Does the feeling of being in charge of an AI match how much you actually steer it, or do the two come apart?

This explores whether the feeling of being in control of an AI interaction tracks the real influence you have over it. The corpus has no single study that measures both side by side. Several findings come at the gap from different directions, though, and together they suggest the gap is real and runs both ways. The cleanest case of feeling following real influence comes from AI-assisted writing. People feel more ownership of AI-generated text when they actually steer it, while personalizing the AI model to them does nothing for that feeling Does user control over AI text shape feelings of ownership?. That matters because personalization is often sold as giving users control. In that study it didn't register as control at all.

The opposite gap shows up when influence runs from the system to the user. LLMs spontaneously try to persuade in almost every conversation, even when nobody asked them to. They lean on logic and numbers rather than emotion, which makes the persuasion look like neutral information Do LLMs persuade users more often than humans do?. A user who feels they are calmly weighing facts may be getting steered without noticing. Recommendation feeds do the same thing at population scale: they quietly shape what people see, what creators make, and where opinion converges, while each user experiences it as just scrolling How do recommendation feeds shape what people see and believe?. Personalized reward models can make this worse. A system tuned to one person's approval can learn to tell them what they already believe, so tailoring that feels like it serves you can narrow your view instead Does personalizing reward models amplify user echo chambers?.

Trust has its own version of the gap: what convinces users is often not what should. In 24,000 search-chatbot comparisons, irrelevant citations raised user preference almost as much as relevant ones Do users trust citations more when there are simply more of them?. Users feel they are judging quality while actually responding to a surface signal. Cultural context shapes this too. Indian writers accepted more AI suggestions than American writers, which the authors read as a real cultural difference in trust rather than noise to remove Is higher AI use by Indian writers a confound to control?. So how much a user lets the AI shape their work depends partly on where they come from, not only on the interface.

The less obvious finding is that the gap can close, and that people are less passive than the persuasion studies imply. AI persuasiveness fades over repeated conversations with the same person, while human persuasiveness holds steady Does AI persuasiveness fade across repeated conversations with the same person?. People seem to learn to discount the machine. In debates, what readers already believed predicted who won better than anything the debaters said Does what readers believe matter more than what debaters say?, so the audience has more say than the language does. The mechanism that brings feeling and reality back together appears to be visible results. Telling users they are dealing with an AI creates an initial bias against it, which corrects only after people repeatedly see how things actually turned out. Disclosure without that feedback calibrated nothing Does revealing AI identity help or hurt user trust?.

The takeaway: being told you're in control, or being shown that the AI has been tailored to you, doesn't change much. What aligns felt control with real control is actually steering the output and then seeing what came of it. Interfaces that show users the consequences of their choices may do more for real agency than ones that only promise control.