Efforts to understand and quantify how a changing climate can impact agriculture often rely on bias-corrected and downscaled climate information, making it important to quantify potential biases of this approach. Here, we use a multi-model ensemble of statistically bias-corrected and downscaled climate models, as well as the corresponding parent models from the Coupled Model Intercomparison Project Phase 5 (CMIP5), to drive a statistical panel model of U.S. maize yields that incorporates season-wide measures of temperature and precipitation. We analyze uncertainty in annual yield hindcasts, finding that the CMIP5 models considerably overestimate historical yield variability while the bias-corrected and downscaled versions underestimate the largest weather-induced yield declines. Misrepresentation of tails of temperature distributions is the main factor leading to differences between CMIP5 and downscaled model products. We also find large differences in projected yields and other decision-relevant metrics throughout this century, with overconfident projects linked to underestimation of temperature extremes. Results highlight challenges facing stakeholders considering different climate information and the inherent trade-offs in resolution, historical accuracy, and projection confidence.