Designing a Safety Strategy for a Consumer Generative AI Product
Company: Google
Role: Product Manager
Category: Product Design & Strategy
Difficulty: medium
Interview Round: Other
As a product manager, how would you approach safety for a consumer product powered by generative AI?
```hint Rank the harms first
A consumer AI product can fail in many ways, and the failures are not equally costly. Decide which harms matter most for the product you choose, and remember that wrongly refusing a legitimate request is also a failure users notice.
```
### Clarifying Questions
- What does the product do: answer questions in a chat, generate images or other media, or take actions such as purchases and messages on the user's behalf?
- Who can use it: adults only, or teenagers and children as well?
- Does it remember users between sessions or read their personal data, such as email, photos or files?
- Is it about to launch, or already live with reported safety incidents to address?
### What a Strong Answer Covers
- A map of the harms this product can cause (harmful content, false answers, privacy leaks, misuse, unsafe actions), prioritized by severity and likelihood
- Layered controls across product design, inputs, the model, retrieved context, outputs and actions, and why no single layer is enough
- How safety is measured before launch and in production, including how often legitimate requests are wrongly refused
- A launch, monitoring and incident-response plan that feeds real failures back into testing
- The trade-off between safety and usefulness, and how it shifts with the audience and the product's capabilities
### Follow-up Questions
- The assistant gains permission to read a user's inbox and reply on their behalf. Which new risks appear, and what changes in your controls?
- After a safety update, wrongful refusals double and weekly usage drops. How do you decide whether to roll back?
- A screenshot of a harmful answer spreads on social media on launch day. What do you do in the first 48 hours?
Overview: Product strategy question on keeping a consumer generative AI product safe. It tests whether a product manager can rank the harms an AI assistant can cause, design layered controls across inputs, the model, outputs and actions, measure safety without over-refusal, and plan launch and incident response.