If an AI lets you clear up confusion your own way, do people pick one favorite method — or mix and match?

This explores what people actually do when an AI system lets them choose how to clear up a misunderstanding, instead of forcing one fixed way to clarify. The most direct evidence points to a surprising answer: people don't settle on one preferred strategy. In the PleaSQLarify study, the system turned ambiguous database requests into visible choices that users could explore. Participants used several different ways of fixing the misunderstanding, often within a single session, and clarification took between two and eight turns. The design lesson is that interfaces should support many ways of resolving ambiguity, not pick the 'right' one Can systems surface ambiguity instead of guessing user intent?.

When people do get asked a question, the type of question matters. Questions that target a concrete gap ('What type of monitor?') consistently beat open requests to restate a goal ('What are you trying to do?'). The reason is that people engage when they can see how their answer will improve the result Which clarifying questions actually improve user satisfaction?. That fits PleaSQLarify's approach of showing the ambiguity directly: in both cases the user is handed something concrete to react to, not asked to explain themselves again.

The less obvious finding is that people often don't want to be asked at all. In a Wizard-of-Oz study of computer-use agents (participants believed they were using an AI that a hidden researcher was actually running), users wanted the agent to quietly resolve low-stakes ambiguity, like which menu to click. They wanted it to stop and check before high-stakes actions like payments. What set their tolerance was the stakes, not how ambiguous the request was When should agents ask before acting on ambiguous commands?. Conversation analysis offers a more formal version of the same idea. Human speakers insert short side-exchanges into a conversation to clarify intent or narrow scope before going on. AI agents that chain tool calls silently skip this step and drift away from what the user meant When should AI agents ask users instead of just searching?.

There is also a quieter problem: much of human clarification doesn't look like a question. Research based on Clark's model of communication levels (attention, hearing the signal, meaning, action) finds that most clarifications are statements, and they are often carried by gaze, gesture or tone, not by wording. Systems that detect clarification by looking for question syntax miss most of it Why do clarification requests look different at each communication level?. So 'what users prefer' may partly be a measurement artifact: we count the repairs our systems can see.

Finally, be careful about trusting preference as the target. People report satisfaction even when they are still confused, and continued engagement tracks real understanding better than a satisfaction score Does user satisfaction actually measure cognitive understanding?. A similar gap appears with reasoning formats: users liked planning-style explanations, but simpler chain-of-thought traces helped them catch errors better Do people prefer the reasoning formats that help them verify?. The repair strategy people reach for and the one that actually fixes the misunderstanding may not be the same. The corpus has little direct evidence comparing preferred and effective repair strategies, so that question remains open.