Our results suggest this is because of entangled perceptions: writers dislike seeming more extreme, but like seeming more confident, and to readers these are two sides of the same coin. AI developers may have little incentive to train more faithful AI assistants, despite risks from distortive AI.