The accuracy of Arctic sea ice surface temperatures Ts derived from advanced very high resolution radiometer (AVHRR) thermal channels is evaluated in the cold seasons by comparing them with surface air temperatures Tair from drifting buoys and ice stations. We use three different estimates of satellite surface temperatures, a direct estimate from AVHRR channel 4 with only correction for the snow surface emissivity but not for the atmosphere, a single-channel regression of Ts with Tair, and Key and Haefliger's (1992) polar multichannel algorithm. We find no measurable bias in any of these estimates and few differences in their statistics. The similar performance of all three methods indicates that an atmospheric water vapor correction is not important for the dry winter atmosphere in the central Arctic, given the other sources of error that remain in both the satellite and the comparison data. The errors are not reduced by regression with both thermal channels and the satellite scan angle. A record of drifting station data shows winter air temperature to be 1.4 ¿C warmer than the snow surface temperature. ''Correcting'' air temperatures to skin temperature by subtracting this amount implies that satellite Ts estimates are biased warm with respect to skin temperature by about this amount. A case study with low-flying aircraft data suggests that ice crystal precipitation can cause satellite estimates of Ts to be several degrees warmer than radiometric measurements taken close to the surface presumably below the ice crystal precipitation layer. An analysis in which errors are assumed to exist in all measurements, not just the satellite measurements, gives a standard deviation in the satellite estimates of 0.9 ¿C, about half the standard deviation of 1.7 ¿C estimated by assigning all the variation between Ts and Tair to errors in Ts. ¿ American Geophysical Union 1995