Pacing the launch of a high-capability, high-risk AI feature against safety needs
Company: OpenAI
Role: Product Manager
Category: Product Design & Strategy
Difficulty: medium
Interview Round: Other
As a PM at OpenAI, you own the launch of a highly capable new feature that also has real potential for harm. How would you balance shipping quickly against the safety constraints it needs?
```hint Not every risk needs the same gate
Separate the possible harms by how severe, how likely and how reversible they are. Ask which ones must be ruled out before anyone outside the company gets access, and which ones you can safely learn about from limited real use.
```
### Clarifying Questions
- What is the capability, and what is its most serious realistic harm: deliberate misuse by bad actors, costly mistakes the system makes by itself, or exposure of private data?
- Who would get it first: developers through the API, consumers in ChatGPT, or enterprise customers under contract?
- What is pushing for speed: a competitor's launch, a partner commitment, or the value of learning from real usage sooner?
- Which safety requirements are already fixed (company risk-evaluation thresholds, legal obligations, public commitments), and which are left to the PM's judgment?
### What a Strong Answer Covers
- What "velocity" and "safety" each mean for this capability, and whether the choice really is a single trade-off.
- A risk assessment that separates harms by severity, likelihood, reach and reversibility, and names the limits no deadline can override.
- A rollout plan with measurable criteria, set in advance, for expanding access or pulling it back.
- Mitigations, monitoring and incident response, and the working practices that let the team move faster without lowering the bar.
- Decision rights and communication: who can block or approve each step, how disagreements escalate, and what is said publicly.
### Follow-up Questions
- A competitor ships a similar capability next week with far fewer restrictions. What, if anything, changes in your plan?
- Red-teaming finds a misuse path you can make harder but cannot close. What do you do?
- After a limited launch, misuse reports are very low. How do you decide whether that is enough evidence to expand?
Overview: An OpenAI product manager strategy question about launching a highly capable new feature that carries real potential for harm. It tests how the candidate weighs shipping speed against safety constraints, assesses risks, sets criteria for expanding access, and decides who can approve or block each step.