Can organizations lose scrutiny capacity while keeping oversight forms?

The sentence sits in the introduction's list of ways quiet failure takes hold: "organizations can retain nominal human oversight while shedding the actual capacity to scrutinize machine recommendations." The failure is not that a reviewer approves a bad output. It is that the review step stays in the process diagram while the organization loses what would make the review real: the expertise, the time, the access to reasons and the standing to disagree.

That is different from rubber-stamping, which the vault already holds at the level of a single decision. Does targeted human oversight beat both full autonomy and exhaustive review? finds that step-by-step oversight floods the human with low-value approvals and degrades into stamping. Rubber-stamping is a behavior of a reviewer inside a process. The paper's claim is about the process itself: a reviewer who once could have scrutinized the recommendation no longer can, and no one recorded the loss because the form still gets signed.

A third limit is coverage, and it is separate from both. How much agent behavior actually gets human review? holds that a reviewer with intact capacity and no fatigue still sees only the decisions that reach them. Capacity and coverage fail differently and would need different remedies, and the paper's sentence is about capacity. The coverage note's own caution applies: tool calls and decisions are different units, so the gap is not a literal ratio.

Why this is hard to see is built into the arrangement. Nominal oversight produces the same artifact as real oversight, a recorded human decision, so an audit of the artifact passes. Only a test of capacity, for example whether a reviewer can catch a planted error, tells them apart, and what such a test would have to be is the open question in How can we measure whether AI errors stay visible and recoverable?. An anchored, tamper-evident approval trail makes the artifact more durable without making it more diagnostic, which is filed as the black box paper anchors human approvals as tamper-evident evidence while the vault says a recorded approval looks the same whether oversight was real or nominal — a durable record may not be an informative one. This is the accountability-diffusion mechanism from How do competent systems quietly undermine safety oversight? in its organizational form.

The implication for design is uncomfortable, because "keep a human in the loop" is the standard answer to agent risk. On this paper's reading it answers the wrong question unless the loop keeps its capacity; see the tension filed at the vault answers agent risk by keeping a human in the loop while the hidden-failures paper says nominal oversight can persist after the capacity to scrutinize is gone — the human-in-the-loop prescription may be hollow. One measured design that does not lean on the reviewer keeping its capacity is Can memory poisoning compromise decision-making even with authorization layers?: the reviewing agent was bypassed in every trial and no unsafe action ran, because an authorization check outside any agent's judgment stood between the approval and the action. That is a machine reviewer in one poisoned-memory pipeline and says nothing about a human loop. The transferable point, on the vault's reading, is only that safety can sit where the reviewer's capacity is not what carries it.

What the excerpt does not give. No organization, sector or study is cited, and no rate at which capacity erodes. It is stated as something organizations "can" do.