Generative models reflect training data biases—stereotypes, skewed dialect handling, uneven refusal rates across groups.
Product responses
- Slice evals by demographic proxies where ethical and legal
- Offer reporting and appeal for blocked prompts
- Document known limitations in UX
- Involve policy and ERG stakeholders—not only engineering
Connect to AI ethics
Carry fairness habits from AI ethics lessons—Generative UX amplifies harm velocity because output is natural language.
Content moderation staffing
High-traffic consumer apps need human review pipelines; B2B devtools still need abuse monitoring.
Important interview questions and answers
- Q: Why slice metrics?
A: Aggregate satisfaction can hide worse experience for minority dialects or names.
Self-check
- What is a slice eval?
- Why involve non-engineering stakeholders?
Tip: Slice evals where policy allows—aggregate CSAT hides disparate quality.
Interview prep
- Slice evals?
Detect disparate quality hidden in aggregate metrics.
- Why policy input?
Fairness definitions are organizational, not purely technical.