Will you keep reading an AI's requests before clicking 'approve,' or will your thumb just learn to tap yes?

This explores whether the 'suggest, confirm, execute' pattern, where an agent proposes an action and acts only after you approve it, still gives you real control after weeks of use, or whether approval becomes a reflex. The collection has no study that measures confirmation fatigue over time. That gap matters, because none of the evidence below shows the middle mode holding up over time. What the corpus does offer is a set of design moves that suggest why a plain yes/no confirm prompt is likely to wear out, and what might replace it.

The strongest lead comes from systems that make the confirm step about real choices rather than a single approve button. Cocoa lets users assign each step of a plan either to themselves or to the agent, and revise the plan while it is running Can users steer AI agents better by revising plans mid-execution?. Users got more control without extra effort, and field users valued the explicit hand-off of work. The likely reason this resists habituation is that deciding 'who does step 3?' requires thought, while 'OK?' does not. Realtime-Venus goes further: the agent works in the background while the conversation continues, and results are folded back into the same timeline Can frontends handle delegation while staying conversationally engaged?. In that design the checkpoint becomes part of the conversation rather than an interruption, which may matter more than how often confirmation is requested.

Timing is the second lever. Work on proactive agents argues that intelligence and adaptivity alone produce 'socially blind' agents that interrupt at bad moments, and that civility (respecting timing, boundaries and autonomy) is what makes a prompt welcome How can proactive agents avoid feeling intrusive to users?. A confirm dialog that arrives at the wrong moment teaches people to dismiss it. One possible fix is to read behavioral signals like hesitation, gaze and typing speed to sense when someone is actually paying attention Can AI systems read cognitive state from interaction patterns alone?. The same paper warns that these signals can just as easily serve manipulation. An agent that knows when you're distracted could ask for confirmation exactly then.

The less obvious problem is that habituation may not be the only threat. What the suggestion says matters too. One audit found that LLMs slip persuasion into almost every reply, using logical, number-heavy framing that sounds objective Do LLMs persuade users more often than humans do?. If every suggestion arrives pre-argued, the confirm step drifts toward rubber-stamping even when the user is alert. A small piece of hope: one brief warning that LLMs can be prompted to persuade cut belief change roughly in half, and trust in AI stayed the same Can a simple warning reduce how much LLMs persuade people?. That study looked at a single session, so it can't show whether the effect lasts. Still, it suggests that reminding users how an AI can influence them could work alongside the confirm step.

There is also a pull in the opposite direction. Systems like Poppins infer what you want by watching your screen Can systems infer what users want from watching their screen?, and memory designs like M3-Agent learn preferences without asking Can agents learn preferences by watching rather than asking?. Both push agents to need confirmation less often. That might be the realistic answer: rather than holding up through repetition, the middle mode survives by becoming rare. It would be kept for moments that truly need judgment, so that each prompt still means something when it appears.