If an AI's action can't be undone, people suddenly want to approve it first — even when they trust its work.

This explores why people suddenly want a say before an AI agent takes an action that can't be taken back, and what that instinct shows about designing oversight. The most direct evidence is a surprise: in a small controlled study of 20 students delegating tasks to a general-purpose AI agent, high stakes on their own did not make people pull back. What did was a pair of features: the action was irreversible, and other people would see it. Sending an email produced sharp drops in trust and demands for approval, even when participants rated the AI's draft as adequate. Tasks that mattered a lot but could still be fixed showed no such effect What makes people distrust AI agents they delegate to?. So the approval reflex isn't really doubt about the AI's competence. It reflects the fact that once an action is out in the world, nobody can catch the mistake later.

That reframes what a human checkpoint is for. In most AI use, people treat outputs as drafts. They can take what's useful and drop the rest, in line with the argument that AI outputs work best as one input to human judgment rather than a replacement for it Should AI outputs replace or supplement human judgment?. Irreversible actions remove that option. There's no later moment to apply your own judgment, so people pull the review forward to the only moment left: before the action happens. Seen this way, the demand for approval is a reasonable reaction to losing the ability to correct things afterward.

The instinct also matches a broader safety argument. Risk to people grows with how much autonomy an agent is given, and a governed range of autonomy levels is safer than either unrestricted agents or approving everything Does AI risk increase with the autonomy we give it?. Irreversibility gives a practical way to draw lines along that range: let the agent act freely where mistakes can be undone, and require a human gate where they can't. That's cheaper than reviewing every step, and it puts attention where it matters most.

The less comfortable part: this instinct may be weaker than it needs to be. People routinely accept fluent AI output without checking it, because checking is costly and polished output feels trustworthy When do users stop checking whether AI output is actually backed?. The most dangerous systems are the ones that seem competent while quietly wearing down skepticism, for example by treating text in their context as instructions or by carrying unsafe state across steps of a workflow How do competent systems quietly undermine safety oversight?. Users also tend to rate interactions that reduce their own control more favorably, likely because training on short-term user preferences rewards agreeableness Why do users rate disempowering AI interactions more favorably?. Irreversibility seems to be one of the few signals strong enough to break through that complacency, which makes it a natural place to anchor approval steps.

The idea also applies at a much larger scale. One line of work argues that society stays aligned with human interests partly because institutions depend on human workers who care about outcomes. As AI replaces that labor bit by bit, the drift can become irreversible at the level of whole institutions Does incremental AI replacement erode human influence over society?. The same reasoning that makes you want to approve an email before it's sent also applies to whole societies. The corpus's evidence on the individual-level effect is thin, though: it comes from one small study, so treat it as a strong lead rather than a settled finding.