AI developers should compromise on model capability to guarantee human control