If we assume that humanity has gained access to effectively infinite computing power, and has put AIXItl or something similar into a copy of the universe, simulated at whatever level unifies quantum mechanics and gravitation into a coherent, leakproof framework, AIXItl would have an extremely small belief that it was inside a simulation. Only if the simplest unification of quantum mechanics and gravity turns out to be "we're in a simulation," would a hyperintelligent AI in a perfect simulation of our universe come to the belief that it's in a simulation.

So, the epistemically perfect AI would come to an incorrect decision. This does not imply a flaw in its method for forming beliefs; it merely implies the tautology that there is no way to find out what there is no way to find out.