AI Ethics

AI ethics is the interdisciplinary study of the moral obligations and societal implications arising from the design, deployment, and governance of artificial intelligence systems. As AI becomes embedded in consequential decisions — credit scoring, medical diagnosis, hiring, criminal sentencing, content moderation — the field has moved from academic philosophy to a mainstream corporate and regulatory concern.

Core Principles

Fairness and Non-discrimination

AI systems must not discriminate on the basis of protected characteristics (race, gender, religion, age, disability, national origin) unless such distinctions are legally justified. Algorithmic fairness is technically complex because multiple mathematical fairness definitions — demographic parity, equalised odds, calibration — are often mutually incompatible.

Key issues:

• Training data reflecting historical discrimination
• Proxy variables correlating with protected attributes
• Feedback loops amplifying initial biases

Accountability

When an AI system causes harm, there must be clear lines of responsibility. The "accountability gap" arises when:

• AI developers disclaim responsibility for third-party deployments
• Deployers argue they are merely using tools
• Users accept responsibility they don't fully understand

Emerging regulatory approaches (EU AI Act, UK AI Bill) assign accountability based on the role in the AI value chain and the risk level of the application.

Transparency and Explainability

• Transparency — openness about when AI is used, what data it was trained on, and how it reaches outputs
• Explainability — providing intelligible reasons for specific decisions (especially important in high-stakes contexts)
• Interpretability — understanding the internal mechanisms of a model

The trade-off between predictive performance and interpretability is real: simpler linear models are interpretable but less powerful; deep neural networks are powerful but opaque. Explainable AI (XAI) methods (LIME, SHAP, attention visualisation) provide post-hoc explanations.

Privacy

AI systems often require large personal datasets for training and inference. Privacy concerns include:

• Collection of personal data without meaningful consent
• Re-identification attacks on anonymised data
• Inferences revealing sensitive attributes not directly provided
• Membership inference attacks (determining if an individual was in training data)

Privacy-preserving ML techniques: federated learning, differential privacy, homomorphic encryption, synthetic data generation.

Safety and Reliability

AI safety encompasses both near-term safety (preventing system failures, adversarial attacks, specification gaming) and long-term safety (AI alignment — ensuring advanced AI systems pursue intended goals even as they become more capable).

Near-term safety practices: red teaming, adversarial testing, formal verification, uncertainty quantification, human-in-the-loop requirements, circuit breakers.

Human Oversight

High-stakes AI decisions should maintain meaningful human oversight. This is sometimes called "human-in-the-loop" (HITL) or "human-on-the-loop" (HOTL) depending on the degree of human involvement. Fully autonomous AI in life-critical contexts (medical diagnosis, lethal autonomous weapons) is generally regarded as ethically problematic.

International Frameworks

| Organisation | Framework | Key Feature | |-------------|-----------|-------------| | OECD | AI Principles (2019) | First intergovernmental standard; widely adopted | | UNESCO | Recommendation on AI Ethics (2021) | UN member states; non-binding | | EU | AI Act (2024) | Binding law; risk-tier classification | | G7 | Hiroshima AI Process (2023) | Voluntary code of conduct | | IEEE | Ethically Aligned Design | Technical standards for practitioners | | ASEAN | Guide on AI Governance and Ethics | Regional non-binding guidance |

Practical Implementation

Organisations embedding AI ethics in practice typically implement:

1. AI ethics policy — board-approved principles aligned to international frameworks
2. Ethics review board — cross-functional committee reviewing high-risk AI deployments
3. Bias audits — systematic testing of model outputs across demographic groups
4. Red teaming — adversarial testing for harmful outputs, jailbreaks, and edge cases
5. Model cards — documentation of a model's intended uses, limitations, and evaluation results
6. Incident response — processes for investigating and remediating AI-related harms

References

1. OECD (2019). Recommendation of the Council on Artificial Intelligence. OECD/LEGAL/0449.
2. UNESCO (2021). Recommendation on the Ethics of Artificial Intelligence. SHS/BIO/PI/2021/1.
3. BNM (2023). Responsible AI in Financial Services — Discussion Paper. Bank Negara Malaysia.
4. Axiata Group (2022). Axiata AI Ethics Framework. Axiata Group Berhad.