When you tell an AI what you want in VR, do you already know — or figure it out mid-conversation?

This explores how what a user wants becomes clearer and more settled over a back-and-forth with an AI, especially in spatial settings like VR where the system can ask clarifying questions and show 3D previews. The corpus's core answer is that intent isn't something users already have and hand over. It firms up during the exchange. How do users actually form intent when prompting AI systems? describes it as progressive constraint resolution: each answer or choice narrows the space of what the user might mean, but their stability fluctuates along the way, so a user can be sure one minute and unsure the next. AI systems stumble partly because they can't see that internal state.

This is why answering a clarifying question can be harder than it looks. Why can't users articulate what they want from AI? argues that people can't articulate what they want in the abstract, and that models make it worse by responding rather than probing. The proposed fix is to change the job. Instead of asking users to imagine and describe a result from scratch, the system offers model-generated options and lets them react. Reacting to a candidate is easier than specifying one, and each reaction pins down another requirement the user hadn't put into words.

Spatial interfaces make this concrete. In a 24-person VR geometry-editing study, Do spatially-anchored previews stabilize VR geometry editing tasks? found that pairing clarification questions with previews anchored in the 3D scene cut the number of conversation rounds and made task progress steadier. Best-case performance didn't improve. My reading is that a preview in place lets a user see a candidate and say 'no, taller' or 'closer to that wall', which works as constrained evaluation. The gain shows up as fewer stalls and false starts rather than a higher ceiling. That is an inference: the study measured task progression, not intent directly.

Other notes suggest why the timing and shape of the questions matter. When should AI agents ask users instead of just searching? borrows a idea from conversation analysis: a short side-question that clarifies intent or scopes the response, asked mid-stream, prevents a misunderstanding instead of repairing it afterward. Can dialogue systems track both speakers' beliefs across turns? models this as both parties tracking each other's beliefs turn by turn, moving from partial to shared understanding. Can conversation shape predict whether it will work? adds that the trajectory of a conversation predicts satisfaction almost as well as reading its text (68% vs 70%), which suggests that how intent matures leaves a visible signature. How does AI context differ from conventional software context? points to a further wrinkle in spatial work. The scene, the history and the user's goal all keep changing, so what the user wants and the thing it's about move together.

The corpus has only that one VR study on this, and no work that tracks how a user's confidence in their intent changes turn by turn in a spatial interface. What it gives you is a well-supported general picture (intent matures through interaction, and previews plus targeted questions make that smoother) and one spatial data point consistent with it.