Responsible AI interview questions: fairness, safety and governance

Twelve common responsible AI interview questions

1. Where does bias in an ML system come from?
2. Name and compare two fairness metrics.
3. Can a model be fair by every metric at once?
4. How would you explain a model's prediction to a user?
5. What are SHAP and LIME, and what are their limits?
6. How do you protect personal data used in training?
7. What is differential privacy?
8. How do you red-team a generative AI system?
9. What is a model card?
10. What does the EU AI Act require for high-risk systems?
11. How would you audit an AI hiring tool?
12. What would you do if a launch deadline conflicted with a fairness concern?

Sources of bias

Bias enters through historical data that reflects past discrimination, sampling that under-represents groups, labels that encode human judgement, proxy features that correlate with protected attributes (postcode for ethnicity, career gaps for gender), and feedback loops where a model's outputs shape its future training data. A good answer pairs each source with a mitigation: re-sampling or re-weighting, better labelling guidelines, removing or constraining proxies, and monitoring outcomes after launch.

Fairness metrics

When base rates differ between groups, several of these criteria cannot all hold at once — a known impossibility result. Interviewers want you to say that choosing a metric is a value judgement that should involve stakeholders, then document the choice.

Explainability and privacy

• SHAP attributes a prediction to features using Shapley values; LIME fits a simple local model around one prediction. Both explain the model, not the world, and can be unstable or misleading with correlated features.
• User-facing explanations should be actionable: the main factors and what would change the outcome.
• Privacy: minimise data, remove or pseudonymise identifiers, restrict access, and consider differential privacy, which adds calibrated noise so individual records can't be inferred from outputs.
• Generative models can memorise training data, so test for regurgitation of personal information.

Safety testing and documentation

Red-teaming means deliberately attacking a system — jailbreaks, prompt injection, harmful content requests, edge-case inputs — before users do, and turning findings into test cases that run on every release. Model cards document a model's intended use, training data, evaluation results across groups, limitations and ethical considerations; datasheets do the same for datasets. Together they make reviews and audits possible.

Regulation, especially in hiring

The EU AI Act classifies AI used in employment decisions as high-risk, with obligations such as risk management, data governance, human oversight, transparency and logging phasing in over 2026 and 2027. New York City's Local Law 144 requires bias audits and candidate notices for automated employment decision tools, enforced since July 2023. Our guide to AI interview laws and candidate rights covers what candidates can ask for, and AI interview bias covers the candidate's view.

Practising

Answer these out loud, in two minutes or less each, and expect a follow-up on every claim. A free AI interview assistant such as CoPilot Interview can run responsible AI questions as a mock interview built from your résumé and score each answer. Prepare one real example where you found or prevented a fairness, privacy or safety problem, and what process changed as a result.

Disclosure: CoPilot Interview is our product, so weigh what we say about it accordingly. Everything else on this page applies whichever tool you use, or none.

Frequently asked questions

What questions are asked in a responsible AI interview?

Sources of bias, fairness metrics and their trade-offs, explainability methods, privacy techniques, red-teaming, documentation such as model cards, regulation, and situational questions about handling fairness concerns under deadline pressure.

What is the difference between demographic parity and equalised odds?

Demographic parity requires similar positive outcome rates across groups. Equalised odds requires similar true positive and false positive rates across groups, which accounts for actual qualifications.

How would you audit an AI hiring tool for bias?

Compare outcomes across demographic groups using appropriate fairness metrics, check for proxy features, test with counterfactual inputs, review the training data, and document results — as laws such as NYC Local Law 144 require.

What is a model card?

A short document describing a model's intended use, training data, evaluation results including across groups, limitations and ethical considerations, so others can judge whether to use it.