AI Governance & Ethics — Cheatsheet

24 Feb, 2026

Quick reference based on Alison: AI Governance and Ethics and some extra follow-up readings

Five Pillars of Responsible AI

1. Bias & Fairness

Where Bias Comes From

Bias Mitigation Techniques

• Data-level — Re-sampling, re-weighting, expanding data collection to underrepresented groups
• Model-level — Fairness-aware algorithms, adversarial debiasing, fairness constraints in objective functions
• Process-level — Regular audits of data and outputs, diverse development teams, independent oversight
• Organizational — Ethics boards, cross-disciplinary collaboration (CS + ethics + law + sociology)

Key Evidence

2. Transparency & Explainability

The Problem

Deep learning models act as "black boxes" — high performance but low interpretability, undermining trust in high-stakes domains (healthcare, criminal justice, finance).

Solutions

Why It Matters

• GDPR grants individuals the right to an explanation for automated decisions
• Transparency enables auditing, accountability, and stakeholder trust
• Regulators increasingly mandate explainability

3. Accountability

Key Principles

• Clear responsibility — Define who owns AI outcomes (design, deployment, operation)
• Human oversight — Maintain ability to intervene and override AI decisions
• Auditability — Systems must be inspectable; keep logs of decisions and reasoning
• Governance frameworks — Regular audits, impact assessments, continuous monitoring

Challenges

• Highly autonomous systems (self-driving, automated trading) blur responsibility lines
• Probabilistic/non-deterministic outputs make tracing causation harder
• Rapid AI evolution outpaces legal frameworks

4. Data Privacy & Security

Privacy Measures

Security Priorities

• Defend against breaches, hacking, and adversarial attacks
• Maintain continuous patching and vulnerability assessments
• Align with ISO standards and cybersecurity best practices
• Partner with cybersecurity specialists

Cultural Practices

• Train all staff on data protection and AI ethics
• Foster a privacy- and security-first culture
• Establish whistleblower channels for reporting unethical behavior

5. Societal Impact

• Job displacement — McKinsey estimates up to 375M workers may need to switch occupations by 2030
• Mitigation — Reskilling/upskilling programs, social safety nets, policy collaboration
• Equitable distribution — Ensure AI benefits are shared broadly, not concentrated
• Democratic integrity — Guard against AI-enabled manipulation (e.g. Cambridge Analytica)

Regulatory Landscape

Case Study Lessons

HealthCore (Healthcare AI)

• Challenge: Predictive analytics on sensitive patient data
• Approach: Privacy impact assessments → anonymization + encryption → explainable AI → independent oversight → staff training
• Lesson: Multi-tiered governance (consent + transparency + security + cooperation) enables innovation while protecting patients

MediTech (Diagnostic Tool Bias)

• Challenge: AI diagnostic tool underperformed for minority patients
• Approach: Expanded data collection → fairness-aware re-weighting → adversarial debiasing → clinician + ethicist collaboration → ongoing audits
• Lesson: Bias mitigation requires coordinated technical, organizational, regulatory, and educational interventions — fixing data alone is not enough

TechNova (Regulatory Compliance)

• Challenge: Navigating GDPR/CCPA while integrating AI at scale
• Approach: Cross-functional team (data science + legal + compliance) → anonymization + consent → explainable models → IEEE-aligned audits → human override mechanisms
• Lesson: Proactive compliance builds trust; embed ethics into governance from day one, not as an afterthought

Cambridge Analytica (Data Misuse)

• What happened: Harvested millions of Facebook users' data without consent for political ad targeting
• Lesson: Informed consent is non-negotiable; robust data governance and transparency prevent manipulation and preserve democratic integrity

PredPol (Predictive Policing)

• What happened: Algorithm perpetuated racial bias by training on historically biased crime data
• Lesson: Biased data in → biased predictions out; independent oversight and transparent algorithmic processes are essential

Watson for Oncology (Healthcare AI Failure)

• What happened: Provided unsafe treatment recommendations based on hypothetical rather than real patient data
• Lesson: Rigorous testing and validation before deployment; continuous monitoring post-deployment; AI must complement — not replace — human judgment

Organizational Action Checklist

• Data — Audit training data for representativeness and historical bias
• Models — Apply fairness-aware algorithms; use LIME/SHAP for explainability
• Governance — Establish an ethics board or committee with cross-disciplinary membership
• Privacy — Implement privacy by design, impact assessments, anonymization, and encryption
• Human oversight — Design systems that allow human intervention and override
• Audits — Schedule regular audits of AI performance, fairness, and compliance
• Compliance — Map to GDPR, CCPA, EU AI Act, and relevant domain standards
• Training — Educate all staff on AI ethics, data protection, and bias awareness
• Diversity — Build diverse, inclusive development teams
• Collaboration — Engage with policymakers, academia, and cross-industry forums (FAccT, IEEE, OECD)
• Monitoring — Continuously monitor deployed systems for drift, bias, and security threats
• Incident response — Maintain whistleblower channels and breach response procedures

References & Links

Regulations & Laws

• GDPR — Official Text (EUR-Lex)
• CCPA — Official Page (CA Attorney General)
• EU AI Act — Official Text (EUR-Lex)

Ethics Frameworks & Principles

• IEEE Ethically Aligned Design / Global Initiative
• Asilomar AI Principles (Future of Life Institute)
• EU Ethics Guidelines for Trustworthy AI (EC Digital Strategy)
• Google's AI Principles
• OECD AI Principles

Standards

• ISO/IEC 42001:2023 — AI Management Systems

Research Organizations & Initiatives

• AI Now Institute
• ACM FAccT (Fairness, Accountability, and Transparency)
• AI4ALL

Explainability Tools

• LIME — GitHub | Docs
• SHAP — GitHub | Docs

Key Research & Case Studies

• Buolamwini & Gebru — "Gender Shades" (2018) | MIT Media Lab
• ProPublica — COMPAS "Machine Bias" (2016) | Methodology
• Obermeyer et al. — "Dissecting Racial Bias in a Healthcare Algorithm" (2019)
• Hofstra et al. — "The Diversity–Innovation Paradox in Science" (2020)
• McKinsey Global Institute — "Jobs Lost, Jobs Gained" (2017)

Privacy Techniques

• Differential Privacy — Wikipedia Overview | Dwork Original Paper (PDF)