The global and tropical means of clear‐sky outgoing longwave radiation (hereinafter OLRc) simulated by the new GFDL atmospheric general circulation model, AM2, tend to be systematically lower than ERBE observations by about 4 W m−2, even though the AM2 total‐sky radiation budget is tuned to be consistent with these observations. Here we quantify the source of errors in AM2‐simulated OLRc over the tropical oceans by comparing the synthetic outgoing IR spectra at the top of the atmosphere on the basis of AM2 simulations to observed IRIS spectra. After the sampling disparity between IRIS and AM2 is reduced, AM2 still shows considerable negative bias in the simulated monthly mean OLRc over the tropical oceans. Together with other evidence, this suggests that the influence of spatial sampling disparity, although present, does not account for the majority of the bias. Decomposition of OLRc shows that the negative bias comes mainly from the H2O bands and can be explained by a too humid layer around 6–9 km in the model. Meanwhile, a positive bias exists in channels sensitive to near‐surface humidity and temperature, which implies that the boundary layer in the model might be too dry. These facts suggest that the negative bias in the simulated OLRc can be attributed to model deficiencies, especially the large‐scale water vapor transport. We also find that AM2‐simulated OLRc has ∼1 W m−2 positive bias originating from the stratosphere; this positive bias should exist in simulated total‐sky OLR as well.