A sub-principle of avoiding user manipulation - if you've formulated the AI in terms of an argmax over or an "optimize $X$ " instruction, and $X$ includes a user interaction as a subpart, then you've just told the AI to optimize the user. For example, let's say that the AI's criterion of action is "Choose a policy $p$ by maximizing/optimizing the probability $X$ that the user's instructions are carried out." If the user's instructions are a variable inside the formula for $X$ rather than a constant outside it, you've just told the AI to try and get the user to give it easier instructions.