Metrics question

OpenAI wants one topline safety metric for frontier model deployments. How would you define it so it is credible for leadership decisions, sensitive enough to detect meaningful changes in harm, and decomposable into drivers that research and engineering teams can act on?

Practice this question out loud. An AI interviewer asks it, follows up like a real interviewer would, and scores your answer. Type or speak.

What this question tests

Metrics design for an ambiguous, high-stakes safety concept that must satisfy both leadership decision-making and engineering actionability.

Confirm with the interviewer what deployment scope this covers, a single model launch or the ongoing fleet of deployed models, since the metric design differs.
Define the topline as a composite: rate of confirmed harmful outputs per some fixed volume of real production interactions, sampled and reviewed against a severity-weighted rubric.
Address credibility for leadership: anchor the composite to independently audited definitions of harm categories, not an internal-only scoring method leadership cannot trust.
Address sensitivity: ensure the sample size and review cadence are large and frequent enough to detect a meaningful shift, not just quarterly spot checks.
Address decomposability: break the composite into drivers, like harm category, deployment surface, and user segment, so research and engineering can trace a topline move to a specific cause.

Proposes a severity-weighted rate rather than a raw incident count, since one number must reflect both frequency and how bad each incident is.
Insists on independent or externally auditable review criteria as the credibility anchor for leadership, since a self-graded metric will not survive scrutiny.
Explicitly designs the metric to decompose by harm category and surface, so a topline change always has a clear, actionable next question for teams to chase.
Names the tradeoff: a single number is inherently a simplification, so it should always ship with its component breakdown attached, never presented alone.

Proposing a single opaque score with no visible decomposition, which leadership can trust but teams cannot act on.
Using raw incident counts without severity weighting, which treats a minor and a severe harm as equivalent.

How would you set the sampling rate to balance cost against statistical sensitivity?
How would you handle a harm category that is rare but catastrophic when it occurs?