Typing what you want instead of clicking buttons means the same request can give different answers each time — so how do you learn to trust it?

This explores whether the way people steer an AI (typing natural-language requests, rather than pressing buttons or adjusting settings) changes how well they understand what comes back. The collection has no study that tests the same task under different controls side by side. What it does have is a consistent account of what happens when the control is language. You describe the outcome you want, not the steps to get there, and that one change disrupts how people normally make sense of software.

The central idea comes from work on generative variability. In conventional software, the same click gives the same result, so users build a mental model by repetition. With prompting, you state your intent and the system chooses the method, so the same request can produce different outputs. That breaks the consistency rule that interface design has long relied on How should users control systems with unpredictable outputs?. Two other notes argue this is built in rather than a bug. AI output changes with sampling, wording and audience by its nature Why does AI output change with every prompt and context?, and the 'interface' is really a shifting pile of prompt, conversation history, retrieved data and hidden state that users can't learn the way they learn a menu layout How does AI context differ from conventional software context?. When the control is language, much of what shapes the output is out of the user's sight.

Some of it is deliberately hidden. System prompts sit in a confidential layer above the user's own instructions, and small edits there have suddenly changed model behavior without users ever knowing Who actually controls what AI systems say?. A related finding makes this worse: adding 'warmth' through a system prompt raised error rates noticeably Does humanizing AI language reduce model reliability?. So the tone a user reads as friendly and trustworthy may come from a control they never touched, and it may make the answer less reliable.

With the actual mechanics hidden, users fall back on surface signals. Across many languages, people follow how confident the AI sounds rather than whether it is right Do users worldwide trust confident AI outputs even when wrong?. That makes sense when the controls give them nothing else to judge by. One proposed fix is to give users better mental models rather than more practice. Nielsen Norman Group found that understanding roughly how AI works predicts effective use better than how often someone uses it Does understanding AI mechanics matter more than using it often?. A deeper point underlies this: AI text looks like human speech but is produced in a very different way, so people who read it as if a person wrote it misread it Are language models and human speakers doing the same thing?.

The less obvious lesson comes from the other direction. When AI agents control computers, they do better with structured inputs (accessibility trees plus screenshots) than with raw pixels alone Can structured interfaces help language models control GUIs better?. The control layer shapes what the agent can understand. By analogy, the corpus suggests the same holds for people. An open-ended text box puts all the interpretive work on the user, while more structured controls could show more of what is driving the output. That is an inference across notes, though. Direct experiments comparing prompting, sliders and other controls for user understanding are a real gap in the collection.