Override rate

How often human reviewers overturn the system's recommendation. A non-trivial override rate is evidence the humans are still deciding; a near-zero override rate is, absent an independent outcome audit, evidence of rubber-stamping. It is how you audit whether a human-signs boundary is real. Treat it as a prompt to investigate, not a verdict on its own: for a decision with a genuinely low base rate of system error, a low override rate can be honest, which is exactly why the audit, not the number alone, is what settles the question.

The proportion of AI recommendations or actions a human overseer changes or rejects; the most diagnostic oversight metric, where a near-zero rate signals automation bias or quasi-automation rather than a flawless AI. The four-eyes principle (two independent humans confirming before a decision takes effect, written into Article 14 for the most consequential biometric cases) is the complementary control for severe-stakes steps.

The frequency with which a human reviewing an agent's proposed action actually rejects or modifies it, named in the primary framework's guidance on auditing human oversight (p.30). A very low override rate can indicate the checkpoint is functioning well, or that nobody is genuinely reviewing before approving; the two look identical from the metric alone and require further investigation to distinguish.

The proportion of cases in which a human reviewer changes an automated system's recommended outcome, tracked over time and per reviewer. A near-zero override rate is one of the strongest, though not conclusive on its own, signals of symbolic rather than genuine review.

The frequency at which humans reject or modify an agent's proposed action at a checkpoint. A low override rate can signal rubber-stamping behavior rather than agent quality, per the framework's own audit guidance (MGF v1.5, pp.29 to 30). Its deep treatment, including countermeasures, belongs to Topic 4.4. (see Topic 4.4)