Forecasts of climate change are inevitably uncertain. It is therefore essential to quantify the risk of significant departures from the predicted response to a given emission scenario. Previous analyses of this risk have been based either on expert opinion, perturbation analysis of simplified climate models or the comparison of predictions from general circulation models. Recent observed changes that appear to be attributable to human influence provide a powerful constraint on the uncertainties in multi-decadal forecasts. Here we assess the range of warming rates over the coming 50 years that are consistent with the observed near-surface temperature record as well as with the overall patterns of response predicted by several general circulation models. We expect global mean temperatures in the decade 2036-46 to be 1-2.5 K warmer than in pre-industrial times under a 'business as usual' emission scenario. This range is relatively robust to errors in the models' climate sensitivity, rate of oceanic heat uptake or global response to sulphate aerosols as long as these errors are persistent over time. Substantial changes in the current balance of greenhouse warming and sulphate aerosol cooling would, however, increase the uncertainty. Unlike 50-year warming rates, the final equilibrium warming after the atmospheric composition stabilizes remains very uncertain, despite the evidence provided by the emerging signal.