How could the climate models possibly be accurate? Models reduce large areas of the Earth to a single point, the grid cell. Think about your own current grid cell, 100 km in horizontal resolution, 1 km vertical and 10 minute time step. What are the values of the physical quantities like temperature, pressure, wind speed, cloudiness, albedo, flowing radiation energies, CO2, humidity, precipitation and so on? One single value, please.
Where do you find the observations that give the initial values from the bottom of the oceans to the limits of the space. What are the observed right values 10 minutes later, 30 years later and 50 years earlier?
Climate models try to take smaller than grid things like hurricanes and clouds into account by parametrization. It is not physics, it is curve fitting. What could e.g. 76% cloudiness mean? One thunderstorm in a corner of the grid or thin cumulus clouds covering the whole area?
Models try to calculate solutions to complex problems like heat transfer using numerical methods instead of using error-less physics based equations.
Ensembles average the values of individual models and runs but systematic errors in parametrization, initial values, numerical mathematics remain. It is not useful to project that the global average temperature in 2100 will be between 15 to 23 C. You must get all the variables right at the same time. This includes among others the sociological model projecting CO2, ocean heat content, rains and humidity at the top of tropics, the famous hot spot.
It might be that models that are selected to the official ensemble are the ones that fit the curves of the past the best and project a politically correct future. Fitting the parametrization based on the surface station temperatures might not be the right way to do it.