False Omission Risk
A conditional risk metric that is the counterpart of false discovery risk, referring to cases where the system should take an action but instead chooses a conservative fallback. In binary decision settings such as allow or block, it measures the error rate over the subset of inputs on which the utility-maximizing action was taken.
Timeline (last 90 days)
The paper's core argument is: relying on probabilistic interfaces to drive deployment decisions is insufficient, and in scenarios lacking a human fallback mechanism, what truly needs to be controlled is the false omission risk on the subset of inputs for which the utility-maximizing action has been taken, rather than classification accuracy or AUC.
False omission risk is a type of conditional risk metric that is the counterpart to false discovery risk. In binary decisions, a false discovery means the system took an action but the action was wrong; a false omission means the system should have taken an action but chose conservative fallback.
Engineering teams can use a scalar risk budget (for example, a false omission rate of no more than 5%) as the deployment threshold, instead of making vague trade-offs among multiple entangled error rates.
All Facts (3)
The paper's core argument is: relying on probabilistic interfaces to drive deployment decisions is insufficient, and in scenarios lacking a human fallback mechanism, what truly needs to be controlled is the false omission risk on the subset of inputs for which the utility-maximizing action has been taken, rather than classification accuracy or AUC.
50%UnverifiedFalse omission risk is a type of conditional risk metric that is the counterpart to false discovery risk. In binary decisions, a false discovery means the system took an action but the action was wrong; a false omission means the system should have taken an action but chose conservative fallback.
50%UnverifiedEngineering teams can use a scalar risk budget (for example, a false omission rate of no more than 5%) as the deployment threshold, instead of making vague trade-offs among multiple entangled error rates.
50%