There is a concern that any hypothetical hyperintelligent AI could emulate its human adversaries and try various strategies against these models to see which one works better. In theory, with a sufficiently good models it would make this AI a perfect conversationalist, capable of talking humans into doing whatever it pleases. This is precisely what you do in this game: try and talk a human into letting you out of the proverbial box by repeating the conversation until he agrees. Also featuring your usual AI alarmist stuff like goal (mis)alignment, utility functions, difference between instrumental and terminal values, metaethics and inappropriate cyberpunk references. Not featuring Paperclip Optimizer because there is a better game about that one.