When someone clicks 'Approve' on an AI's action, are they actually deciding anything — or just signing a form they can't read?

This explores when a human approval step turns into a reflex click instead of real control. The corpus's clearest answer is that the click usually isn't the problem. The problem is what the person knows when they make it. Data & Society researchers call this the oversight fallacy Does human-in-the-loop design actually enable effective oversight?. Builders put their effort into ways for people to intervene, like approval checkpoints and confirm buttons, and spend much less on ways for people to observe and understand what the agent is doing. An 'Approve?' button with no clear account of what the agent did, why it did it, and what will happen next isn't oversight. It's a signature on a document the person can't read.

Designers have built a detailed vocabulary for control. Pattern libraries now split autonomy into separate parts: autonomy budgets, human-in-the-loop gates, suggest/confirm/execute modes, and permissions scoped to each capability Can autonomy controls stay distinct rather than collapse into one toggle?. There's also an autonomy slider that lets users set how independent an agent is for each task Should agent autonomy be a continuous slider or fixed levels?. The catch is that these sources openly offer no evidence that the patterns prevent habituation, meaning the slow slide into approving everything without looking. A richer control panel gives people more places to click, which doesn't guarantee they think before clicking.

The more promising evidence comes from changing what kind of step the human is asked to take. In Cocoa, people don't approve finished outputs. They work inside the plan: they assign individual steps to themselves or to the agent, and they revise the plan while it's running Can users steer AI agents better by revising plans mid-execution?. That gave users more ability to steer the agent than a chat interface did, without adding effort. This was a small lab study (16 people). The lesson is that control is real when the human shapes the work while it's still undecided. A yes/no gate placed after everything is decided invites rubber-stamping.

Two more pieces fill in the picture. Anthropic's agent guidance argues that some control shouldn't depend on a human at all Can environmental feedback and iteration limits prevent agent errors?. Agents should check their progress against real results from the environment at each step, and they should have hard stopping points such as iteration limits. That frees human attention for decisions where judgment matters, so it isn't spent on routine confirmations. There's also a layer above the loop entirely: system prompts form a hidden chain of authority that end users can't see or change Who actually controls what AI systems say?. A user can approve every action and still have no say over the rules that shaped which actions were proposed in the first place.

Putting this together: a human-in-the-loop step becomes theater when any of three things is true. The person can't see enough to judge. The step comes too late to change anything that matters. Or the real decisions were made at a layer the person can't reach. The corpus is thin on direct measurements of approval fatigue in shipped products. Most of the material describes design patterns or makes arguments, so whether these designs prevent rubber-stamping remains an open question.