Best I think would be if the warning came implicitly as part of the game, and a little ways into it.
For example: The player sees one NPC Alex warn another NPC Joe that failing to drink the Potion of Feather Fall will mean he's at risk of falling off a ledge and dying. Joe accepts the advice and drinks it. Soon after, Joe accidentally falls off a ledge and dies. Alex attempts to rationalize this result away, and (as subtly as possible) shrugs off any attempts by the player to follow conversational paths that would encourage testing the potion.
Player hopefully then goes "Huh. I guess maybe I can't trust what NPCs say about potions" without feeling like the game has shoved the answer at them, or that the NPCs are unrealistically bad at figuring stuff out.
Exactly -- that's the kind of thing I had in mind: the player has to navigate through rationalizations and be able to throw out unreliable claims against bold attempts to protect it from being proven wrong.
So is this game idea something feasible and which meets your criteria?
This thread is for the discussion of Less Wrong topics that have not appeared in recent posts. If a discussion gets unwieldy, celebrate by turning it into a top-level post.