Detecting and Containing LLM Hallucinations in a Customer-Facing AI Product
Company: Google
Role: Product Manager
Category: Machine Learning
Difficulty: medium
Interview Round: Other
A generative AI system is live, and actual customers depend on its answers, but it sometimes produces confident statements that are false or that no source supports. As the product manager, explain how you would handle these hallucinations.
```hint Not every wrong answer costs the same
Before choosing fixes, sort the failures by how much harm a wrong answer causes and by whether the customer could easily spot it, because that sorting decides where to spend effort.
```
### Clarifying Questions
- What does the system do for customers (answer questions from the company's documents, handle support cases, draft content, or take actions on their behalf), and in which domain?
- Does the team use a third-party model through an API or train its own, which decides whether fixes can go into the model or only around it?
- Is this a response to a specific incident, or a standing plan for a product that is about to scale?
- Are the customers consumers, or businesses that pass the output on to their own users?
### What a Strong Answer Covers
- A working definition of hallucination for this product, and a classification by severity and by how easily a customer can verify the output
- Measurement: an evaluation set, sampling of production output and user signals that together give a hallucination rate per use case
- Layered mitigations across scope, grounding in sources, generation and verification, and product design such as citations, declining to answer and handoff to a person, with what each costs
- Incident response when a harmful hallucination reaches customers
- The balance between helpfulness and risk (over-refusal, latency, cost) and how the remaining risk is communicated to customers
### Follow-up Questions
- Stricter grounding cut hallucinations sharply, but the assistant now declines to answer twice as often and satisfaction fell. What do you do?
- How would you explain the remaining hallucination risk to a business customer's legal team?
- Which hallucination metric goes on the team's main dashboard, and how do you stop it from being gamed?
Overview: An AI product question asking how a product manager would deal with hallucinations in a generative AI system that real customers rely on. It tests how hallucinations are defined and measured, layered mitigations such as grounding and declining to answer, incident response, and balancing helpfulness against cost and risk.