An Oversight Requirement a Random Number Generator Could Satisfy
A one-line test for rules that mandate human oversight, and the two requirement elements that no capability and no noise source can supply.
There is a one-line test for any rule that mandates human oversight of an automated system: could a random number generator satisfy the requirement as written? Read the requirement, replace the human with a source of noise that approves at some rate, interrupts on some schedule, and occasionally escalates, and ask whether the system remains compliant. If it does, the requirement was never protecting anything. It was mandating presence, and presence is the one property of oversight that costs nothing to counterfeit.
The diagnosis behind this test is not new, and the people who own it should be read first. Ben Green surveyed dozens of policies requiring human oversight of government algorithms and found the humans cannot perform the functions the policies assign them, so the mandates function as legitimation: the flawed system ships, now wearing a human signature (Green 2022). Ben Wagner called the resulting arrangements quasi-automation, systems that keep a human in the loop as a rubber stamp precisely so that the loop escapes the legal category of a fully automated decision (Wagner 2019). The empirical flank keeps filling in: overseers exhibit automation bias, defer to the system they are supposedly checking, and in some settings a human in the loop makes outcomes worse. None of this needs restating. What needs stating is why the standard repairs still fail the one-line test, and the two elements a passing requirement must contain.
The standard repairs strengthen the human. Give the overseer epistemic access, causal power, self-control, and fitting intentions, and the arrangement counts as effective oversight. The fullest statement of this capacity framework is Sterz et al. 2024. Meaningful-human-control accounts add a tracing condition: some human in the chain must properly understand the system and acknowledge being a legitimate target of blame (Santoni de Sio and van den Hoven 2018). These conditions bind where presence does not, but they are still requirements on the overseer’s capacities and self-understanding, not on their exposure. A maximally capable overseer with nothing at stake passes every capacity condition and remains an excellent rubber stamp, and acknowledging oneself as a blame target is compatible with a career in which no miss has ever cost the acknowledger anything. Capacity can be present, awareness can be sincere, and the requirement can still be satisfied by an arrangement in which oversight failure lands on no one who signs.
Here is the part I have not seen stated as doctrine, though the evidence for it has been accumulating in an adjacent field for decades. A substantial fraction of what oversight measurably contributes may not be judgment at all. It may be perturbation. The studies that find human feedback improves an automated process measure against the same baseline, no feedback, and that baseline cannot distinguish the value of the overseer’s intelligence from the value of being interrupted at all. Optimization has treated this as standard technique forever: random restarts rescue search from local minima, and the random number generator gets no credit for judgment. The same point has lived in programming folklore since before the field had a name: explaining a bug to a rubber duck fixes it, and the duck contributes cadence, not insight. I have found no study that includes a randomized-interrupt arm. Until it does, every measured benefit of a human in the loop is an unresolved mixture of two things with opposite policy implications: a component that can be bought from a noise source at the price of randomness, and a component that exists only because a particular party holds the criterion and eats the miss.
That split is one term short, and the missing term decides how the test should be written. Between noise and ownership sits an interrupt policy trained on a particular overseer’s history of interruptions. Such a policy is not random, because its timing carries information: the policy has learned where one person’s threshold sits, and fires there. Nor does it own anything, because a schedule extracted from someone’s record is still enforcing a threshold that person set. Timing is learnable. The criterion is not. That is the sharpest form of the distinction, because it survives the case where the machine’s behavior is indistinguishable from the overseer’s.
So the one-line test has a harder version, and the harder version is the one worth applying. A random number generator is the weak falsifier: a requirement can clear it and still be empty, because clearing a noise source is a low bar. Ask instead whether a policy trained on the overseer’s own past decisions could satisfy the requirement as written. Nearly every requirement that survives the noise test fails that one, and the reason is visible once stated. Requirements describe when the human acts. Almost none describe what the human answers for, and the timing is exactly the part a trained policy reproduces.
Institutions that take oversight seriously already split the mixture. Risk-limiting election audits draw ballots at random. Tax authorities audit at random. Athens allocated most public offices by lot. In every case, the randomness is the mechanism and never the authority: the lottery decides which ballots get checked precisely so that no official’s discretion can be captured, while the decision that elections shall be audited, to what confidence level, and what happens when the audit fails, is owned by a named institution that answers for it. Noise inside an owned criterion is mature oversight design. Noise as the criterion is nothing at all.
So the requirement worth writing has two elements, and neither is presence. First, criterion ownership: a named party holds the success criterion, including the right to revise it, and that right is located, not distributed into a diagram where every box can point to another. Second, outcome-bearing: the owning party can be worse off when oversight misses, in a way that arrives without anyone’s cooperation. The audit question that operationalizes both is short: who was worse off at the last miss? An arrangement that cannot answer with a name has automated its oversight already, whatever the org chart says. And the one-line test returns as the falsifier: a random number generator can be present, can interrupt, can escalate at a calibrated rate, and can even improve outcomes. It cannot own a criterion, because a source whose misses land on no one can administer a standard but not own one. Any requirement a random number generator could satisfy has confused the part of oversight that was always automatable with the part that is not an input at all.
The practical consequence is a reallocation, not a retreat. If part of oversight’s benefit is perturbation, buy it as perturbation: randomized checks, scheduled interrupts, sortition-shaped review rosters, all cheap, all capture-resistant, all already best practice somewhere. Spend the scarce thing, a human who owns the criterion and bears its misses, exactly where noise fails, on deciding what the system is for, noticing that the criterion itself has rotted, and answering for the miss no enumeration foresaw. Presence can be automated. Capacity can be simulated. Perturbation can be randomized. What cannot be conjured, by any capability at any price, is a party with something to lose, and oversight requirements should be written for exactly that element, because it is the only one the machines cannot supply.
The decomposition behind this essay, including the placebo-arm experimental design and the property-rights argument for why criterion ownership is not an input, is developed in a working paper.