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 | BNM FATE Framework — Bank Negara Malaysia's Fairness, Accountability, Transparency, and Ethics (FATE) framework (2023) is Malaysia's most developed sector-specific ethical AI guidance. It requires financial institutions deploying AI in credit decisions to:

• Conduct fairness testing across demographic groups

• Provide explainable reasons for adverse decisions

• Establish AI model risk governance committees

• Maintain human override mechanisms

MDEC Responsible AI programme — MDEC runs a Responsible AI assessment tool and advisory service for Malaysian companies. The "AI Trustworthy Assessment" helps companies identify and mitigate ethical risks before deployment. Academic research — Universiti Malaya's Department of AI and Universiti Teknologi Malaysia (UTM) have active AI ethics research groups. The Malaysian Communications and Multimedia Content Forum (CMCF) has published guidance on AI-generated content disclosure. Cultural considerations — Malaysian AI ethics discourse must address:

• Racial sensitivity (Bumiputera/non-Bumiputera considerations in credit scoring, hiring AI)

• Religious compliance (halal certification for AI in food, Islamic finance applications)

• Multilingual fairness (BM, English, Chinese, Tamil — AI systems may perform unequally across languages)

• Rural-urban digital divide (AI benefits concentrated in urban, English-proficient populations)

Corporate practice — Large Malaysian enterprises including Maybank, Axiata, and Petronas have published Responsible AI principles. Axiata's AI Ethics Framework (2022) is one of the most detailed published by a Malaysian company.

Practical Implementation

Organisations embedding AI ethics in practice typically implement:

AI ethics policy — board-approved principles aligned to international frameworks

Ethics review board — cross-functional committee reviewing high-risk AI deployments

Bias audits — systematic testing of model outputs across demographic groups

Red teaming — adversarial testing for harmful outputs, jailbreaks, and edge cases

Model cards — documentation of a model's intended uses, limitations, and evaluation results

Incident response — processes for investigating and remediating AI-related harms

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.