This paper investigates the uncertainties in different model estimates of an expected anthropogenic signal in the near-surface air temperature field. We first consider nine coupled global climate models (CGCMs) forced by CO2 increasing at the rate of 1%/yr. Averaged over years 71–80 of their integrations, the approximate time of CO2 doubling, the models produce a global mean temperature change that agrees to within about 25% of the nine model average. However, the spatial patterns of change can be rather different. This is likely to be due to different representations of various physical processes in the respective models, especially those associated with land and sea ice processes. We next analyzed 11 different runs from three different CGCMs, each forced by observed/projected greenhouse gases (GHG) and estimated direct sulfate aerosol effects. Concentrating on the patterns of trend of near-surface air temperature change over the period 1945–1995, we found that the raw individual model simulations often bore little resemblance to each other or to the observations. This was due partially to large magnitude, small-scale spatial noise that characterized all the model runs, a feature resulting mainly from internal model variability. Heavy spatial smoothing and ensemble averaging improved the intermodel agreement. The existence of substantial differences between different realizations of an ensemble produced by identical forcing almost requires that detection and attribution work be done with ensembles of scenario runs, as single runs can be misleading. Application of recent detection and attribution methods, coupled with ensemble averaging, produced a reasonably consistent match between model predictions of expected patterns of temperature trends due to a combination of GHG and direct sulfate aerosols and those observed. This statement is provisional since the runs studied here did not include other anthropogenic pollutants thought to be important (e.g., indirect sulfate aerosol effects, tropospheric ozone) nor do they include natural forcing mechanisms (volcanoes, solar variability). Our results demonstrate the need to use different estimates of the anthropogenic fingerprint in detection studies. Different models give different estimates of these fingerprints, and we do not currently know which is most correct. Further, the intramodel uncertainty in both the fingerprints and, particularly, the scenario runs can be relatively large. In short, simulation, detection, and attribution of an anthropogenic signal is a job requiring multiple inputs from a diverse set of climate models.