Faithful Where It Can Be Checked: Auditing a Reflection Agent Against Its System Prompt in a Randomized Trial

Abstract

Conversational agents are increasingly used to guide reflection. A recent randomized trial compared a GPT-4o career reflection agent with the same program in a static journaling survey. Agent participants ended less committed to their career plans and more doubtful. We coded all 17,930 turns from its two studies, checked our coding against human coders and linked conversations to the trial's surveys. The rules the agent followed were the easy-to-check ones, like a reply length cap. Told not to flatter, it praised participants in half of its turns; told to challenge gently, it almost never did, and such a break leaves no visible trace. The behavior tied to the worse outcome was the demand to decide: the survey posed each decision once, while the agent asked again when participants hesitated, and those pressed most ended most doubtful. Our findings inform reflection agent design and the writing of checkable instructions.