Although most current coupled general circulation models (GCMs) exhibit some sort of ENSO signal, there are still many areas for improvement. For example, the models generally simulate El Niño events with frequencies which are too high, structures which extend too far to the west, and a large diversity of amplitudes. Moreover, simulating the correct ENSO properties with the right balance of mechanisms and feedbacks is still a challenge. Several recent studies using ocean-atmosphere GCMs suggest that the atmospheric component, and in particular the deep convection scheme, plays a dominant role in the modeled ENSO. To help elucidate these findings, the two main atmosphere feedbacks relevant to ENSO, the Bjerknes positive feedback (μ) and the heat flux feedback (α), are here analysed in 12 coupled GCMs from the CMIP3 database. We find that the models generally underestimate both feedbacks, leading to an error compensation. The strength of α is inversely related to the ENSO amplitude in the models and the latent heat and shortwave flux components of this feedback dominate. While the latent heat feedback is primarily responsible for this inverse relationship, errors in the shortwave flux feedback are the main cause of the model diversity in the overall α. In the tropical Pacific, the shortwave flux feedback is intrinsically linked to the large-scale vertical motion, with SST anomalies in the East Pacific coupled to changes in the amount of deep convection/subsidence and cloud cover. We thus propose that an improved atmosphere-ocean heat flux feedback in the models can only be achieved by correcting the errors in the convection/cloud physics responsible for the biases in the shortwave flux feedback.