Between 1985 and 2014 observations show an expansion of Southern Ocean sea-ice. This phenomena is not simulated in CMIP6 Atmosphere-Ocean General Circulation Models (AOGCMs). Here we quantify the impact of this discrepancy on radiative feedback and simulated global temperature trends.

Using this relationship between and , for each model we apply three values of to demonstrate the impact of Southern Ocean feedback biases on the estimated . First, the estimated from the historical ensemble mean, second, a modified where we replace the Southern Ocean (55–78°S) with values from the amip-piForcing ensemble mean, and third, a modified where we replace the Southern Ocean with values from the amip-piForcing ensemble mean. These values of and the associated are indicated for each model as a dot, vertical marker, and an arrowhead respectively in Figure 3d and are recorded in Table 1. Here the Southern Ocean substitution has been performed for and to capture both the direct impact of the sea ice biases through surface albedo changes as well as any other related local processes such as changing clouds (Cesana et al., 2025).

The zonal mean of and over the Southern Hemisphere higher latitudes is shown in Figures 3b and 3c respectively, where black vertical lines indicate the region over which the amip-piForcing values have been substituted when modifying the historical . In Figure 3b observed values of have also been included, using observed values of and and taking from the ensemble mean of all CMIP6 models analyzed. This was done given the effective radiative forcing used for the IPCC AR6 is only provided as a global mean timeseries. A clear negative feedback can be seen in the observed estimate, confirming that in this region the amip-piForcing experiment is able to capture feedback processes consistent with those seen in the real world.

We see that had the coupled historical experiments simulated the observed changes in sea ice, assuming all other feedbacks remain unchanged, this would impact the global temperature trends by approximately 0.04 0.03 K (multi-model-mean). For HadGEM3-GC3.1-LL, this accounts for just under one third of the bias in historical temperature trends compared to observations. We see that depending on the model, when substituting in just the Southern Ocean , 12%–29% of the total disparity in global is accounted for between the historical and amip-piForcing experiments. This increases to 18%–57% when other local changes are considered and the Southern Ocean is substituted (Table 1). In the multi-model-mean, substituting all local Southern Ocean processes changes global-mean by 0.25 , whereas substituting just changes global-mean by 0.18 (Table 1), that is, contributes 72% to the total difference over the Southern Ocean.