摘要:The proxy for temperature ( δ signal) in ice cores is stored in the snow/ice during precipitation events and hence reflects the temperature at which precipitation is formed (here approximated by the inversion temperature T i ) weighted with the accumulation. Results from a 14 year integration (1980–93) with a regional atmospheric model (RACMO, Δ X = 55 km) show that the annual mean accumulation-weighted inversion temperature ( T i,w ) and the annual mean T i are not covariant in time at four out of the five deep-drilling sites considered, mainly due to year-to-year variations in the seasonality of precipitation. As a consequence, the surface temperature ( T s,core ) derived from RACMO output, using a method analogous to the retrieval of the surface temperature from ice-core δ signals, deviates from the directly modelled surface temperature T s on interannual time-scales. Results from a 5 year sensitivity integration, forced with a 2 K temperature increase, show an 18% overestimation of the increase in T s,core relative to the increase in T s due to a change in the relationship between the inversion strength and the surface temperature in a different climate regime. Similar errors may occur in deriving the temperature difference between Last Glacial Maximum and present-day climate from δ signals in ice cores.