AI & Technical question

OpenAI is preparing to launch a multimodal feature that can take in and generate audio, images, and video. How would you build a launch-readiness risk framework that maps abuse cases, model failure modes, harm severity, and mitigation coverage, and then decide which risks must be blocked before launch versus launched with monitoring and fallback controls?

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

AI safety and technical judgment on building a structured risk framework for a fundamentally more capable and harder-to-moderate multimodal launch.

Build the framework as a matrix: map each abuse case, like non-consensual imagery, voice cloning for fraud, or violent content generation, against modality, audio, image, video, and against severity of potential harm.
For each cell in the matrix, assess current mitigation coverage, like classifier accuracy or watermarking effectiveness, and flag gaps where coverage is weak relative to severity.
Set a hard rule: any high-severity abuse case with weak mitigation coverage must be blocked before launch, not shipped with a promise to monitor and fix later.
Allow lower-severity or well-covered risks to launch with monitoring and fallback controls, like rate limiting or rapid takedown mechanisms, rather than blocking the whole capability over a minor gap.
Define success: percentage of identified high-severity abuse cases with verified mitigation coverage before launch, targeting effectively complete coverage as the gate, not a partial best-effort standard.

Structures the framework as a concrete matrix, abuse case by modality by severity, rather than a vague list of concerns.
Sets an explicit, non-negotiable gate: high-severity gaps block launch, which directly answers the question's ask for a block-versus-monitor decision rule.
Proposes real, specific mitigations, like watermarking and rapid takedown, rather than generic promises to be careful.
Names a concrete pre-launch coverage target for high-severity risks, making the framework auditable rather than a subjective judgment call.

Proposing to launch broadly with monitoring for all risk categories, including high-severity ones, treating post-launch fixes as sufficient.
Listing abuse cases without a structured way to weigh severity against actual mitigation coverage.

How would you validate that a mitigation like watermarking is actually effective before relying on it as a gate?
What would you do if a high-severity abuse case had no available mitigation at all?