AI Ethics and Governance

Why AI Ethics Is Engineering, Not Philosophy

AI ethics is sometimes framed as an abstract philosophical concern — disconnected from the practical work of building systems. This framing is wrong and dangerous. Ethical failures in AI are engineering failures with concrete, measurable consequences: a hiring algorithm that systematically rejects qualified candidates from certain demographic groups; a credit scoring model that perpetuates historical lending discrimination; a medical imaging AI that performs well on majority demographic groups and poorly on underrepresented ones; a content moderation system that censors minority languages at higher rates than the majority language.

These are not hypotheticals. Amazon scrapped an AI recruiting tool that downgraded résumés containing the word "women's." ProPublica found that the COMPAS recidivism prediction tool used in US courts was twice as likely to falsely flag Black defendants as future criminals compared to white defendants. Google Photos infamously labelled photographs of Black people as gorillas.

Each of these failures had an engineering root cause: biased training data, inappropriate proxy variables, insufficient evaluation on subgroups, or misaligned optimisation objectives. Addressing AI ethics means building better data pipelines, better evaluation frameworks, better monitoring systems, and better organisational processes — not just writing a values statement.

The FATE Framework: Fairness, Accountability, Transparency, Explainability

The FATE framework provides a practical structure for ethical AI implementation:

**Fairness**: AI systems should not discriminate against individuals or groups based on protected characteristics. Operationalising fairness requires choosing a mathematical fairness definition — demographic parity (equal positive prediction rates across groups), equalised odds (equal true and false positive rates), or individual fairness (similar individuals treated similarly) — and accepting that these definitions are mathematically incompatible with each other in most practical settings. Fairness decisions are value judgements that require domain expertise and stakeholder involvement, not just optimisation.

**Accountability**: Clear ownership of AI system decisions. Who is responsible when an AI makes a consequential error? Accountability requires documented model governance (who approved deployment?), audit logs (what data was used? what was the decision?), and defined escalation paths (how do affected individuals contest a decision?).

**Transparency**: Honest disclosure of when AI is being used, what data it processes, and what its limitations are. Transparency obligations are now legally mandated in many jurisdictions — GDPR's right to explanation, the EU AI Act's transparency requirements for high-risk systems, and national consumer protection frameworks.

**Explainability**: The ability to provide human-understandable reasons for AI decisions. Highly complex models (large neural networks) are inherently opaque — this is the "black box" problem. Techniques like SHAP values, LIME, and attention visualization provide post-hoc explanations, but the field of inherently interpretable models (decision trees, linear models, rule-based systems) is sometimes the right architectural choice for high-stakes applications.

The Regulatory Landscape: EU AI Act and Beyond

The **EU AI Act** (fully effective 2027) is the world's most comprehensive AI regulation. It takes a risk-tiered approach:

**Prohibited AI practices**: Social scoring systems, real-time biometric surveillance in public spaces, subliminal manipulation systems. These are banned outright.

**High-risk AI systems**: AI used in critical infrastructure, education, employment, credit scoring, law enforcement, and border control. These require conformity assessments, mandatory registration, human oversight mechanisms, and rigorous documentation. Violations carry fines up to 7% of global annual revenue.

**Limited-risk systems**: Chatbots and AI-generated content must be disclosed as AI. This is a transparency obligation, not a capability restriction.

**Minimal-risk systems**: AI-enabled spam filters, recommendation engines — subject only to voluntary codes of conduct.

**GDPR** imposes additional constraints on AI systems processing personal data of EU residents: a legal basis for processing, data minimisation requirements, the right to human review of automated decisions (Article 22), and the right to erasure. For AI systems trained on personal data, GDPR compliance requires careful data governance throughout the pipeline.

Beyond the EU, the US AI Executive Order (2023), UK AI Safety Institute, China's generative AI regulations, and Brazil's AI Bill all create a patchwork of compliance requirements for globally operating organisations.