AI systems are making high-stakes decisions about people’s lives right now, and they’re getting it wrong in ways that disproportionately harm marginalized groups. Amazon’s hiring tool (2018): trained on historical resumes, the system learned to penalize resumes containing the word “women’s” and preferred candidates from all-male colleges. Amazon scrapped it. COMPAS recidivism: a risk assessment algorithm used in US courts was found to falsely label Black defendants as high-risk at nearly twice the rate of white defendants. Healthcare AI: UnitedHealth’s nH Predict system reportedly had a 90% error rate while denying elderly patients coverage for post-acute care. Facial recognition: MIT’s Gender Shades study (Buolamwini & Gebru, 2018) found error rates of 0.8% for light-skinned men vs. 34.7% for dark-skinned women.

Most AI ethics frameworks converge on five core principles: Fairness — AI systems should not discriminate against individuals or groups based on protected attributes (race, gender, age, disability). Transparency — people affected by AI decisions should understand how and why decisions are made. Accountability — there must be clear responsibility for AI outcomes, including mechanisms for redress when things go wrong. Privacy — AI systems should respect data rights, minimize data collection, and protect personal information. Safety — AI systems should be reliable, robust, and should not cause harm. These principles appear in the EU AI Act, OECD AI Principles, IEEE Ethically Aligned Design, and corporate AI ethics guidelines from Google, Microsoft, and others.

Unethical AI is expensive: Regulatory fines — the EU AI Act imposes fines up to €35 million or 7% of global annual turnover for violations. Lawsuits — class-action suits against biased AI in hiring, lending, and healthcare are increasing. Reputational damage — Amazon’s biased hiring tool and Apple Card’s gender discrimination made global headlines. Lost customers — users abandon products they don’t trust. Employee attrition — engineers leave companies that deploy harmful AI (Google’s Timnit Gebru firing led to significant talent loss). The business case is clear: investing in responsible AI upfront is cheaper than cleaning up harms after deployment.

AI ethics is not one person’s job — it’s a shared responsibility across the entire AI lifecycle: Data scientists & engineers — choose appropriate training data, test for bias, implement fairness constraints, document model limitations. Product managers — define acceptable risk levels, ensure user disclosure, design appeal mechanisms. Leadership — set organizational values, fund ethics infrastructure, establish governance. Legal & compliance — interpret regulations, ensure compliance, manage risk. Affected communities — should be consulted in design, have mechanisms for feedback and redress. Regulators — set guardrails, enforce standards, protect public interest. The most common failure mode is diffusion of responsibility: everyone assumes someone else is handling ethics.

Multiple organizations have published AI ethics frameworks: OECD AI Principles (2019) — adopted by 46 countries, emphasizing inclusive growth, human-centered values, transparency, robustness, and accountability. EU AI Act (2024) — the world’s first comprehensive AI law, using a risk-based approach with four tiers from minimal to unacceptable risk. IEEE Ethically Aligned Design — technical standards for ethical AI, including well-being metrics and data agency. NIST AI Risk Management Framework (2023) — US voluntary framework for managing AI risks across the lifecycle. Corporate frameworks — Google, Microsoft, Meta, and others publish their own AI principles, though enforcement varies.

2016: ProPublica’s investigation of COMPAS reveals racial bias in criminal justice AI. 2018: Amazon scraps biased hiring tool; MIT Gender Shades study exposes facial recognition disparities; Timnit Gebru and Joy Buolamwini publish groundbreaking bias research. 2019: OECD AI Principles adopted by 46 countries. 2020: Google fires Timnit Gebru over a paper on LLM risks, sparking industry-wide debate. 2023: NIST AI RMF published; ChatGPT and generative AI raise new ethical questions about hallucination, copyright, and deepfakes. 2024: EU AI Act enters into force (August 1). 2025: Prohibited AI practices banned (February); general-purpose AI obligations apply (August). 2026: High-risk AI obligations fully enforceable (August 2).

Myth: “AI is objective.” Reality: AI reflects the biases in its training data and the choices of its designers. An algorithm trained on biased historical data will reproduce and amplify those biases. Myth: “Ethics slows down innovation.” Reality: Ethical failures slow down innovation far more — through lawsuits, regulatory action, and lost trust. Building ethics in from the start is faster than retrofitting. Myth: “We just need more data.” Reality: More biased data produces more confident biased predictions. Data quality and representativeness matter more than quantity. Myth: “Ethics is subjective.” Reality: While some ethical questions are debatable, many harms are measurable (disparate impact, accuracy gaps across demographics) and can be addressed with engineering rigor.

This course covers AI ethics from principles to practice: Chapters 1–4 (Foundations): Why ethics matters, how bias enters AI systems, formal fairness definitions and metrics, and practical bias mitigation techniques. Chapters 5–7 (Transparency & Privacy): Explainability and interpretability (SHAP, LIME), privacy and data rights (GDPR, differential privacy, federated learning), and LLM-specific ethical challenges (hallucination, copyright, deepfakes, alignment). Chapters 8–10 (Governance & Regulation): The global regulatory landscape (EU AI Act, US policy), AI safety and alignment, and building ethical AI in practice (governance frameworks, impact assessments, ethics review boards).