Metrics question

An early rollout of a multimodal model is driving strong user retention, but harmful and policy-sensitive image and audio outputs are also rising. What safety and user-experience metrics would you define, how would you instrument them, and what thresholds would trigger scaling the rollout, limiting it, or pausing it?

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 and threshold design for managing a live rollout where growth and harm signals are moving in opposite directions.

Define safety metrics: rate of confirmed policy-violating outputs per some fixed volume of generations, segmented by modality, image versus audio, since risk profiles differ.
Define user-experience metrics: retention and engagement, tracked alongside, never in isolation, so growth is always read next to the safety trend.
Propose instrumentation: automated classifiers flagging likely violations in near real time, backed by a sampled human review to catch what classifiers miss and to calibrate classifier accuracy over time.
Set explicit thresholds: a defined safety-metric increase over a rolling window, say a week, triggers automatic limiting of rollout percentage, and a further increase triggers a full pause pending investigation.
State the guiding principle clearly: retention gains never justify continuing to scale past a breached safety threshold, since the threshold exists precisely to prevent that tradeoff being made informally under growth pressure.

Proposes concrete rolling-window thresholds tied to specific actions, limit rollout or pause, rather than a vague commitment to watch metrics closely.
Segments safety metrics by modality, recognizing image and audio harms differ in kind and require separate tracking, not one blended number.
Combines automated classifiers with sampled human review, addressing the real risk that automated detection alone misses subtle or novel harms.
States the governing principle explicitly, that safety thresholds are non-negotiable regardless of retention gains, which directly answers the tradeoff the question poses.

Proposing to track retention and safety separately without a stated rule for what happens when they move in opposite directions.
Relying solely on automated classifiers with no human review, missing subtle or evolving harm patterns.

How would you set the specific threshold for pausing versus just limiting rollout?
What would you do if human reviewers and automated classifiers frequently disagreed on borderline cases?