Evidence-Grounded RAG Assistant for TikTok Misinformation Reviews
Company: TikTok
Role: Product Manager
Category: Product Design & Strategy
Difficulty: medium
Interview Round: Other
You are a product manager at TikTok. Design a system built on retrieval-augmented generation (RAG) that helps the content moderation team spot misinformation and decide how to handle it.
```hint Decide what the model stands on
Before drawing the pipeline, ask what evidence each judgment should rest on, who keeps that evidence current as new claims spread, and what the system should do when it finds nothing relevant.
```
### Clarifying Questions
- Should the system support human moderators as a review aid, or also act on videos automatically, for example at upload?
- Which misinformation areas come first: health, elections and civic processes, crises and breaking news, or every topic?
- Which parts of a video can the system read: speech, on-screen text, captions and hashtags, comments, user reports?
- Which markets and languages must the first version cover?
### What a Strong Answer Covers
- Which moderation roles the system serves and which step of their workflow it speeds up or makes more consistent.
- What the retrieval corpus contains, how sources are vetted and kept fresh, and how claims are pulled out of short videos that mix speech, text and visuals.
- How outputs turn into actions (label, limit distribution, remove, escalate), where humans stay in the loop, and how creators can appeal.
- Metrics that weigh catching misinformation against wrongly acting on legitimate content, plus a pre-launch evaluation plan.
- Risks such as claims with no evidence yet, hallucinated citations, evasion tactics and uneven quality across languages.
### Follow-up Questions
- A false claim about a breaking news event goes viral before any fact-check exists. What does your system do in the first hours?
- Reviewers begin approving the system's draft verdicts without reading the evidence. How would you detect that, and what would you change?
- How would you evaluate the system before launch when labeled misinformation is scarce and changes every day?
Overview: A product design question for a TikTok product manager: design a retrieval-augmented generation system that helps content moderators detect misinformation in short videos and decide how to act on it. It tests reviewer workflow design, evidence sourcing and freshness, human-in-the-loop enforcement, and metrics that weigh accuracy against over-removal.