1. Fairness, privacy, and safety

You've walked the interviewer through your comment moderation model: it scores every new comment and hides anything above 0.8. Then comes the question: "How do you keep this model from harming users, leaking data, or being gamed?"

That question is about the constraints a model has to meet besides accuracy. It must not make systematically worse mistakes for one group of users (fairness). It must not expose the personal data it was trained on or logged (privacy). And it has to keep working when people try to fool it or corrupt its training data (adversarial robustness). Human review and explanations sit on top as the product's safeguards for the cases the model still gets wrong.

The basic controls are cheap and always apply: measure errors per group, keep only the personal data the model needs, and send uncertain cases to people. Heavier privacy techniques that trade accuracy for protection are usually not worth that cost unless you release statistics or train on very sensitive data.

In most machine learning (ML) design rounds this is a few minutes near the end, after evaluation and serving. Trust-and-safety, lending, and hiring roles probe it much deeper. A mid-level answer names where bias comes from, compares error rates per group, drops unneeded personal data, and routes uncertain cases to people. A senior answer picks the fairness metric by the harm, tells gaming apart from poisoning, and says honestly when differential privacy or federated learning is worth its cost. At any level, raising one of these risks before you're asked is a strong signal.

We'll use one example throughout: a community app that runs a toxic-comment classifier on about 5M comments a day. It's trained on past moderator decisions plus user reports, its users write in several English dialects, including AAE (African American English), and moderators review a queue of borderline comments.

This tutorial covers the technical controls inside one predictive model. Safety of generated outputs, such as large language model (LLM) guardrails and red-teaming, belongs to an LLM course, and governance paperwork like model cards and launch sign-off is the process organizations wrap around these decisions.