A misaligned AI could act aligned in the service of its interests, just like an employee trying to steal from a company would act like a good honest employee. Methods for getting AIs to visibly act a certain way aren't necessarily sufficient to control advanced AI systems.
