Draft a philosophical argument addressing the 'Value Loading Problem' in AI alignment. Specifically, analyze the difficulty of defining 'human happiness' as a maximization function for an AI without causing perverse instantiation (e.g., the experience machine drug scenario). Discuss Coherent Extrapolated Volition (CEV) and propose a hypothetical framework for an AI to distinguish between stated preferences and true interests, referencing instrumental convergence.