IPCC 2007

This is the money quote.
IPCC 2007
“These studies highlight some common biases in the simulation of clouds by current models (e.g., Zhang et al., 2005). This includes the over-prediction of optically thick clouds and the under-prediction of optically thin low and middle-top clouds. However, uncertainties remain in the observational determination of the relative amounts of the different cloud types (Chang and Li, 2005). For mid-latitudes, these biases have been interpreted as the consequence of the coarse resolution of climate GCMs and their resulting inability to simulate the right strength of ageostrophic circulations (Bauer and Del Genio, 2006) and the right amount of sub-grid scale variability (Gordon et al., 2005). Although the errors in the simulation of the different cloud types may eventually compensate and lead to a prediction of the mean CRF in agreement with observations (see Section 8.3), they cast doubts on the reliability of the model cloud feedbacks. For instance, given the nonlinear dependence of cloud albedo on cloud optical depth, the overestimate of the cloud optical thickness implies that a change in cloud optical depth, even of the right sign and magnitude, would produce a too small radiative signature. Similarly, the under-prediction of low- and mid-level clouds presumably affects the magnitude of the radiative response to climate warming in the widespread regions of subsidence. Modelling assumptions controlling the cloud water phase (liquid, ice or mixed) are known to be critical for the prediction of climate sensitivity.”
There is more. I truncate only to get a readable comment in. As I recall, possibly from Soden somewhere different models arbitrarily choose a range for absorbtion in thes clouds which can be double another model. I hope most people would agree a doubling or halving gives critically different results and quite a different but predictable emergent ECS.
This is where the input that later forms the ECS occurs.
Worse it implies that 2 wrong assumptions may be being made that somehow compensate each other in that they then produce the right ECS.