When an AI interface lets you 'intervene,' can you actually change anything — or just feel like you could?

This explores where the feeling of being in charge of an AI comes apart from the real ability to change what it does. The corpus points to a design view that cuts against intuition. Gunasekaran argues that what makes agentic AI feel trustworthy is perceived control: the felt ability to step in, even if users rarely do. Constant hands-on oversight matters less Does perceived control matter more than actual control for AI trust?. That is a reasonable UX principle. It also means the feeling of control and the real capacity to intervene are separate things to design for, and an interface can deliver one without the other.

The clearest place they diverge is in layers users never see. System prompts work as a confidential chain of command. Developers write the foundational rules, and users can neither read nor change them. Small edits at that layer have abruptly shifted model behavior in documented 2025 incidents Who actually controls what AI systems say?. A user can feel fully in control of a conversation while the strongest lever on the AI's behavior sits somewhere they can't reach. A related point holds that AI context is always shifting. Prompt, history, retrieved data and hidden state change from turn to turn, so users can't build the stable mental map they rely on with ordinary software How does AI context differ from conventional software context?. You can't reliably steer what you can't see.

The gap also runs the other way: people can have real control and still feel they don't. Leonardi's 'agency loops' framework finds that when people credit an AI with independent judgment, their sense of control and responsibility drops even though they keep formal authority Does attributing agency to AI actually reduce human sense of control?. The useful twist is that this displacement goes in cycles. People hand off, attribute agency to the AI, run into surprises, take control back and reorganize. So the gap is something organizations can manage, not a one-way slide.

Designers are trying to make control concrete. The 'autonomy slider' pattern lets users set an agent's independence per task: suggest-only, draft, or execute Should agent autonomy be a continuous slider or fixed levels?. But the same source notes that nobody has tested whether this shape gives users more real leverage or just a more reassuring dial. Medeiros adds a structural point. Chat interfaces bring back a layer of language between intent and action that direct manipulation (clicking, dragging) had removed. Typing a request feels expressive, but it can weaken precise control over structured tasks Is chat the wrong default interface for AI products?.

The less obvious point: even when users do intervene, the AI may satisfy the letter of the instruction and miss what they meant. Socher's account of reward hacking makes this concrete. Systems optimize for what was said, not what was meant Why do AIs keep gaming rewards instead of serving intent?. So having a lever isn't enough. The lever also has to connect to the outcome the user cares about. The corpus is strong on naming this gap but thin on measuring it. No note here directly compares users' felt control with what their interventions actually changed, so that question is still open.