AI Ethics & Governance

AI Safety & Security asks whether the system can be attacked or misused. This branch asks the broader questions: is it fair, transparent, lawful, accountable, and appropriate to deploy at all?

Mental model

Governance turns social and legal obligations into accountable decisions, evidence, controls, and review throughout an AI system's lifecycle. Ethics asks what ought to be done; governance defines who decides, under which criteria, and how the decision can be audited or challenged.

Roadmap: responsible AI foundations

• Responsible AI landscape maps fairness, privacy, transparency, safety, accountability, and social impact.
• Bias and fairness: sources and types explains where unfairness enters the pipeline.
• Fairness metrics and impossibility tradeoffs covers group metrics, individual fairness, and incompatible criteria.
• Measuring and mitigating bias turns fairness concerns into audits, slices, interventions, and monitoring.

Transparency, privacy, and documentation

• Transparency and explainability covers SHAP, LIME, interpretability, and limits.
• Model cards and documentation records intended use, data, evaluation, limits, and risks.
• Privacy, consent, and data rights connects AI governance to data rights across the lifecycle.

Governance and oversight

• The EU AI Act and risk tiers summarizes the risk-based structure and engineering implications.
• AI governance frameworks maps NIST AI RMF, ISO/IEC 42001, OECD principles, and internal governance.
• Accountability and human oversight makes responsibility explicit across design, release, and operation.

Broader impacts

• Societal and labor impact covers deployment effects beyond model metrics.
• Environmental cost of AI covers energy, water, hardware, and carbon-aware decisions.

Connects to: Interpretability · AI Safety and Security · Privacy and PII in Datasets

Core sources

• EU AI Act text — authoritative legal text for the European risk-based framework.
• NIST AI Risk Management Framework — govern, map, measure, and manage functions with supporting profiles.
• Model Cards for Model Reporting — structured disclosure of intended use, evaluation, and limits.
• Datasheets for Datasets — documentation questions covering dataset motivation, composition, collection, and maintenance.
• Fairness and Machine Learning — technical definitions, trade-offs, and limits of fairness interventions.