Frameworks for Responsible AI
Responsible AI is the development and use of AI systems in ways that are ethical, transparent, and accountable, and that respect human values and rights. It's not a separate ethics process bolted onto the side of an ML project — it's integrated into every stage of the work. The chapters we've covered so far (sourcing, EDA, preprocessing, engineering) all touch one or more of its pillars.
Seven Pillars
- Ethics and value alignment. Designing and deploying AI aligned with ethical principles, human rights, and societal values like fairness, privacy, security, and wellbeing.
- Transparency and explainability. Making systems and their decision processes understandable to stakeholders.
- Privacy and data governance. Responsibly managing data used to train models. Protecting privacy. Collecting and using data ethically.
- Fairness and non-discrimination. Mitigating bias. Not discriminating against individuals or groups on the basis of sensitive attributes.
- Human oversight and control. Building in the ability for humans to intervene and maintain meaningful control over critical decisions.
- Robustness and safety. Ensuring systems are tested, monitored, and reliable. Risk management.
- Accountability and governance. Clear lines of accountability. Audit trails. Governance frameworks that outlast any individual engineer's tenure.
You're joining a team building a hiring recommendation tool that will help screen resumes for a large company. Which of the seven pillars do you think are most critical to address before the tool ships — and what specific questions would you ask?
A useful organizing framework for three of those pillars is FAT: Fair, Accountable, Transparent. Each term sounds simple. None of them is.
Fairness — No Single Universal Definition
There is no single universal definition of fairness. If you think about how you define fairness, it might be very different from how your neighbor describes it, or how your mom describes it.
Two useful distinctions:
- Individual fairness. Similar individuals should receive similar outcomes. A system is individually fair if two people who are alike in all relevant ways get the same result.
- Group fairness. Different groups should experience similar rates of positive outcomes or similar error rates. A system is group-fair if the average outcome doesn't differ systematically across protected groups.
These two definitions are often in tension. A system can be individually fair but group-unfair. A system can be group-fair but individually unfair. Choosing between them is a value judgment, not a technical decision — and should be made explicitly and documented.
Accountability means clear lines of responsibility for outcomes. Users have recourse if they identify issues. Three key questions: Who is responsible for system performance? On what set of values and laws is the system based? What recourse do users have if the system is not behaving in accordance with those values and laws?
Transparency means users have visibility into data usage and model functioning. Document your training data, feature engineering choices, evaluation methodology, known failure modes, and deployment constraints. Transparency is what makes accountability possible — you can't hold anyone responsible for what isn't documented.
A recidivism prediction tool has equal false positive rates across racial groups (group fair), but two individuals with identical criminal histories receive different risk scores because one has a family member with a prior conviction (individually unfair). Which statement is most accurate?