Intercomparison Project (CMIP5) simulations exhibits substantial biases, bolstering little confidence in future ET projections. Despite poor predictive skill and large bias of ET from the global climate models, the information content necessary to calculate ET offline is available in the models’ archived outputs: temperature T, water vapor pressure e, atmospheric pressure P, and surface net radiation R. A relatively simple three-source energy balance model [Penman–Monteith (PM)], along with the mean annual cycle of remotely sensed vegetation properties, can then be used to reconstruct ET with a substantially reduced bias relative to in situ turbulent heat flux measurements. This methodology is used here to reconstruct ET projections from 2006 through 2100 over North America using output from selected CMIP5 models and to attribute projected ET trends to specific atmospheric controls. CMIP5 ET exhibits substantial bias in annual ET relative to in situ flux measurements across North America (38%–73%; 2006–15), but ET reconstructed from the CMIP5 meteorology with the PM method greatly reduces this bias (−8% to +14%). Present-day North American ET is more sensitive to changes in atmospheric demand for ET (temperature and water vapor pressure) than energy limitation (net radiation), and to a lesser extent vegetation properties (leaf area index). Accordingly, ET is projected to increase 0.26–0.87 mm yr−1 yr−1 over North America through 2100 driven primarily by trends in temperature.