All the models are build assuming recent changes are driven by greenhouse gas concentrations. All of them understate the effect of soot deposition on the cryosphere. All of them are trying to do long-term predictions on a chaotic dynamical system (already impossible without complete initial data) for which we don't know all the inputs and linkages. GIGO.
It would be interesting to turn modern AI loose on the problem of climate modeling -- see what the version that best retrodicts the observed data comes up with for predictions. They'd still be something to take with a whole shaker full of salt, since it's still a chaotic dynamical system, but at least it wouldn't be driven by dubious assumptions.