What are the uncertainties involved in climate simulation? Firstly, we must be aware that it is impossible to establish the accuracy (or reliability1, dened as the extent to which accurate results are obtained in a given domain-see Chapter 3) of a climate model for future prediction simply because we cannot establish this accuracy based on repeated trials. For simulations of the past, the situation is different: There is one historical realisation of Earth’s climate, changing over time, with which climate-simulation outcomes can be compared. However, from determining the accuracy of a climate model for the past, we cannot derive its accuracy for the future. As an example, the paradigmatic Intergovernmental Panel on Climate Change (IPCC, 2001) diagram of simulations of the evolution of climate over the past 150 years is presented in Figure  6.1; results are shown for different climate-model inputs (external ‘radiative forcings’ [see page 99, rst note] on the climate system, either of natural or anthropogenic origin).* In Figure 6.1a, the grey band denotes an ensemble of historical simulations with one climate model in which no anthropogenic emissions are included (clearly counterfactual).† The result is a signicant mismatch between the model and the observations for the most recent decades. In Figure 6.1b, the model includes no natural external forcings from volcanoes and the sun (also clearly counterfactual). Now, the result is a signicant mismatch around the middle of the century. Figure 6.1c shows the result in which both natural and anthropogenic forcings are included. The model comes close to the observations, but how reliable1 is the model that produces this result? We have no way of determining this; we certainly cannot conclude from Figure 6.1c that the model is reliable1

for attributing the causes of climate change. The parameterisations in the model may be tuned to give a good t in Figure 6.1c, and maybe some processes that are important in reality are not included in the model. The most we can do quantitatively is to apply sensitivity and uncertainty analysis to estimate-but not determine-the reliability1 of the model.