New AI study shows when LLMs should ask questions, not infer preferences

A new research paper titled "When to Ask a Question: Understanding Communication Strategies in Generative AI Tools" (arXiv:2605.11240) by Charlotte Park, Kate Donahue, and Manish Raghavan addresses a critical design challenge in large language models: when should the AI ask for more information versus inferring what the user wants from minimal input? The authors note that generative AI’s flexibility in allowing users to provide as little or as much context often forces the model to infer details based on distributional knowledge, which can disadvantage users with atypical preferences and perpetuate majority biases.

The researchers build a formal mathematical model that captures the trade-off between user burden (annoyance from too many questions) and preference representation (accurately reflecting the user's true needs). They show that correlated preferences across users can be leveraged to determine the optimal amount of information to solicit before generating content. Their empirical evaluation confirms that strategic elicitation can mitigate the systematic biases of pure inference, enabling AI tools to serve diverse users more equitably while maintaining efficiency. The work has direct implications for designing conversational AI, recommendation systems, and any generative tool that must balance personalization with usability.