Ch8. AI Ethics and the Future — Principles for Responsible AI

Why AI Ethics Matters

AI systems already influence decisions in loan approvals, hiring screening, criminal risk assessment, medical diagnosis, and autonomous driving — decisions with direct real-world consequences for people’s lives.

Powerful tools demand serious responsibility. AI ethics is the field of principles and practices governing how AI should be developed and deployed.

AI Bias

Why AI Systems Become Biased

Training data bias: AI learns patterns from data. Biased data produces biased models.

Real-world cases:

Facial recognition bias: MIT research (2018) found error rates of 1% for white men vs. 35% for Black women in major commercial systems
Hiring AI: Amazon built a résumé screening AI that rated women applicants lower because historical hiring data skewed male → system was scrapped
Criminal risk assessment: The COMPAS recidivism tool was found to overestimate recidivism risk for Black defendants

GIGO Principle: “Garbage In, Garbage Out.” A model’s quality is bounded by its training data’s quality. Biased data → biased AI. Data curation is as important as algorithm design.

Mitigation Approaches

Collect diverse data: Represent varied demographics — race, gender, age, geography
Measure fairness metrics: Monitor performance disparities across groups
Diverse development teams: Different perspectives are needed to spot different biases

Explainable AI (XAI)

A black-box AI can’t explain why it made a decision. “Why was my loan denied?” “Why did you flag this scan as cancer?”

Why explainability matters:

Trust: Users accept decisions they can understand
Debugging: Identify why decisions go wrong
Regulatory compliance: EU AI Act mandates explainability for high-risk AI

Key XAI Techniques

LIME (Local Interpretable Model-agnostic Explanations): Approximates complex model predictions locally using simple interpretable models.

SHAP (SHapley Additive exPlanations): Game theory-based method that computes each feature’s contribution to a specific prediction.

Attention Visualization: Visualize where Transformer attention focuses — which words or image regions most influenced the output.

Privacy and AI

AI requires large-scale data that frequently contains personal information.

Differential Privacy: Add calibrated mathematical noise to training data so individual records cannot be inferred from the model. Used by Apple, Google.

Federated Learning: Rather than centralizing all user data, train locally on each device and share only model weight updates. Google uses this for Gboard next-word prediction — your typing data never leaves your phone.

AGI and the Alignment Problem

Artificial General Intelligence (AGI): AI capable of matching human performance across all cognitive tasks. Superintelligence: AI surpassing human performance across all domains.

When and whether these will be achieved is genuinely disputed among experts.

The Alignment Problem: As AI systems become more powerful, ensuring they reliably act in accordance with human values and intentions becomes the critical challenge — and the central open problem of AI safety research.

In 2023, hundreds of AI researchers and executives signed a statement calling AI risk “a global priority alongside other societal-scale risks such as pandemics and nuclear war.”

The EU AI Act: The World’s First Comprehensive AI Law

Passed by the EU in 2024, the AI Act classifies AI systems by risk:

Risk Level	Examples	Regulation
Unacceptable	Social scoring, real-time biometric surveillance	Banned
High-risk	Hiring, credit scoring, medical devices	Strict conformity requirements
Limited	Chatbots, deepfakes	Disclosure obligations
Minimal	Spam filters, AI games	Self-regulation

Key Concept Cards

AI Bias ★★★★★ : Biased training data → biased AI. Real harm documented in facial recognition, hiring, and criminal justice. Diverse data and fairness metrics are essential.

Explainable AI (XAI) ★★★★☆ : Techniques to explain why black-box AI made a decision. LIME and SHAP are leading methods. Mandated for high-risk AI under the EU AI Act.

The Alignment Problem ★★★★★ : Ensuring powerful AI systems reliably pursue human values — the central open problem of AI safety. Core research agenda at OpenAI, Anthropic, and DeepMind.

Practice Quiz

Q1. Why is explainability especially critical for AI in medical diagnosis?

Without understanding why an AI reaches a diagnosis, physicians cannot verify it, correct it when wrong, or explain it to patients. Medicine is a high-stakes domain where a wrong decision can cost a life — the physician must be able to interrogate the AI’s reasoning and integrate it into their final judgment. Regulatory frameworks also increasingly mandate explanations for high-risk AI decisions.

Q2. Why is federated learning better for privacy than traditional centralized training?

Traditional training aggregates all users’ raw data on a central server. Federated learning keeps data on the user’s device; only model weight updates (gradients) are sent to the server. The raw personal data never leaves the device. This is particularly valuable for sensitive domains like healthcare and finance — and makes large-scale learning possible without centralizing personal information.