AI Ethics, Fairness, and Safety
Responsible AI
Bias and Fairness
AI systems can reflect or amplify social biases present in data, labels, or deployment environments. Fairness requires attention to data collection, metric choice, and the real-world context of use.
Transparency and Explainability
Some AI systems are easy to explain, while others behave like black boxes. Explainability aims to make model behavior understandable to developers, users, and auditors.
Model limitations matter
A model may perform well on average while still failing badly for specific groups, edge cases, or high-stakes situations.
Why is fairness important in AI?
Fairness work aims to reduce discriminatory or harmful behavior across groups.
Correct answer: It helps reduce harmful unequal outcomes
What is a black-box model?
Black-box behavior can be hard to audit, debug, or explain.
Correct answer: A model whose internal decision process is difficult to interpret.
Responsible AI Practices
| Practice | Purpose | Example |
|---|---|---|
| Bias testing | Find uneven performance | Check error rates across groups |
| Documentation | Record intended use and limits | Model cards or datasheets |
| Human oversight | Allow review of critical decisions | Approval before deployment |
| Monitoring | Detect drift and failures | Track post-launch performance |
What is model drift?
Drift occurs when the environment or data distribution changes after training.
Correct answer: Changes in data or behavior that reduce performance over time
Why is monitoring important after deployment?
Real-world deployment is not static, so ongoing checks are essential.
Correct answer: Because model performance can change as data, users, or conditions change.