Many Earth system models (ESMs) approximate surface emissivity as a constant. This broadband approximation reduces computational burden, yet biases longwave (LW) atmospheric fluxes and heating by neglecting the spectral structure of surface emissivity and atmospheric absorption. These biases are largest over surfaces with strongly varying emissivity and minimal atmospheric opacity (e.g., due to water vapor and clouds). Our study focuses on liquid water, ice, and snow surfaces. We use LW spectral emissivity ε(λ) calculated via the Fresnel equations and validated against a dataset of spectral surface emissivity. We flux-weight and bin ε(λ) into 16 spectral bands accepted by an offline single-column atmospheric radiative transfer model (RRTMG_LW) commonly used in ESMs (including E3SM and CESM). We quantify flux and heating biases introduced by broadband emissivity assumptions in comparison with the 16-band spectrally resolved case for three different surface types, three standard atmospheric profiles, and for the key drivers surface temperature, cloud water path, and atmospheric water vapor. In addition, we devise and test novel greybody and semi-spectral methods of representing ε(λ) with the goal of reducing biases while preserving computational efficiency. We find that typical broadband assumptions artificially cool Earth’s surface, thereby stabilizing the lower troposphere. LW upwelling flux is overestimated by 4.5 W/m2 (~1.4%) at the bottom of a mid-latitude winter atmosphere over an ice surface, and by 3.3 W/m2 (~1.4%) at the top of atmosphere. Lastly, we find that a semi-spectral approach (five bands instead of 16) reduces biases by up to 99% relative to the broadband approximation.