There's a lot of power in using the Wizard of OZ prototyping technique to shape the intents and personality of a conversational UI interaction, especially as it is "natural", voice-driven interaction. Does the "fake" nature of the interaction have to ever be disclosed to the research subject? It seems to be something more easily done hands on and in persona to avoid user frustration after the disclosure.
Of should the user research subject ever be "let in" on what's behind the mirror? Do they need to know?