Almost every regime that governs automated decisions relies on a human at some point.
An overseer who can disregard or override the system. A reviewer of an automated decision. A caseworker exercising a statutory discretion. A professional making an individual assessment.
Each regime expects that involvement to be genuine. None says what evidence would show that it is.
That gap falls to the organisation rather than the regulator. Under Article 14 of the AI Act, Article 22 of the GDPR, the public law duty to exercise a discretion and professional duties of individual assessment, whether a review was genuine is a question of fact, and in practice the organisation will be expected to show that it was.
An organisation that has told a regulator, a court or a customer that a human reviews the output has made a commitment it will be asked to evidence.
Departure rate is the proportion of decisions in which the person with authority over the outcome disagreed with the system's output and acted to change it: declining what it approved, approving what it declined, or selecting differently from its ranking.
A departure changes the outcome, not merely the route to it; adjusting a score without changing the decision does not count. Compute it per system and per reviewer, because the aggregate hides the individual. A team at four per cent overall may contain four reviewers at fifteen per cent and eleven at zero.
It comes from logs that, in our experience, most organisations already keep. For one system, it takes one analyst a day.
Across enough decisions, a rate at or near zero has several explanations. The model may be accurate on this case mix. A threshold may be routing only clear-cut cases to the reviewer. The reviewer may examine each case and reach the same outcome. The population may have few borderline decisions.
Or the reviewer may be confirming the output without examining it.
The number looks the same in every case. The organisation should be able to say which explanation applies and how it knows. If it cannot, a challenger will argue that the review was nominal, and there will be no evidence with which to answer.
Direction matters more than level. A rate that falls steadily over several quarters, with no change to the model or the case mix, indicates reviewers deferring more and more to the system, and the fall can show in the series before it shows in complaints or audit findings.
Every report should also state the base-rate limitation. Departure rate measures how often reviewers respond, and that can only be interpreted against how often the system is actually wrong, which organisations generally do not know.
On its own, departure rate is a weak signal. Two measures from the same logs make it stronger: time per decision, read as a distribution, and coverage, the share of decisions that received any human attention. Coverage shows whether anyone looked, time how long they looked, and departure whether looking changed anything. They show where review may be nominal; they do not prove it.
Sampled independent review carries the most weight. A second competent reviewer, blind to the system's output, decides a sample of approved cases afresh. Only this measure tests whether agreement with the system was right, and it is usually skipped because it takes the most effort.
Recorded reasons on departure show whether an override engaged with the case. Abstention and escalation show whether the reviewer had any route other than a forced decision.
The measures are read against three conditions: the reviewer's competence to judge the decision, the information needed to depart and the authority to act. Where accepting takes one click and overriding takes a written justification and a manager's sign-off, a low rate is a rational response to the design, and the number is measuring the interface.
A minimum departure rate rewards disagreement rather than sound review. A reviewer who overrides correct outputs to show independence improves the metric and worsens the decisions. Track the distribution, investigate changes against model performance and case mix, and set no quota.
Report it without overclaiming. An organisation may say what it measured and what it did about the results. On these measures alone, it may not say that oversight is effective.
Green (2022) surveyed 41 policies that prescribe human oversight of government algorithms and found they rest on an uninterrogated assumption that people can oversee algorithmic decision-making effectively. Automation bias affects experts as well as novices, and the literature does not support treating training as a reliable standalone safeguard. The findings are not uniform, but their weight is that human oversight cannot be relied on to work as policy assumes.
Human oversight is therefore itself a dependency of the system. The more consequential the system, the less sensible it is to treat a reviewer as an assumed means of correcting its errors. The question to ask of each system is whether it should be relying on a person to catch the error at all; where it should not, redesign it so that it needs less oversight.
Where review remains, manage it as a performance control. Correctness alignment scores each reviewer's overrides against what actually happened, over a horizon that matches the decision: days for a fraud flag, a year or more for a loan.
Computing the number creates evidence that could be used against the organisation. But the number already exists in the logs, and declining to compute it means the first person to do so is likely to be someone with an adverse interest. Measuring turns an open-ended, unmanaged exposure into a bounded, dated problem with a remediation plan.
If you are deciding whether to run the measurement, start with section 4 of the full paper, Departure Rate, for the definition and section 5 for how to read the number, then section 6 for the companion measures. If you already run it, section 8 sets out what you may and may not claim, and section 11 how to move to correctness alignment. Section 10 is for anyone deciding which systems should rely on review at all, and the closing section, Where to start, and what it costs, gives the staged path.
