The more AI agents run on their own, the worse humans get at stepping in — and approval buttons alone won't fix that.

This explores what happens to human supervisors as AI agents run longer without them. The corpus says oversight capability drops, and it drops in two ways at once. When an agent acts on its own for longer, the person in charge sees less of what it did and why. Over time, relying on the system also wears away the skills supervision depends on: knowing what's going on, judging well, and knowing the domain Does granting agents more autonomy undermine human oversight?. The second effect is the less obvious one. The overseer isn't just badly placed to watch; with long enough disuse, they become a weaker overseer.

The common fix is to add approval checkpoints, and it misses the problem. Data & Society researchers found that builders invest heavily in the ability to intervene (approve, reject, pause) but put far less into the tools that show what the agent actually did. A checkpoint gives you a button. It doesn't give you the understanding to know when to press it Does human-in-the-loop design actually enable effective oversight?. A 48-person study shows how this plays out. People overseeing computer-use agents mostly checked whether the agent followed its own plan, not whether its actions were safe Do overseers focus on agent execution or action safety?. An overseer who has drifted out of touch tends to judge whether the agent stuck to its process, not whether the outcome was safe.

The obvious response is to review everything, and that fails too. In a research-automation system, step-by-step human review produced a 50% accept rate, compared with 25% for full autonomy. Routing only uncertain, high-stakes decisions to a person reached 87.5%. Constant interruptions make people rubber-stamp, which is its own kind of skill loss Does targeted human oversight beat both full autonomy and exhaustive review?. That points toward oversight designed around specific moments. Deloitte proposes placing humans at deliberately chosen handoff points, backed by audit logs and cryptographic receipts How should organizations supervise autonomous AI agents?. Other research treats 'how much should the agent decide alone' as a design question with three layers: the model's policy for when to escalate, the software around it that enforces that policy, and the interface that shows users what they've delegated How should AI agents balance autonomy with user oversight?.

The skill loss matters because the human is often what makes an agent usable at all. A 90%-reliable agent can be fine when a person catches its errors and unacceptable when it runs unattended Should reliability requirements change with agent autonomy?. If longer autonomy quietly erodes that person's ability, the safety margin shrinks without anyone noticing. This is one reason some researchers argue risk rises steadily with autonomy and that fully autonomous agents shouldn't be built Does AI risk increase with the autonomy we give it?. Traditional governance tools like monitoring and incentives also struggle to make up the gap when agents decide quickly and opaquely Can traditional governance tools actually control AI agents?.

At the scale of whole societies, the same pattern appears as 'gradual disempowerment.' Institutions stay aligned with human interests partly because they depend on people who care how things turn out. As AI takes over that work, the capacity to notice and correct drift weakens, possibly past the point of recovery Does incremental AI replacement erode human influence over society?. The fact you might not have expected: skilled oversight doesn't stay in reserve while it goes unused. It decays. The corpus's main answer is to keep humans actively involved in the work instead of watching from the sidelines Should AI systems stay collaborative rather than fully autonomous?. The corpus has little direct measurement of how fast overseer skills decline, so that remains an open question.