“Current AI agents assume users have well-formed preferences and ask clarifying questions when tasks are underspecified. Research argues this approach is flawed—users often lack domain knowledge to specify preferences without agent guidance. The solution is for agents to actively teach users rather than passively elicit existing preferences.”
Key Takeaways
- Existing agents incorrectly assume users have expert-level domain knowledge and preferences
- Users need agents to help construct informed preferences, not just clarify existing ones
- This requires agents to teach domain concepts before eliciting user preferences
Agents must help users learn domain knowledge to form informed preferences.
trending_upWhy It Matters
This research challenges a fundamental assumption in AI agent design that limits their usefulness for everyday users. By shifting from preference elicitation to preference construction, agents can serve broader audiences who lack domain expertise. This could significantly improve user experience and expand AI agent adoption across non-specialist populations.
FAQ
How do agents currently handle underspecified tasks?
They default to asking clarifying questions, assuming users already understand their preferences—an unrealistic assumption for domain novices.
What's the practical benefit of agents teaching users?
Users can make informed decisions by learning relevant domain knowledge first, enabling agents to help construct well-reasoned preferences rather than just guessing at existing ones.
