We estimate the causal impact smart thermostats have on energy use using data from a field experiment in which treated households were randomized into free installation of a smart thermostat. We combine this experimental data with 18 months of high-frequency data on household energy consumption in the form of more than 16 million hourly electricity use records and almost 700 thousand daily observations of natural gas consumption. We model the effect of a smart thermostat on energy consumption using a difference-in-differences instrumental variables (DDIV) specification. In contrast to advertised savings based on engineering models, we find no evidence that smart thermostats have a statistically or economically significant effect on energy use. This result is robust to the inclusion of numerous controls and when the model is estimated on various subsamples (e.g., by hour). We explore potential mechanisms using almost four million observations of system events including user interactions with their smart thermostat. Results indicate that user behavior dampens energy savings and explains the discrepancy between estimates from engineering models and those from the field.