Paper urges AI agents to help construct user preferences

The paper challenges the 'expert user' assumption in AI agents, arguing users often lack domain knowledge. It proposes agents should help construct preferences rather than just elicit them, leading to better alignment.